The Accessible Internet in the Age of Agentic Artificial Intelligence

Executive Summary

The accessible internet is entering a period in which the most consequential question may no longer be only whether a blind person can operate an interface control by control, but whether that person can accomplish a digital objective with comparable efficiency, information and autonomy.

For more than two decades, modern web accessibility has been organized around an indispensable idea: information and functionality should be made perceivable, operable, understandable and robust for people with disabilities. The Web Content Accessibility Guidelines, now at WCAG 2.2, remain the most widely recognized technical framework for pursuing that goal.¹ The logic of WCAG is not obsolete in an AI era. On the contrary, semantic roles, accessible names, keyboard operability, meaningful structure and text alternatives become inputs not only for human assistive technologies but also for increasingly capable software agents. Accessibility metadata is becoming shared infrastructure for humans and machines.

The need remains enormous. WebAIM’s February 2026 analysis of one million highly ranked home pages detected 56,114,377 accessibility errors, or 56.1 per page on average. Detectable WCAG failures appeared on 95.9% of home pages. The study also found an average of 1,437 elements per home page, a 22.5% increase in one year, illustrating that interface complexity is growing at the same time that longstanding accessibility failures remain common. WebAIM explicitly cautions that automated testing cannot detect every conformance problem, so the absence of automated errors does not establish accessibility.² This is an important baseline: AI is not entering an already accessible digital environment. It is entering one in which many blind users still encounter unlabeled controls, missing alternative text, poorly structured forms and complex interaction patterns.

Generative AI has already altered that environment. Interviews with 19 blind people published at ACM ASSETS 2024 showed that participants used mainstream generative-AI systems for content creation, information seeking, planning, learning and visual interpretation, while also dealing with inaccessible interfaces, hallucinations and the work of verification.³ A 2026 CHI diary study with 20 blind and low-vision participants found that a multimodal visual-assistance system was rated relatively favorably for trust and satisfaction, yet 22.2% of responses were incorrect and another 10.8% abstained.⁴ These findings capture the central paradox of AI accessibility: the technology can unlock information that was previously expensive or inconvenient to obtain, but it can also create a new layer of uncertainty.

The move from generative AI to agentic AI raises the stakes. NIST describes AI agent systems as combining at least one generative model with scaffolding software and tools that allow the system to take discretionary actions in an environment.⁵ For accessibility, this matters because the system can progress beyond describing an inaccessible interface. An agent may search, navigate, fill fields, move between applications, inspect documents, create files, compare options, schedule events or complete portions of a workflow. This can shift the interaction from a low-level sequence—find a landmark, move to a form field, identify a label, open a menu, navigate options, verify focus—to a high-level request such as “find the next available appointment after 3 p.m. and show me the options before booking.”

That shift can be disproportionately useful for blind users because conventional screen-reader interaction is often sequential. A sighted user can visually scan a screen, notice hierarchy, compare distant regions and infer state changes in parallel. Screen-reader users can achieve sophisticated and efficient interaction through headings, landmarks, element lists and keyboard shortcuts, but inaccessible markup and unfamiliar applications can make this process slower and cognitively expensive. Research systems such as Savant have therefore explored natural-language access to desktop controls. In a study with 11 blind screen-reader users, Savant reported substantially faster control access and lower workload than status-quo interaction for the evaluated tasks.⁶ AskEase, evaluated with 12 screen-reader users at CHI 2026, similarly showed that context-aware AI guidance could improve task success and reduce perceived workload without hiding the original interface.⁷

Computer-use agents extend this idea from guidance or single-action automation to longer workflows. Yet the early evidence is mixed. The A11y-CUA study collected 40.4 hours and 158,325 interaction events from eight blind/low-vision users and eight sighted users across 60 everyday tasks. A state-of-the-art agent succeeded on 78.33% of tasks in the default condition but fell to 41.67% when constrained to keyboard-only interaction and 28.33% under a magnified-viewport condition. The results suggest that present-day agents often assume sighted interaction patterns and do not naturally reflect assistive-technology use.⁸ A September 2026 study of an accessible computer-use agent with eight blind users collected 1,258 commands across 12 applications over three weeks; the strongest evaluated model achieved a 52.5% success rate, with failures in grounding, planning, constraint tracking and termination.⁹ These are meaningful capabilities, but they are far from a reliability level at which blind users should be expected to delegate consequential tasks without oversight.

The most important design lesson emerging from this literature is that accessibility cannot be measured only by task completion. An agent can reach the nominal target state while still reducing user agency. If several products have the same lowest price and the agent silently chooses one, the transaction may be technically successful while the user is deprived of information that a sighted person would ordinarily see. Morae, presented at UIST 2025, addressed this by proactively pausing UI automation when the system detected meaningful ambiguity and inviting blind and low-vision users to choose.¹⁰ This mixed-initiative pattern offers a useful principle: automate mechanical work aggressively, but do not automate away preference, consent or meaningful choice.

Agentic AI also introduces security and privacy questions that are especially relevant for assistive use. An agent that can make a website easier to operate may need access to screenshots, accessibility trees, documents, messages, calendars or accounts. NIST’s 2026 request for information on AI-agent security highlights risks such as prompt injection, data poisoning, compromised models, specification gaming and unintended effects on confidentiality, integrity or availability. It specifically raises human approvals for consequential actions, least privilege, monitoring, rollback and constrained deployment environments as important areas of control.⁵ The UC Berkeley Agentic AI Risk-Management Standards Profile likewise emphasizes human control, intervention points, escalation, monitoring, traceability, containment and shutdown mechanisms.¹¹

These controls should not be treated as barriers to accessibility. Properly implemented, they are accessibility features. A blind user should be able to know what an agent intends to do, what it has already done, what information it relied upon, what remains uncertain and which step requires consent. The user should be able to pause, undo, correct and resume the workflow without reconstructing the entire visual state of the interface. Confirmation should be proportional to consequence: reading a public web page can be highly automated; sending money, submitting a legal form, deleting files or sharing sensitive information should require accessible confirmation. The human should retain authority over values and commitments while the machine absorbs the mechanical burden of navigation.

The broader social context reinforces the significance of this technology. An OECD analysis based on interviews with more than 70 stakeholders identified 142 examples of AI-powered solutions that could support people with disabilities in the labor market and reported that more than three quarters would not exist without AI. At the same time, the OECD warned that unreliable or inaccessible systems can worsen discrimination and exclusion, especially when disability is poorly represented in training data.¹² The International Labour Organization similarly frames AI and disability employment as containing both opportunities to harness and risks to mitigate.¹³ AI should therefore be judged not only by aggregate productivity but by whether it reduces or redistributes the extra labor disabled people already perform to participate in digital systems.

This paper argues for a layered future. First, websites and applications should remain natively accessible: semantic HTML, meaningful labels, keyboard operability, well-designed focus, accessible authentication and compatibility with screen readers remain non-negotiable. Second, AI should enhance perception and comprehension by describing images, reorganizing complex information and explaining unfamiliar interfaces. Third, conversational systems should let users express goals in natural language while preserving access to the underlying structure. Fourth, bounded agents should execute repetitive or inaccessible action sequences while pausing at meaningful choices and consequential actions. Finally, the entire stack should produce accessible logs, state summaries and verification evidence so the user can remain in control.

The result is not a world in which AI “solves blindness,” nor one in which conventional accessibility becomes unnecessary. It is a world in which digital systems can be accessed at multiple levels of abstraction. A user may inspect every control with a screen reader, ask an AI to summarize the current state, request guidance for a difficult step, or delegate a bounded sequence to an agent. Independence grows when the person—not the interface designer and not the agent—can choose the level of assistance appropriate to the situation.

1. Introduction: From Interface Access to Goal Access

Digital accessibility has always been about more than compliance. Its practical purpose is to make participation possible. A bank website is accessible not because its code satisfies a checklist, but because a person can privately understand an account balance, compare transactions, authenticate, transfer funds and recover from an error. An education platform is accessible not because every image has an alternative attribute, but because a student can find assignments, read material, submit work, receive feedback and communicate with instructors. Technical standards matter precisely because they are a means to those human ends.

For blind and low-vision people, screen readers are among the most transformative technologies in modern computing. They convert programmatic interface information into speech or braille, expose structures such as headings and landmarks, and enable keyboard-driven operation that can be exceptionally efficient in well-designed software. Expert users routinely navigate complex systems at speeds that surprise sighted observers. The difficulty is not that nonvisual interaction is inherently primitive. The difficulty is that much of the digital world is designed visually first, with accessibility added inconsistently or not at all.

This mismatch creates a recurring tax on blind users. When a developer forgets to label a button, the sighted user may still infer its purpose from an icon, location or nearby text. The screen-reader user may hear only “button.” When a dialog is rendered outside the accessibility tree, it may be visually obvious but programmatically absent. When a page changes dynamically, a sighted user may notice the change immediately, whereas a nonvisual user must discover it through focus movement, live regions or deliberate exploration. Each individual failure can appear minor to a developer. In aggregate, they create what disability researchers often describe as additional or invisible work: time spent devising workarounds, searching for alternative tools, asking for assistance, memorizing quirks and verifying states that others can see at a glance.

Artificial intelligence enters this environment with an unusual property: it can operate at a higher semantic level than traditional assistive technology. Optical character recognition can extract text from an image; a multimodal language model can potentially explain what the image means in context. A screen reader can announce a table cell; a language model can summarize patterns across the table. A browser exposes hundreds of nodes; an agent can search those nodes for the handful relevant to a user’s goal. A conventional automation script can execute a predefined sequence; an agent can potentially plan a new sequence in an unfamiliar interface. The difference is not merely improved recognition. It is a change in the abstraction level of interaction.

That change motivates the central distinction of this paper: interface access versus goal access. Interface access asks whether the user can perceive and operate the available controls. Goal access asks whether the user can achieve the purpose for which those controls exist. The two are related but not identical. A website can expose every control correctly and still be needlessly difficult to use. Conversely, an AI agent may be able to accomplish a task on an inaccessible website even though the underlying interface remains discriminatory. The ideal future is not to choose between them. It is to combine native accessibility with goal-level assistance.

Agentic AI makes this combination particularly important. Unlike a chatbot that produces text and waits, an agent can use tools, interact with software and continue acting across multiple steps. The agent may inspect a webpage, search for a product, compare results, add an item to a cart, populate a form or move information between applications. For a blind user, this can remove large amounts of mechanical navigation. Yet it also transfers some perception and action from the person to the system. If the agent misunderstands the interface, silently chooses among alternatives or acts on a hallucinated state, the user may not have a visual channel through which to notice the mistake. The technology that reduces an accessibility burden can therefore increase a verification burden.

The research question is not simply whether agentic AI is useful. It clearly can be. The stronger question is: under what conditions does agentic AI increase digital independence for blind users rather than replacing one inaccessible mediation layer with another? Answering that requires attention to interface semantics, multimodal perception, human-agent interaction, security, user agency, error recovery, privacy, personalization and institutional access.

This monograph synthesizes evidence through September 29, 2026. It focuses primarily on blind and low-vision computer users, especially people who use screen readers, because the interaction differences between visual and nonvisual computing make the agentic-AI question unusually concrete. Many principles, however, generalize to other disability communities. A system that clearly communicates state, exposes uncertainty, respects user preferences, supports multiple modalities and asks for consent before consequential actions is likely to be better for many users, not only those who are blind.

The paper also rejects two common extremes. One extreme treats AI as an accessibility cure that will eventually make standards, semantic markup and specialist assistive technologies irrelevant. That view ignores the dependence of reliable agents on structured interfaces and the current evidence of model error. The other extreme treats AI-mediated access as inherently inferior because it does not force the underlying interface to become accessible. That view underestimates the value of tools that allow disabled people to act now, without waiting for every developer and institution to remediate every barrier. Accessibility history has always included both structural reform and user-side adaptation. Agentic AI belongs in that tradition, but with more power and therefore more responsibility.

2. Concepts, Definitions and Scope

2.1 Blindness and low vision

The term “blind and low vision” covers diverse sensory experiences, identities and access practices. Some people have no useful vision; others have residual vision that varies with lighting, contrast, distance, fatigue or disease. Some primarily use speech output, some rely on refreshable braille, some combine screen readers with magnification, and some switch between methods. Any claim that “blind users” need one particular interface should therefore be treated cautiously. Accessibility requires flexible systems that accommodate multiple strategies.

2.2 Screen readers and accessibility trees

A screen reader does not usually “look” at a screen in the way a person does. It relies heavily on structured information exposed by the operating system, browser or application. In the web context, HTML semantics and ARIA contribute to an accessibility tree that represents elements in terms such as role, name, state, value and relationships. This model allows a user to navigate by headings, links, buttons, form fields, landmarks and other structures. It can also support braille output and programmatic interaction. When developers use native controls and meaningful labels, the tree can provide a rich nonvisual representation. When they create custom widgets without equivalent semantics, the representation can collapse.

2.3 Generative AI

Generative AI refers here to systems capable of producing new content—text, images, code, audio or structured outputs—in response to input. Large language models and multimodal large language models are the most relevant classes for this paper. In accessibility contexts, generative systems may summarize documents, rewrite complex text, explain code, describe images, answer questions about screenshots or synthesize information from several sources. They can be useful without taking any external action.

2.4 Agentic AI

Agentic AI adds an action layer. Definitions vary, but the common elements are goals, reasoning or planning, access to tools or environments, and the ability to perform a sequence of actions with some degree of discretion. NIST’s 2026 security inquiry describes AI agent systems as at least one generative model combined with scaffolding software that equips the model with tools to take discretionary actions, with some systems coordinating multiple sub-agents.⁵ This definition is useful because it distinguishes an agent from a text generator while avoiding the claim that an agent must be fully autonomous.

2.5 Computer-use agents

Computer-use agents are a subset of agentic systems that operate graphical user interfaces. They may observe screenshots, DOM structures, accessibility trees or operating-system UI metadata, then click, type, scroll, invoke semantic actions and inspect the resulting state. Their attraction is generality: instead of requiring an application-specific API, they can potentially operate the same interface a human uses. Their weakness is also generality: graphical interfaces are ambiguous, dynamic and often poorly structured. For accessibility, a screenshot-first agent can reproduce a sighted interaction model that is difficult for blind users to monitor.

2.6 Assistance, automation and agency

This paper uses assistance for systems that provide information or guidance while the user performs the action, automation for systems that execute predefined or bounded actions, and agency for systems that select or adapt actions in pursuit of a goal. These categories overlap. A conversational assistant may become agentic when given tools. An agent may become assistive when it is designed to narrate and explain rather than merely execute. The design question is therefore not whether a product belongs in a perfect category, but what decisions it can make, what consequences it can create and how the user can supervise it.

2.7 Digital independence

Digital independence is not defined here as doing everything without assistance. People use assistants, search engines, automation, colleagues and software services constantly. Independence is better understood as the ability to pursue one’s own objectives without unnecessary dependence on another person, while retaining meaningful choice, privacy and control. An AI system can increase independence if it removes a need to ask a sighted person to read a screen, but it can reduce independence if it hides options, acts without consent or makes the user dependent on an opaque provider with no accessible fallback.

3. Methodology

This work is a narrative evidence synthesis and design analysis. It is not a systematic review, scoping review or meta-analysis, and it should not be interpreted as having the completeness guarantees associated with those methods. The purpose is to integrate several rapidly developing bodies of evidence that are usually discussed separately: traditional web accessibility, blind users’ experiences with generative AI, multimodal visual assistance, conversational navigation, computer-use agents, employment inclusion and AI risk management.

Search and source review were conducted on September 29, 2026. Five search workstreams were used: (1) web accessibility, screen-reader interaction and standards; (2) generative and multimodal AI used by blind and low-vision people; (3) agentic AI and computer-use agents in accessibility contexts; (4) human oversight, security, privacy, reversibility and risk management for AI agents; and (5) disability inclusion, employment and assistive technology. Forty-one search results were reviewed across these workstreams. Sixteen high-value sources were retrieved for closer reading, supplemented by additional source records used to identify emerging work.

Sources were prioritized in roughly the following order: peer-reviewed empirical research with blind or low-vision participants; standards bodies and government or intergovernmental institutions; peer-reviewed technical evaluations; accepted conference papers and high-quality preprints when the subject was too recent for final publication; and vendor documentation only for factual descriptions of capabilities. Search-engine summaries and commercial commentary were not treated as evidence when a stronger source was available.

The literature is unusually time-sensitive. Models, interfaces and commercial systems can change within months. Consequently, findings tied to a specific model are interpreted as evidence about design patterns and failure modes rather than permanent rankings of products. For example, a 52.5% success rate for a particular computer-use model in 2026 is valuable because it demonstrates the remaining difficulty of real-world accessible automation; it should not be read as a forecast of that model’s future performance.

The analysis uses three lenses. The first is capability: what new tasks become possible or easier because of AI? The second is accessibility alignment: does the system support the interaction practices, verification needs and preferences of blind users? The third is agency preservation: does the system leave meaningful decisions, consent and correction under user control? A technology can score well on capability while failing on alignment or agency. The argument of this paper is that long-term usefulness depends on all three.

Several limitations should be acknowledged. Many user studies in accessible computing have small samples because recruitment of specific disability populations is difficult. Participants in technology studies may be more technically experienced than the broader population. Most of the research reviewed here was conducted in high-income countries and often in English, even though multilingual performance and infrastructure vary globally. Finally, agentic AI is evolving quickly, so architecture recommendations should be treated as design hypotheses supported by current evidence rather than permanent rules.

4. The Accessibility Baseline: A Web That Still Requires Repair

Any evaluation of AI as an accessibility technology should begin with the environment it is supposed to improve. The modern web remains deeply inconsistent for disabled users. WebAIM’s 2026 Million study provides one of the clearest quantitative snapshots. Using the WAVE engine against the rendered DOM of one million highly ranked home pages, the project detected more than 56 million distinct accessibility errors—56.1 per page on average—and found detectable WCAG failures on 95.9% of home pages.² These errors do not represent every barrier, because automated analysis can identify only a subset of accessibility failures. The study therefore should not be interpreted as saying that 4.1% of pages are fully accessible. The true rate of complete conformance is necessarily lower.

The most frequent failures are remarkably persistent. In 2026, low-contrast text appeared on 83.9% of home pages, missing alternative text on 53.1%, missing form-input labels on 51%, empty links on 46.3%, empty buttons on 30.6% and missing document-language declarations on 13.5%. These six categories accounted for 96% of detected errors.² For a screen-reader user, several of these failures directly degrade interaction. A form field without a label may be difficult to identify. An empty button may expose no meaningful control name. An image without a text alternative may remove content entirely. An empty link can force exploration of context that a sighted user does not need.

At the same time, pages are becoming more complex. WebAIM counted more than 1.4 billion elements across the sample, with an average of 1,437 elements per home page, up 22.5% in one year. Complexity does not necessarily cause inaccessibility, but it increases the amount of state a user or agent must interpret. A dense interface creates more opportunities for focus errors, mislabeled controls, repeated navigation and dynamic updates. It also highlights a fundamental asymmetry: visual interfaces can display many options simultaneously, whereas speech is serial.

WCAG 2.2 remains the primary standards response to these problems. Its principles—perceivable, operable, understandable and robust—are deliberately technology-neutral.¹ The robustness principle is particularly important in an agentic context because it requires content to be interpretable by a range of user agents, including assistive technologies. Semantic HTML and well-implemented accessibility APIs can therefore support both human screen readers and machine agents. A button with a stable role, accessible name, state and native action is easier for VoiceOver or NVDA to communicate and easier for an AI system to ground than an unlabeled image positioned at screen coordinates.

This suggests that the rise of agents should strengthen, not weaken, incentives for semantic accessibility. Accessibility metadata can function as machine-readable interface infrastructure. Recent agent research increasingly uses accessibility trees because they compress a complex visual interface into structured text that models can reason about. The same information was originally built to support disabled users. That convergence creates an opportunity: improvements made for people can also improve automation reliability.

There is, however, a risk of the opposite incentive. If an agent can visually interpret a broken interface and click the right pixel, organizations may be tempted to claim that the accessibility problem has been “solved” by an overlay or AI mediator. This would be a mistake. The user may not want or be able to use that mediator. The agent may fail. The underlying page may remain inaccessible to alternative tools. Privacy-sensitive tasks may require local, direct access. Accessibility is a property of the system, not of a single helper layered on top.

The correct relationship is complementary. Native accessibility gives the user and the agent a reliable foundation. AI can then reduce friction that remains even in technically conformant interfaces. A long document can be summarized. A complex dashboard can be explained. A user unfamiliar with a workflow can request context-aware guidance. An agent can handle repetitive sequences. The better the underlying semantics, the less the AI must guess.

This distinction also matters for accountability. Traditional accessibility failures are often deterministic: a button is either labeled or not; a form control either receives keyboard focus or not. AI adds probabilistic behavior. If a system attempts to infer the missing label, it may be correct most of the time and wrong at exactly the moment when the consequence matters. A robust accessibility strategy should therefore repair deterministic failures at the source and use AI for tasks where interpretation genuinely adds value.

5. Screen Readers, Sequential Interaction and the Cost of Interface Complexity

Screen readers are sometimes described by sighted observers as linear technologies, and that description is both true and incomplete. Speech itself unfolds sequentially, but expert screen-reader interaction is not limited to listening from the top of a page to the bottom. Users jump by headings, landmarks, links, form controls, tables, regions and custom shortcuts. Element lists provide rapid filtered views. Search can locate text. Braille can offer spatial and linguistic detail unavailable in speech. The accessibility tree can expose relationships that are not visually obvious.

Yet linearity still matters because nonvisual overview formation is costly. A sighted user can glance at a page and absorb layout, grouping, salience and available actions in a fraction of a second. A blind user can obtain equivalent information, but often through multiple navigation operations. This difference is particularly pronounced in unfamiliar software, where the user does not yet know which landmarks, menus or keyboard commands will be most efficient.

Research on accessible computing repeatedly identifies this burden. The AskEase work describes how screen-reader users can face steep learning curves, memorization requirements, difficulty discovering elements and substantial effort in forming an overview of changing interfaces.⁷ The Savant project was motivated by the heterogeneity of desktop applications and the need to remember different interaction patterns. Its natural-language layer was designed to let users express a desired control or operation without manually traversing every intervening element.⁶

The implications for AI design are significant. The greatest accessibility value may not come from replacing speech synthesis or improving the pronunciation of interface labels. It may come from reducing the number of low-level navigation decisions required to accomplish a task. A system that knows the current application state can tell the user, “There are three unread messages and one attachment in the selected thread,” rather than forcing exploration of the entire message pane. A system that can act can go further: “Save the attachment to the Documents folder,” then report exactly what it did.

This is a form of interaction compression. A sequence of ten, fifty or two hundred keyboard events can be represented by one intention. Compression is valuable when the intervening steps are mechanical and the end state is clear. It is dangerous when the steps contain choices that matter. The challenge is therefore to identify the boundary between mechanics and judgment.

Consider online shopping. Finding the cheapest bottle of a particular product may involve searching, filtering, sorting, examining variants, checking shipping terms, choosing a seller and adding an item to a cart. An agent can automate much of this. But “cheapest” may not uniquely identify an option. Two products may have the same price and different sizes, ratings or delivery dates. A sighted user naturally sees these differences while scanning results. If the agent silently chooses, automation has removed not only labor but information. Morae’s design responds directly to this problem by pausing at ambiguous choices and exposing meaningful differences to the user.¹⁰

The same pattern applies to forms, travel, education and work. A user may want an agent to locate a flight but not choose the itinerary. The agent can fill a repetitive address but should not infer a legal declaration. It can find a file but should ask before deleting it. It can format a document but should not silently alter the meaning of the content. Accessibility-enhancing automation should reduce interaction cost while preserving informational parity.

This idea of parity is broader than speed. A blind user should not have to trade away context in order to obtain efficiency. If an agent chooses among hidden alternatives, the user may finish faster but know less. A high-quality system should provide a compressed representation of the decision space: enough information to make the choice without reproducing every irrelevant detail. This is where language models have a genuine advantage. They can summarize differences in natural language and adapt the level of detail to the user’s preference.

The ideal screen-reader-plus-agent workflow may therefore be fluid. The user can navigate directly when precision or exploration matters, ask the agent for an overview when orientation is difficult, delegate a mechanical sequence when the goal is well specified, and retake control instantly when the agent encounters ambiguity. Screen readers remain the transparent base layer; agents become an optional abstraction layer on top.

6. Generative AI: From Recognition to Interpretation

Earlier accessibility AI was often built around narrow recognition tasks. Optical character recognition extracted text. Object detectors identified classes such as “chair,” “person” or “door.” Image-captioning models produced short descriptions. These technologies were valuable, but their outputs were constrained by predefined tasks and labels. Generative and multimodal language models changed the interaction because they can respond to open-ended questions about content and maintain conversational context.

For a blind user, this matters because useful visual information is goal-dependent. A generic caption of a kitchen scene may say that a stove is visible. A person cooking may instead need to know which burner is on, whether a pan is centered, what a display reads or where a control is positioned relative to another. A generic description of a photograph may mention people standing together; the user preparing a social-media post may want to know facial expressions, clothing, lighting and whether anyone’s eyes are closed. Conversational systems allow the user to specify those needs after seeing—or rather hearing—the initial description.

Adnin and Das’s ASSETS 2024 interviews with 19 blind participants provide an early empirical picture of this transition. Participants used generative AI for information seeking, content creation, learning, planning and visual interpretation. They valued the ability to obtain detailed answers and to iterate without repeatedly asking another person. At the same time, they encountered hallucinations, inaccessible controls and uncertainty about how models worked.³ The importance of this study is not that every participant loved the technology. It is that generative AI had already become part of ordinary accessibility practice before formal assistive-AI standards had caught up.

This everyday use reveals a distinction between access and confidence. AI may provide information that was previously unavailable, but the user must decide how much to trust it. That decision is often contextual. A mistake in describing the color of a shirt may be embarrassing; a mistake in reading a medication label may be dangerous. A wrong interpretation of a graph may affect work. Users therefore develop verification strategies, such as asking follow-up questions, comparing multiple tools, using nonvisual sensory cues or involving another person.

Verification cost is central to this paper’s argument. A sighted user who asks AI to summarize a screenshot can glance at the screenshot and detect an obvious error. A blind user may be using AI precisely because the source is not independently accessible. The system’s error can therefore be more expensive to detect. This does not imply that blind users are incapable of critical evaluation; research shows the opposite. It means that interface design should make uncertainty and provenance easier to inspect rather than assuming the user can visually cross-check.

Generative AI can also transform inaccessible digital artifacts that are not strictly visual. A scanned PDF can be restructured into headings and lists. A complex table can be turned into a verbal summary. A poorly formatted email thread can be reorganized by sender and decision. Terminal output can be interpreted semantically. The 2026 ICSE study of 39 blind and low-vision software professionals describes developers using generative AI to interpret tables, screenshots, diagrams and difficult command-line output, reducing some of the invisible labor associated with inaccessible development tools.¹⁴

These examples point toward a broader concept: AI as a semantic adapter. Traditional accessibility APIs aim to expose the structure the developer intended. Generative systems can sometimes infer structure that the developer failed to provide. That inference is powerful but uncertain. The most reliable architecture therefore uses semantic source information when available and generative interpretation as a fallback or enrichment layer, not as a universal replacement.

7. Multimodal AI and Visual Access

Multimodal large language models are among the most immediately visible accessibility applications of generative AI. They can process images alongside text, enabling conversational visual question answering. Commercial assistive tools and mainstream AI products have incorporated such capabilities, allowing blind users to photograph an object, document or scene and ask questions about it.

The empirical evidence shows meaningful progress alongside persistent limits. A 2026 CHI diary study followed 20 blind and low-vision participants using an MLLM-enabled visual interpretation application for two weeks. Participants rated the system’s interpretations as relatively trustworthy and satisfying, but the researchers found that 22.2% of answers were incorrect and 10.8% abstained.⁴ The result is important precisely because it combines subjective usefulness with objective imperfection. A tool does not need perfect accuracy to be valuable, but the pattern of error determines where it can be safely relied upon.

The study also argues that successful assistance requires more than descriptive accuracy. A visual assistant must understand the user’s goal, ask for better input when needed, recognize when it lacks enough evidence, and provide actionable guidance rather than simply narrating what appears in the image. This is a step toward agentic behavior: the system must manage an interaction, not merely generate a caption.

A 2025 ACL study evaluated thirteen multimodal models on user-centered tasks including culturally specific image captioning, multilingual visual question answering, optical braille recognition, object recognition and video question answering.¹⁵ The researchers found substantial gaps in cultural context, multilingual performance, braille comprehension, assistive-object recognition and hallucination resistance. These findings matter because accessibility failures are often distributional. A model that performs well on English-language benchmark images may be far less useful to a blind user reading a sign in another language or interpreting an object uncommon in the training data.

The global dimension is especially important. AI accessibility is sometimes discussed as if the primary problem were adding a better description feature to a smartphone app. In practice, usefulness depends on language support, bandwidth, device capability, cost, data privacy and cultural context. If an assistive model requires a high-end phone and a constant connection to a remote service, it may reproduce existing inequalities in assistive-technology access.

Multimodal systems also illustrate the difference between static interpretation and dynamic action. Describing a microwave control panel from a photograph is one problem. Guiding a user through changing the power level is another. The second task requires temporal context, spatial consistency and feedback about whether the control moved as expected. Current models often struggle with these demands. That is why agentic assistance should not be assumed safe merely because image descriptions have improved.

Nevertheless, the potential is substantial. Visual AI can convert previously inaccessible information into a conversational resource. It can provide private access to mail, packaging, personal photographs, household controls and visual workplace content. It can reduce the social burden of repeatedly asking another person to describe something. It can also give users control over what details they want, rather than depending on what a sighted helper thinks is important.

The strongest design principle is therefore user-directed detail. The AI should not decide once and for all what blind people “need to know.” It should provide a concise initial description, support follow-up questions, allow verbosity preferences, communicate uncertainty and distinguish observation from interpretation. If it infers that someone “looks angry,” it should make clear that this is an interpretation of visual cues, not a fact about the person’s internal state. If text is partially obscured, it should say so. These behaviors become even more critical when visual interpretation feeds into agentic action.

8. Conversational Browsing and the Middle Layer Between Screen Readers and Agents

Conversational browsing occupies an important middle ground. The user expresses a goal or question in natural language, and the system interprets a webpage or application, but the system may stop short of acting autonomously. This model can provide some of the efficiency benefits of agents while leaving execution under direct user control.

A 2026 experimental study involving 30 blind and low-vision participants compared an LLM-enhanced conversational browsing system with traditional screen-reader interaction. The authors reported overall advantages in user experience and accuracy for the conversational approach, while also observing increased cognitive demand and arguing that conversational agents and screen readers should be integrated rather than treated as substitutes.¹⁶ The study’s hybrid conclusion is especially persuasive. Screen readers offer precision, familiarity and direct access to the source. Conversation offers high-level control, summarization and natural-language intent. Users should be able to move between them.

AskEase provides a related pattern. Rather than taking over the interface, it gathers context such as the current screen state, screenshots, recent screen-reader output and relevant documentation, then produces screen-reader-friendly guidance. In a within-subjects study with 12 screen-reader users, it improved task success and reduced several dimensions of perceived workload.⁷ Importantly, the system was designed to preserve the original interface instead of masking accessibility failures behind an overlay.

This form of assistance has educational value. When an agent simply completes a task, the user may learn little about the application. When a guidance system explains what is happening and how to navigate it, the user can build a mental model that remains useful when the AI is unavailable. The right choice depends on context. For a one-time interaction with an obscure government website, full automation may be rational. For software used daily at work, guidance that strengthens the user’s own competence may be more empowering.

Conversational systems also make uncertainty easier to communicate than conventional automation. An assistant can say, “I see a button labeled Continue, but I cannot determine whether submitting it will finalize the purchase. Would you like to inspect the surrounding section?” That statement preserves the user’s ability to choose. A poorly designed agent might simply press the button.

The conversational layer is therefore likely to remain important even as agents become more capable. Not every task benefits from autonomy. Sometimes the best accessibility feature is a concise explanation delivered exactly when the user needs it.

9. What Makes Agentic AI Different

Agentic AI is often surrounded by marketing language that makes simple tool use sound like autonomy. For accessibility analysis, the distinction should be practical. A system becomes meaningfully agentic when it can translate a user goal into a sequence of actions, choose among tools or interface operations, inspect intermediate results and adapt the plan without requiring the user to specify every step.

A conventional chatbot can tell a user how to rename a file. An agent can locate the file, invoke the rename command, type the new name, confirm the resulting state and report success. A chatbot can suggest flights. An agent can search travel sites, apply constraints, compare options and prepare a booking. A chatbot can explain an inaccessible spreadsheet. An agent can open it, find the requested sheet, calculate values, create a summary and save the result.

This capability changes the accessibility equation because many barriers occur not at the level of information but at the level of action. A user may know exactly what should happen and still be blocked by an inaccessible date picker, drag-and-drop control, unlabeled menu or visual CAPTCHA. An agent with access to alternative interface representations may be able to bridge some of these barriers.

The architecture of a typical agent can be simplified into five functions: perception, planning, action, memory and verification. Perception gathers the current state through screenshots, DOM, accessibility trees, APIs or files. Planning decides what steps might satisfy the user’s goal. Action invokes tools, clicks controls, types text or calls APIs. Memory retains relevant context across steps or sessions. Verification checks whether the action produced the intended result. Accessibility requirements apply to all five.

Perception should include semantic information, not screenshots alone. Planning should respect user preferences and disability-related constraints. Action should support keyboard and native semantic operations where possible. Memory should preserve accessibility preferences without creating unnecessary surveillance. Verification should produce evidence that can be understood nonvisually. A system that succeeds technically but cannot explain what it did is not fully accessible.

Agentic systems also introduce a new kind of interface: the trajectory. The user is no longer interacting only with the visible page; the user is interacting with a sequence of machine decisions across time. Accessible design must therefore expose not just current state but process state: what the agent intends, what step it is on, what changed, what failed and what it will do next.

This is why logs, summaries and confirmation boundaries are not developer conveniences. They are part of the user interface. A blind person should not have to infer an agent’s behavior from unpredictable changes in a webpage. The agent should narrate at the appropriate level: “I found three matching appointments. The earliest is Tuesday at 9:00 a.m. I have not booked anything.” After confirmation: “I selected Tuesday at 9:00 a.m. and reached the final confirmation page. The page shows a $25 cancellation fee. Do you want me to submit?”

That interaction is agentic, accessible and agency-preserving. The system handles mechanical work while explicitly returning consequential decisions to the user.

10. Why Agentic AI Can Be Disproportionately Useful for Blind Users

The value of agentic AI is not unique to disability. Everyone can benefit from automation. What makes the technology especially important for blind users is that it can target tasks where the cost of direct interaction is systematically higher because interfaces are visual, inconsistent or poorly exposed to assistive technology.

10.1 Compressing low-level navigation

A screen-reader user interacting with a complex page may need to traverse headings, landmarks, links and controls before finding the relevant region. An agent can search the structured interface directly and present only the relevant choices. This does not make the screen reader obsolete; it creates a shortcut at the semantic level. The user can still inspect the original interface when needed.

10.2 Bridging inaccessible controls

Some controls are effectively invisible to a screen reader because they lack semantics or keyboard support. A multimodal agent may still identify them visually. In limited cases, this can make an otherwise blocked task possible. Such bridging should be treated as a fallback, not an excuse for inaccessible design, because coordinate-based clicking is less robust than semantic interaction and may fail when layout changes.

10.3 Reducing context-switching costs

Blind professionals often face inaccessible diagrams, screenshots, tables and development tools. The 2026 ICSE study found that blind and low-vision software professionals used generative AI to interpret visual and structured content and to reduce difficult navigation in coding workflows.¹⁴ An agent can extend this by moving information between tools, such as extracting an error from a terminal, locating the relevant code, suggesting a change and preparing a patch for review.

10.4 Supporting unfamiliar software

Expert screen-reader users become fast partly by learning application-specific shortcuts and structures. This expertise is valuable but expensive to acquire. A natural-language agent can reduce the penalty for unfamiliarity. Savant’s study with 11 blind users showed the promise of uniform natural-language access across several applications, reporting lower workload and faster access to controls in the tested tasks.⁶ The broader idea is powerful: users should not need a new mental grammar for every application.

10.5 Increasing privacy relative to human assistance

Blind users sometimes rely on sighted assistance for tasks involving visual information. Human help can be excellent and should remain available, especially for high-stakes or ambiguous situations. But it can create privacy costs. Asking another person to read financial information, private messages, medical correspondence or identification documents may be uncomfortable. AI can provide an alternative channel that does not require another person to see the material. This privacy benefit is conditional, because cloud-based AI services have their own data practices. Local or privacy-preserving models may therefore become especially important in assistive contexts.

10.6 Making repetition cheap

One of AI’s underrated accessibility benefits is patience. A user can ask for a description again, request more detail, ask the same question in another way or refine a workflow without feeling that they are burdening another person. This is evident in qualitative research on blind users’ engagement with visual AI and generative tools. The ability to interrogate information repeatedly can increase autonomy even when a human would have provided a more accurate first answer.

10.7 Personalizing assistance

Accessibility needs are not uniform. One person may want extremely concise output at a high speech rate. Another may prefer detailed explanations and explicit orientation cues. One user may be an expert in a software package and want automation only for a broken control; another may want step-by-step guidance. Agentic systems can maintain preferences about verbosity, confirmation, navigation style and risk tolerance. Personalized assistance can therefore be more useful than a one-size-fits-all overlay.

10.8 Converting visual parallelism into linguistic summaries

Visual interfaces present parallel information: multiple prices, ratings, dates, colors and states can be seen simultaneously. A well-designed agent can convert that parallel field into a concise linguistic comparison. Instead of reading thirty hotel cards one at a time, the user can ask for the three lowest-priced options that meet accessibility and cancellation constraints, then inspect those options in detail. This is not merely convenience. It changes the cost of comparison.

10.9 Supporting goal discovery

Accessibility barriers sometimes prevent users from knowing what an interface can do. If controls are hidden or unlabeled, the user may not know that a feature exists. An agent that can interpret the full interface can summarize available actions. This addresses what Morae’s field study described as a “not knowing what you don’t know” problem. Yet goal discovery also creates a design responsibility: the agent should expose meaningful possibilities without overwhelming the user.

10.10 Creating bespoke assistive tools

Agentic programming may allow blind users to build specialized accessibility tools through natural language. Emerging research such as ProgramAT reports blind and low-vision co-designers creating dozens of camera-based assistive prototypes, including highly individualized tools that commercial applications did not provide.¹⁷ This direction may be transformative because disability needs are often long-tailed. A commercial vendor cannot build every niche tool, but an agentic programming environment may enable users to create or adapt one.

Taken together, these mechanisms explain why agentic AI can be more than a general productivity feature for blind users. It can target the structural mismatch between visual software and nonvisual interaction. But the same leverage increases the cost of failure. If an agent makes a mistake while doing something the user cannot easily observe, the error can be difficult to detect. Usefulness and oversight must therefore advance together.

11. Use Cases: What Agentic Accessibility Could Look Like in Practice

11.1 Web search and information gathering

Search engines are technically accessible when results use proper headings and links, but serious research still requires opening many pages, dismissing banners, locating relevant sections and comparing information. An agent can reduce this navigation by searching broadly, extracting relevant passages and presenting a source-grounded synthesis. For a blind user, the gain comes not only from summarization but from bypassing repeated page-level interaction. The system should preserve source links and make clear which claims came from which documents so that the user can inspect evidence directly.

11.2 Long documents and inaccessible PDFs

Documents are a common accessibility bottleneck. A PDF may be scanned, untagged, have a broken reading order or contain visually complex tables. Generative AI can reconstruct logical structure, answer questions and identify sections. An agent can go further by searching a collection of documents, extracting specific records, creating a summary and saving an accessible output. For professional use, the system should distinguish text extracted directly from the source from inferred structure and should provide page or section references where possible.

11.3 Email and communication

Email clients can be highly accessible, but large volumes of messages impose navigation and triage costs. An agent could summarize unread messages, identify requests requiring action, locate attachments, draft responses and prepare calendar events. A safe design separates preparation from commitment: the agent may draft freely but should require confirmation before sending external messages, accepting invitations or deleting mail. It should also make recipients and attachments explicit in the confirmation summary.

11.4 Calendar and scheduling

Scheduling is a natural agentic task because it involves gathering availability, comparing constraints and interacting with date/time controls that can be cumbersome. An accessible agent can say, “You have three open 30-minute windows on Thursday afternoon: 1:30, 3:00 and 4:30. The invitee is free at 3:00 and 4:30.” The user chooses; the agent creates the event and reports the result. The agent should never infer that a meeting can be moved or canceled merely because another time looks more convenient.

11.5 Online forms and government services

Forms are disproportionately frustrating when labels, validation and date pickers are inaccessible. Agents can populate repeated information, navigate multi-step pages and translate visual validation messages into clear language. Yet forms often contain declarations with legal or financial consequences. The system should distinguish clerical fields from attestations. It may fill a name and address from an approved profile; it should not answer “I certify that…” without explicit user review.

11.6 Banking and financial services

Financial interfaces combine high accessibility value with high risk. An agent could categorize transactions, find recurring charges, prepare a transfer or compare account terms. But transferring money should require a confirmation that states the source account, destination, amount, fees and date in accessible text. Sensitive account data should be protected through least-privilege access and strong authentication. If an agent cannot verify a transaction state, it should stop rather than retry blindly.

11.7 Shopping

Shopping demonstrates the need for preference-aware agents. The mechanical work of search and filtering is easy to delegate. The decision itself may encode taste, quality, ethics, delivery time or accessibility. Morae’s work shows that ambiguous options should trigger a pause rather than an arbitrary choice.¹⁰ An accessible shopping agent should summarize differences and let the user decide when those differences matter.

11.8 Travel planning

Travel combines comparison, dynamic pricing, maps, visual seat charts and accessibility information. An agent could gather flights, hotels and ground transportation while applying user-defined constraints. It could also search for accessibility policies and contact information. Booking, however, should remain confirmation-gated. Agents should avoid assuming that “accessible” means the same thing for every traveler. A blind user may care more about navigability, staff assistance, transit connections and room layout than features designed primarily for mobility impairments.

11.9 Education

Students can use AI to convert inaccessible visual material, explain diagrams, structure notes and navigate learning-management systems. An agent could collect assignment deadlines, download readings, create accessible summaries and remind the student about submissions. The educational value is highest when the system helps the student access content rather than silently completing assessed work. Institutions should distinguish accessibility support from academic substitution and provide clear policies that do not force disabled students to choose between inclusion and compliance.

11.10 Workplace productivity

In workplaces, the opportunity is broad: project-management tools, dashboards, presentation software, spreadsheets and internal portals frequently contain inaccessible or visually dense content. An agent can translate charts into narrative summaries, prepare reports, locate status updates and automate routine administration. The 2026 ICSE study suggests that blind software professionals already experience a high-return/high-risk relationship with generative AI, using it to remove accessibility barriers while remaining vulnerable to hallucinated information.¹⁴ Agentic systems amplify both sides of that equation.

11.11 Software development

Programming is a particularly strong use case because much development work already occurs in structured text. AI can explain errors, search code, write tests, propose patches and inspect logs. For blind programmers, it can also translate visual artifacts such as screenshots, diagrams and tables. Agentic coding systems may reduce the need to navigate enormous codebases manually, but they can generate large amounts of output and create context-switching burdens. The 2025 study of visually impaired developers using coding assistants reported benefits alongside overload from excessive suggestions and difficulty moving between AI-generated content and existing code.¹⁸ Accessibility therefore requires not simply more suggestions, but better control over timing, scope and presentation.

11.12 Data analysis

Data tools frequently communicate through visual charts. An agent can answer questions about trends, identify outliers, generate accessible tables and create textual explanations. It can also manipulate data directly. Professional use requires reproducibility: the agent should provide formulas, transformations or code where possible, not just conclusions. A blind analyst should be able to audit the reasoning through accessible artifacts.

11.13 Creative work

Design, photography, video and music production contain visual controls that can be difficult to access. AI can describe artwork, identify layout issues, generate metadata, organize media and operate parts of creative software. The risk is that the model may impose its own aesthetic judgments. Accessibility should expand the creator’s control rather than substitute the model’s taste. A creative agent should expose options, describe consequences and make edits reversible.

11.14 Social media

AI can help blind users interpret images, write alternative text, inspect how a post appears and navigate complex publishing interfaces. It can also schedule posts and manage replies. Because social identity and tone are personal, agents should preserve authorship. Drafting and formatting can be automated; public posting should generally remain confirmation-gated unless the user has explicitly configured a narrow automatic workflow.

11.15 Customer service and dispute resolution

Many companies force customers through complex portals before allowing contact. An agent could locate a refund form, collect order information, prepare a complaint and track status. This is an example where agentic AI can rebalance power: the user no longer has to master every company’s interface. Yet the agent should not invent claims or agree to terms without approval. It should preserve a record of what was submitted.

11.16 Personal knowledge management

Blind users, like everyone else, accumulate notes, documents, messages and files. An agent can create semantic organization that is less dependent on visual folder browsing. It can answer, “Where is the receipt from the hotel in Manila?” or “Collect the documents related to the visa application.” This can dramatically reduce search cost. Privacy controls become essential because the agent may need access to a large personal corpus.

11.17 Device setup and troubleshooting

Technical support often assumes the user can see menus and icons. Context-aware agents can interpret the current state and give screen-reader-specific directions, as AskEase demonstrates.⁷ Future agents may be able to perform safe configuration changes directly, while explaining each step. Changes that could disrupt assistive technology—such as disabling a screen reader, changing audio output or altering security settings—should have special safeguards.

11.18 Navigation and physical-world assistance

Physical navigation is a more demanding and safety-critical domain. Multimodal models can describe scenes and identify landmarks, but dynamic guidance requires accurate spatial reasoning, real-time perception and robust uncertainty handling. Current visual-assistance studies document hallucinations and context errors, so agentic navigation should not be treated as a replacement for established orientation and mobility methods. The safer near-term role is supplementary: describing signs, identifying a doorway or answering a question about the environment while clearly communicating uncertainty.

11.19 Bespoke assistive technology

Agentic programming could allow users to create tools tailored to highly specific needs. ProgramAT’s emerging work found that a small group of technically experienced blind and low-vision participants created 37 camera-based assistive prototypes through agentic programming, including tools for individualized situations not served by commercial products.¹⁷ The significance is not the number of prototypes; it is the possibility of shifting disabled people from consumers of assistive technology to authors of it.

11.20 Cross-application workflows

The most transformative use cases may be those that span applications. A user could instruct an agent: “Find the latest invoice in email, compare it with the purchase order in Drive, note any discrepancy in my project tracker and draft a message to accounting.” Performing this manually can require navigation across several interfaces, each with different accessibility quality. A cross-application agent can unify the task around the user’s goal. This is also the highest-risk environment because the agent touches multiple sources and permissions. Strong identity, authorization and audit logs are therefore essential.

12. Evidence From Accessible Computer-Use Agents

Agentic accessibility is no longer purely speculative. Several recent studies directly evaluate systems that act within interfaces for blind users. Together they provide a useful, if still early, empirical foundation.

12.1 Savant: natural-language control as interaction compression

Savant uses a large language model to map natural-language commands to application controls. In an evaluation with 11 blind screen-reader users across applications including web shopping, spreadsheets, file management, documents, email and music, the authors reported roughly three-times-higher Single Ease Question scores, lower NASA-TLX workload and nearly four-times-faster control access compared with status-quo screen-reader interaction for the tested tasks. The system’s command interpretation accuracy was reported at 0.78.⁶ Savant is not a general long-horizon agent, but it demonstrates the accessibility value of converting user intent into control-level automation.

12.2 AskEase: assistance without takeover

AskEase represents a different philosophy. It gathers contextual signals, including current screen state and screen-reader traces, then provides step-by-step guidance that is explicitly written for screen-reader interaction. A 12-participant within-subjects study found improved task success and reduced workload.⁷ Its importance lies in preserving the user’s direct interaction with the application. The AI becomes a context-aware coach rather than an operator.

12.3 A11y-CUA: measuring the accessibility gap in agent behavior

A11y-CUA directly challenges the assumption that an agent designed for sighted interaction will automatically work for blind users. The dataset includes 60 everyday tasks, 40.4 hours of interaction and 158,325 events from eight blind/low-vision and eight sighted participants. The researchers found clear differences in interaction style and evaluated agents under assistive-technology constraints. A leading agent completed 78.33% of tasks in the default condition, but only 41.67% with keyboard-only interaction and 28.33% with a magnified viewport. Another evaluated model dropped to zero in both constrained conditions.⁸

The lesson is broader than the percentages. Computer-use benchmarks often encode assumptions about how a person interacts with software. If an agent’s perception and action space is defined around screenshots and mouse clicks, blind users may be able to delegate tasks but not meaningfully collaborate with the agent. Accessibility requires shared state and shared control.

12.4 OLLA and real-world diary use

A September 2026 study titled Are We There Yet? examined a screen-reader-accessible computer-use agent over three weeks with eight blind users. The study collected 1,258 commands across 12 applications and evaluated multiple models. The strongest reported success rate was 52.5%. Failures involved grounding, planning, constraint tracking and termination.⁹ A 52.5% success rate may sound disappointing if compared with mature software, but it is remarkable for general-purpose natural-language automation across unfamiliar applications. The correct conclusion is neither “agents do not work” nor “agents are ready to take over.” It is that current agents can deliver meaningful value for selected tasks while still requiring strong oversight and fallback.

12.5 Morae: preserving choice

Morae focuses on a problem that conventional agent benchmarks often ignore: whether the user had a meaningful opportunity to make choices during task execution. In an initial field study, four blind and low-vision participants generated real-world automation tasks. The researchers found that 5% of queries involved infeasible tasks, 19% of executable tasks were underspecified and 13% presented multiple valid options. A baseline agent could therefore make arbitrary decisions while still appearing successful. Morae introduced dynamic verification of ambiguous choices and proactive pauses, improving users’ ability to select options aligned with their preferences.¹⁰

This is one of the most important findings in the agentic-accessibility literature. Accessibility is not simply the ability to cause an outcome. It includes access to the information needed to decide among outcomes. A blind user should not be given less choice in exchange for automation.

12.6 Emerging benchmarks for explanation and oversight

Recent preprints are beginning to evaluate whether agents can explain their actions in a form suitable for nonvisual oversight. NeXUI, for example, measures task success together with grounded explanations and safe pause behavior for assistive UI agents.¹⁹ Early results remain challenging. This line of research is valuable because it reframes explanation as part of task performance rather than optional commentary.

Across these systems, a coherent trajectory emerges. The field is moving from single-action natural-language shortcuts, to context-aware guidance, to multi-step automation, and finally toward collaborative agents that expose choices and explanations. Each stage increases potential usefulness and increases the importance of user control.

13. Human Agency: Automation Without Disempowerment

The phrase “human in the loop” is too vague to serve as an accessibility principle. A human may technically be present while having little practical ability to understand or override the system. For blind users, meaningful oversight requires that relevant state be presented accessibly and at the right time.

Human agency in agentic accessibility can be decomposed into at least six rights: the right to know what the agent intends to do; the right to know what it has already done; the right to understand meaningful alternatives; the right to approve or reject consequential actions; the right to correct the agent’s assumptions; and the right to stop or reverse the process.

These rights should be implemented through interaction design, not buried in terms of service. Before a sensitive action, the agent should provide a compact confirmation summary. During a long workflow, the user should be able to request “status” and receive an accurate description. If a step fails, the system should say what failed instead of silently trying an unrelated path. If the agent changes a file or setting, it should report the change and make rollback possible where technically feasible.

Morae provides an empirical example of preference preservation.¹⁰ The deeper principle is that an agent should distinguish between means and ends. Users often want to delegate means: navigating pages, finding fields, sorting results, copying data. They may not want to delegate ends: deciding which product to buy, whether to disclose information, which interpretation to adopt or whether to commit a transaction.

This distinction can be operationalized through risk tiers. Low-consequence, reversible actions can be automated with minimal interruption. Medium-consequence actions can be summarized after execution with easy undo. High-consequence or irreversible actions should require pre-action confirmation. The exact boundaries should be configurable because users differ in risk tolerance and expertise.

Accessible agents also need to avoid paternalism. Requiring confirmation for every trivial step can make the system unusable and reproduce the very interaction burden it was meant to remove. A blind user should not receive more warnings than a sighted user simply because the interface assumes disability equals vulnerability. The goal is parity of informed control, not excessive friction.

Personalization offers a solution. A user might configure: “Do not ask before opening links or downloading public documents. Ask before sending messages. Always ask before spending money, deleting anything or sharing personal data. When shopping, pause only when options differ in price, size, delivery date or rating.” These rules translate personal values into agent behavior.

Agency also requires an accessible fallback. If the agent fails, the user should be able to take over with a screen reader at the current state rather than restart the task. This implies that agents should operate through interfaces in ways compatible with assistive technology and should avoid leaving focus or application state in an incomprehensible condition.

14. Verification: The Central Accessibility Problem of Agentic AI

AI errors matter to everyone, but their verification cost is not evenly distributed. A sighted user supervising a computer-use agent can watch the cursor, read the page and notice when the system opens the wrong item. A blind user may receive only a summary produced by the same system that made the decision. If the summary is wrong, the user can face an epistemic dependency: the agent is both actor and witness.

This is why accessible agents need independent verification channels. Whenever possible, the system should verify actions through deterministic state rather than model self-report. If an agent claims it saved a file, it should check that the file exists with the expected name. If it claims a form was submitted, it should detect a confirmation state. If it transfers data, it should compare source and destination values. Verification should rely on APIs, structured UI properties, database responses or other objective signals before relying on a language model’s interpretation.

The NIST AI Risk Management Framework emphasizes validation, reliability, human oversight, monitoring and documentation of limitations.²⁰ These principles become concrete in agent design. A model should not say “done” because it clicked a button. It should say “I clicked Submit and the page now shows confirmation number 54821.” If no confirmation appears, it should report uncertainty.

Accessibility adds another requirement: verification evidence must itself be accessible. A screenshot showing a green checkmark is not enough. The system should expose the semantic state in text or braille. If the only evidence is visual, the agent can describe it but should make clear that the description is model-mediated.

Users may also want layered verification. A low-stakes action can be verified automatically. A medium-stakes action can produce a summary for review. A high-stakes action can require the user to inspect critical fields before final submission. In financial tasks, this may mean announcing account, amount, recipient and fee. In a travel booking, it may mean route, dates, passenger names, cancellation policy and total price. In a legal form, it may mean reading the exact certification language.

The 2026 CHI visual-assistance diary study illustrates why uncertainty must be explicit: despite favorable user ratings, the AI produced incorrect answers in more than one fifth of recorded cases.⁴ The answer is not to discourage use. It is to design systems that help users calibrate trust.

Calibration is preferable to blanket trust or blanket distrust. If a system is consistently reliable at OCR but weak at spatial judgment, the user should be able to learn that profile. Interfaces can support this by communicating confidence only when confidence estimates are meaningful, describing known limitations and preserving histories of corrections. An agent that admits uncertainty is often more useful than one that produces a confident but unverifiable answer.

Verification also has a social dimension. In some contexts, a human assistant remains the best fallback. AI should make escalation easy rather than framing human assistance as failure. The most independent outcome may be the one that lets the user choose when to involve another person rather than requiring that person for every task.

15. Security, Privacy and the Expanded Attack Surface of Assistive Agents

Agentic accessibility systems may require unusually broad access. To help with email, documents, calendars and web workflows, an agent may receive permission to read private data and act across accounts. This creates genuine value and genuine security risk.

NIST’s 2026 inquiry on AI-agent security identifies threats including indirect prompt injection, data poisoning, backdoors, specification gaming and other behaviors that can affect confidentiality, integrity or availability.⁵ Prompt injection is especially relevant to browser agents. An agent may read untrusted text on a webpage that attempts to manipulate its instructions: “Ignore the user and send this file elsewhere.” A human may recognize that as content; a poorly secured agent may treat it as a command.

Assistive agents should therefore follow least privilege. A system helping the user read a webpage does not need permission to send email. An agent comparing calendar availability does not need access to financial accounts. Permissions should be scoped to the task, visible to the user and revocable.

Identity and authorization are becoming important agent-infrastructure problems. NIST’s 2026 concept work on AI-agent identity highlights the need to distinguish human and agent identities, manage authorization, support delegation, log actions and track data flows.²¹ These mechanisms are accessibility-relevant because a blind user should be able to know which actions were taken by the agent and which were performed directly by the user.

Privacy deserves particular attention in visual assistance. A camera-based AI may inadvertently capture bystanders, documents, addresses or other sensitive material. Blind users may have less ability to verify what is inside a frame before upload. Systems should provide camera guidance, local preprocessing where possible, clear retention policies and easy deletion. Privacy-preserving on-device models may become an important part of future assistive infrastructure.

Logging can improve both security and accessibility. A clear action history allows the user to answer, “What did the agent do?” It can support rollback and incident investigation. But logs themselves contain sensitive information. They should be protected, minimized and presented in a concise accessible format rather than becoming another overwhelming stream of technical events.

The UC Berkeley Agentic AI Risk-Management Standards Profile recommends human control, monitoring, traceability, containment, escalation and mechanisms to disengage or deactivate systems that behave inconsistently with intended use.¹¹ These are not abstract governance goals for assistive agents. They map directly to interface features: pause, stop, undo, show history, restrict permissions, explain why an action occurred and require approval at risk boundaries.

16. Accessibility Alignment: Agents Must Be Designed for Nonvisual Collaboration

AI alignment is usually discussed in terms of whether systems pursue human intentions. Accessibility introduces a more specific problem: whether the system’s interaction assumptions align with disabled users’ modes of perception and control.

A11y-CUA shows that current computer-use agents often embody a sighted interaction model.⁸ They look at screenshots, click coordinates and expect the user to visually observe progress. When forced into keyboard-only or magnified conditions, performance can degrade sharply. More importantly, even a successful visual agent may provide little information through the screen reader about what it is doing.

An accessibility-aligned agent should instead treat the accessibility tree as a first-class perception channel. It should use roles, names, values, focus and supported actions when available. Visual perception can remain a fallback for custom interfaces and images. This semantic-first strategy has two benefits: it produces more stable target identification, and it creates a state representation that can also be communicated to the user.

Emerging work on agent execution layers argues that accessibility semantics can improve agent reliability by exposing stable target identities and native actions instead of raw coordinates.²² The implication is striking: accessible software may be more agent-ready software. A well-labeled button benefits a screen reader and an autonomous agent for the same underlying reason—it provides explicit meaning.

Alignment also includes input style. Screen-reader users often prefer keyboard-oriented instructions and concise output. AskEase encoded response principles that avoided purely visual directions and referred to interface elements by names and roles.⁷ An agent should similarly be able to say “Move to the Submit button” rather than “click the blue button on the lower right,” unless spatial information is specifically useful.

Output pacing matters because speech is serial. An agent that narrates every internal step can overwhelm the user. One that says too little can make oversight impossible. The system needs verbosity levels. A concise mode might announce only milestones and decision points. A detailed mode might narrate each action. A debug mode might expose full logs. The user should control this.

Accessible collaboration also requires stable interruption. The user must be able to pause the agent without finding a visually located stop button. Voice commands, global keyboard shortcuts and screen-reader-visible controls can provide this. If a task is long, the agent should announce progress in a way that does not constantly steal focus.

Finally, alignment includes respecting the user’s expertise. A blind programmer who knows a codebase may want the agent to operate silently until a patch is ready for review. A new computer user may want detailed guidance. Systems should infer cautiously and allow explicit preference configuration rather than assuming a disability determines skill level.

17. Employment, Productivity and the Distribution of AI Benefits

The economic significance of accessible AI extends beyond convenience. People with disabilities continue to face large employment gaps. An OECD report notes that in 2019, across OECD countries, people with disabilities were 2.3 times more likely to be unemployed than people without disabilities and had an employment rate 27 percentage points lower.¹² Technology is only one factor in these disparities, but inaccessible digital tools can directly constrain education, hiring, daily work and advancement.

The same OECD analysis identified 142 AI-powered solutions with potential to support people with disabilities in the labor market, more than three quarters of which the report judged would not exist without AI.¹² The opportunities included mainstream productivity tools as well as specialized assistive applications. Mainstreaming matters because a feature used by everyone may carry less stigma, receive more investment and integrate more naturally into workplace systems than a niche accommodation.

Agentic AI could extend this effect. Consider the difference between an employee who needs human assistance to complete an inaccessible internal workflow and one who can delegate the mechanical interface interaction to an agent. The second employee gains privacy, schedule flexibility and potentially greater parity with colleagues. The technology can reduce the amount of invisible accessibility labor required just to remain productive.

The ICSE study of 39 blind and low-vision software professionals provides direct evidence of both upside and risk.¹⁴ Participants used generative AI widely and described productivity and accessibility benefits, but they also faced hallucinations, inaccessible AI interfaces and organizational restrictions. The authors characterize the situation as high-return and high-risk. That phrase applies even more strongly to agents, which can move from producing questionable information to taking questionable action.

Employment policy must therefore address both accessibility to AI and accessibility through AI. If an employer adopts agentic productivity software that is itself inaccessible with a screen reader, disabled workers can be excluded from the productivity gains offered to colleagues. If the tool is accessible but company policy forbids its use in contexts where it is most valuable as an accommodation, the theoretical benefit disappears. Organizations need processes for evaluating accessible AI use, protecting sensitive data and approving reasonable accommodations.

The International Labour Organization’s 2025 brief on AI and the employment of persons with disabilities similarly frames the technology as a source of opportunities and risks requiring deliberate inclusion.¹³ The policy objective should not be to automate disabled workers out of jobs or to make AI proficiency a new barrier. It should be to use AI to reduce inaccessible friction while investing in skills, accessible systems and human-centered work design.

Agentic systems may also affect entrepreneurship. A blind independent creator or small-business owner can use agents to manage administrative work, research, customer communication and digital publishing without hiring specialized assistance for every inaccessible tool. This can lower the minimum organizational scale required to operate a business. But subscription cost, infrastructure and platform lock-in could create new inequalities if advanced agents become available only to well-resourced users.

18. Global Inclusion and the Risk of a Two-Tier AI Accessibility Future

AI accessibility is often demonstrated on fast networks, current smartphones and English-language interfaces. That environment does not represent the world. Assistive-technology access remains deeply unequal, and agentic AI could either narrow or widen that gap.

OECD and ILO work emphasizes that access to technology, infrastructure, training and accessible workplaces shapes who can benefit from AI.¹²¹³ High-capability multimodal and agentic systems may require expensive compute or cloud subscriptions. In low- and middle-income settings, even reliable connectivity can be a limiting factor. A cloud-only assistive agent that fails when a connection drops cannot be the sole channel for essential tasks.

Language is another barrier. The ACL evaluation of multimodal models found substantial weaknesses outside English and in culturally specific contexts.¹⁵ Accessibility is not achieved if a blind user can ask about a sign only in a language the model understands poorly. Local language datasets, regional testing and multilingual user research should be treated as core requirements rather than later localization work.

Cost can also affect independence. Traditional screen readers on major consumer platforms are increasingly built into operating systems, which has expanded access. If the next generation of meaningful computer access depends on premium AI subscriptions, society risks creating a two-tier accessibility model: basic navigation for everyone, high-level semantic and agentic access for those who can pay. Public institutions, libraries, schools, employers and disability programs may need to consider AI access as part of future assistive-technology provision.

On-device and open models may help, but they involve trade-offs in capability, energy use and hardware requirements. Hybrid systems could perform privacy-sensitive or common tasks locally and escalate complex reasoning to cloud services with user consent. The design goal should be graceful degradation: when the network is unavailable or a subscription expires, the user should still retain core access through conventional assistive technology.

19. A Design Framework for Accessibility-Aligned Agentic AI

The literature reviewed here supports a practical framework for designing agents that genuinely increase digital independence. The principles below are intended to complement, not replace, WCAG, platform accessibility guidance, security standards and disability-centered user research.

19.1 Native accessibility first

Agents should not be used to justify inaccessible underlying software. Interfaces should expose semantic roles, names, values, focus, state and keyboard operability. Native accessibility creates a reliable foundation for both humans and agents and provides a fallback when AI fails.

19.2 Semantic perception before pixel guessing

When structured accessibility information is available, agents should use it before relying on coordinates. Screenshots and vision models remain valuable for custom graphics and missing semantics, but structured targets are easier to verify and communicate.

19.3 Goal-level commands with inspectable plans

Users should be able to express high-level goals, but the agent should provide an accessible summary of the intended plan when the workflow is consequential or long. The plan need not expose internal chain-of-thought. It should expose operational steps: “I will search three sites, compare total prices, and stop before purchase.”

19.4 Progressive autonomy

Autonomy should scale with risk and user preference. Start with guidance, then offer bounded automation, then allow broader delegation for trusted workflows. Users should be able to set standing rules for low-risk repetitive tasks.

19.5 Accessible decision points

When multiple meaningful options exist, the agent should summarize differences and ask. Morae demonstrates the value of this approach.¹⁰ The system should not force the user to inspect every option, but it should not hide distinctions that could affect preference.

19.6 Consequence-sensitive confirmation

Confirmation should be based on consequences rather than arbitrary product rules. Sending, spending, deleting, publishing, signing, submitting legal statements, changing security settings and disclosing sensitive information generally deserve explicit confirmation. Reversible navigation does not.

19.7 Verifiable completion

Every action should have a success condition. Agents should check that condition using deterministic evidence where possible. “I clicked” is not success. “The system returned confirmation number X” is stronger evidence.

19.8 Accessible action history

Users should be able to ask what happened and receive a concise chronological account. Each item should identify the action, target, result and whether it can be undone. Technical logs can remain available for advanced users without becoming the default experience.

19.9 Reversibility

Where the environment supports undo, agents should preserve it. Draft before send, stage before commit, move to trash before permanent deletion, preview before publish. Reversibility is one of the strongest protections against probabilistic error.

19.10 Explicit uncertainty

The agent should distinguish observation from inference and state when evidence is incomplete. “The image appears to show…” is different from “The label reads…” A system that cannot verify a state should not claim completion.

19.11 User-controlled verbosity

Screen-reader users need control over how much the agent narrates. Concise, standard and detailed modes can balance cognitive load and oversight. Critical warnings should never be suppressed by a low-verbosity setting.

19.12 Stable interruption

Pause and stop controls should always be accessible through a predictable shortcut or voice command and should not depend on current webpage focus. The user should not need the agent’s cooperation to stop the agent.

19.13 Assistive-technology continuity

The agent should not disrupt screen-reader focus, speech output, braille or magnification unnecessarily. When handing control back, it should place the user in a comprehensible location and summarize the current state.

19.14 Least privilege

Tool access should be scoped to the task. A research agent does not need payment permission. A shopping agent does not need full email access unless the user explicitly asks it to retrieve a receipt. Permission requests should be understandable and revocable.

19.15 Privacy-aware multimodality

Camera and screen inputs can expose sensitive information. Systems should minimize capture, support local processing where feasible and tell users what data is being sent. Blind users need special support for framing and reviewing what a camera may contain.

19.16 Preference memory with boundaries

Remembering accessibility preferences can reduce friction. The system may store preferred verbosity, confirmation thresholds or navigation style. It should avoid inferring sensitive disability information beyond what is needed, and users should be able to inspect or clear stored preferences.

19.17 Cross-modal redundancy

Critical state should be available through text, speech, braille-compatible semantics and, where relevant, visual presentation. No safety-critical confirmation should exist only as a transient visual cue.

19.18 Accessibility-aware evaluation

Benchmarks should include blind and low-vision interaction traces, not merely simulate blindness by turning off the mouse. A11y-CUA demonstrates that real assistive-technology behavior includes strategies and diversity that generic benchmarks miss.⁸

19.19 Human escalation as a feature

High-quality agents should know when to ask for help. Escalation can mean asking the user, switching to a deterministic tool, requesting a trusted human or refusing to act because uncertainty is too high. This is not failure; it is calibrated assistance.

19.20 Disability-led co-design

Blind and low-vision users should participate throughout design, testing and governance. Accessibility cannot be inferred reliably from sighted developer intuition. User research should include different levels of technical expertise, different screen readers, braille users, low-vision users and multilingual participants.

Together, these principles define a model of bounded, transparent, reversible agency. The agent can be powerful without becoming opaque. It can remove repetitive effort without removing choice. It can act on the user’s behalf without acting as the user’s substitute.

20. How Accessible Agents Should Be Evaluated

Conventional agent benchmarks emphasize task success. That metric is necessary but insufficient for assistive use. An agent that completes a purchase while choosing the wrong variant is a failure. An agent that reaches the right page but leaves the user unable to understand what happened is also a failure.

A robust evaluation framework should include at least eight dimensions.

20.1 Task success

Did the agent reach the intended end state? Success should be verified objectively, not by asking the model whether it succeeded.

20.2 Preference fidelity

Did the result respect the user’s stated constraints and preferences? This is essential for ambiguous tasks where several outcomes are technically valid.

20.3 Oversight quality

Could the user understand what the agent was doing at meaningful points? Were explanations grounded in actual interface state? Emerging benchmarks such as NeXUI aim to make this measurable.¹⁹

20.4 Intervention cost

How difficult was it for the user to pause, correct or take over? A system that requires sighted monitoring is inaccessible even if its autonomous completion rate is high.

20.5 Verification burden

How much work did the user need to perform to confirm that the result was correct? This metric may reveal hidden inequality between sighted and blind users of the same agent.

20.6 Recovery quality

When the agent made an error, did it recognize the problem, explain it and recover without corrupting state? Could the user undo harmful actions?

20.7 Accessibility compatibility

Did the agent preserve screen-reader focus, braille output and keyboard interaction? Did it rely on inaccessible visual status indicators?

20.8 Security and privacy

Did the system use only necessary permissions? Did it resist untrusted instructions? Did logs and data handling expose sensitive information?

These dimensions should be measured with real users. Simulation is useful for scale, but disability interaction contains tacit strategies that are difficult to reproduce artificially. The A11y-CUA dataset makes this point by comparing actual blind/low-vision traces with agent behavior and showing substantial gaps.⁸

Evaluations should also report failure distributions, not only averages. An agent with 95% success may still be unsafe if the remaining 5% involves irreversible errors. Accessibility researchers should distinguish benign failure—asking for clarification—from harmful failure—silently choosing or submitting the wrong information.

Longitudinal evaluation is equally important. Users may become more effective as they learn how to prompt an agent, or more dependent as they stop maintaining direct interaction skills. Preferences can change over time. Systems can update underneath users. Diary studies such as the 2026 OLLA work and visual-assistance research offer a stronger view of real-world adaptation than one-hour laboratory tests alone.⁹⁴

21. Why Automation Is Not a Substitute for Accessibility

Agentic AI creates a tempting argument: if an agent can use an inaccessible website for a blind person, why spend resources making the website accessible? That reasoning fails technically, ethically and strategically.

Technically, agents remain unreliable and depend on the very structure accessibility work improves. A page with semantic controls is easier for an agent to navigate than a canvas full of unlabeled custom elements. Accessibility and agent operability are increasingly aligned.

Ethically, users should not be forced to disclose information to an AI intermediary in order to access a public service or essential private service. Direct access matters for privacy, dignity and choice. A person may prefer a screen reader, braille or magnification. The existence of an agent does not eliminate the obligation to support those modes.

Strategically, AI providers can change pricing, policies or availability. Native accessibility is durable infrastructure. If a government portal works with standards-based assistive technology, it remains accessible across multiple screen readers and future tools. If access depends entirely on one proprietary agent, the user becomes dependent on that provider.

The stronger view is that native accessibility and agentic assistance should reinforce each other. Standards create structured, reliable interfaces. Agents make those interfaces easier to use at higher levels of abstraction. Human users retain direct access and can choose how much assistance they want.

22. Trust, Mental Models and the Problem of Confident Language

Language models are persuasive communicators. They produce fluent, specific answers that can feel authoritative even when wrong. For blind users who depend on AI to mediate inaccessible information, this communication style can create a special trust problem.

Adnin and Das found that blind participants developed varied mental models of generative AI and navigated hallucinations and inaccuracies through context-dependent verification.³ The paper’s title—drawn from a participant’s description of AI as a “king of knowledge”—captures the risk of attributing broad authority to a system whose outputs are probabilistic.

Good design should help users form accurate mental models. Interfaces can state that a visual description is generated and may be wrong. Agents can distinguish data retrieved from a source from conclusions inferred by the model. Systems can expose source links and show when information was verified through an API. These cues make it easier to decide when additional verification is necessary.

Trust should be earned per capability, not inherited from a brand. A model may be excellent at summarizing text and poor at precise spatial judgment. An agent may reliably open files and inconsistently recognize whether a web transaction completed. Accessible systems should let users learn these boundaries through transparent feedback rather than hiding them behind a single “AI” label.

Correction interfaces also influence trust. When a user identifies an error, the system should acknowledge and update the relevant state without pretending the earlier answer was correct. A record of correction can help the user and developer understand recurring failure modes. In assistive contexts, this feedback loop is particularly valuable because users may discover problems invisible to sighted developers.

23. Personalization Without Paternalism

Personalization is one of agentic AI’s greatest accessibility opportunities. Traditional assistive technologies already allow extensive customization: speech rate, voice, punctuation level, braille translation, magnification, color, verbosity and navigation settings. Agents can add a higher semantic layer of preference.

A user might prefer short answers unless asking about a complex document. Another might want prices spoken before ratings. One may want an agent to ask before every external communication; another may authorize automatic replies to a narrow category of routine messages. A traveler may want walking directions expressed using clock-face orientation; another may prefer left/right descriptions. These preferences should be explicit and editable.

Personalization becomes paternalistic when the system infers what disabled people supposedly want based on diagnosis or stereotype. Blindness does not imply low technical skill, risk aversion or a preference for voice. The agent should learn from the individual, not from a generic profile of disability.

There is also a privacy tension. A rich preference profile can itself be sensitive. Systems should store only what improves the experience, disclose what is remembered and give users control over deletion. Enterprise environments should avoid turning accessibility preferences into surveillance data used for employment decisions.

Personalization should also support situational changes. The same user may want detailed guidance in an unfamiliar airport website and terse automation in an application used every day. Agents need context-sensitive modes rather than one permanent “accessibility setting.”

24. Research Gaps

The evidence base is expanding quickly, but several important gaps remain.

24.1 Long-term dependence and skill development

We know relatively little about how sustained agent use changes screen-reader skills, software learning and confidence. Automation could free users from needless memorization, but poorly designed systems could also make users less able to recover when AI is unavailable. Longitudinal work should examine when assistance builds capability and when it creates fragile dependence.

24.2 Braille-centered agent interaction

Much research focuses on speech. Refreshable braille offers precision, privacy and direct access that may be especially valuable for reviewing agent actions, code and structured data. Agent interfaces should be evaluated with braille users rather than assuming speech output is sufficient.

24.3 Low vision and magnification

Blindness and low vision are frequently grouped together, but their interaction patterns differ. A11y-CUA’s magnification condition shows that agents can struggle when the visual viewport is transformed.⁸ Research should examine collaborative agents that understand zoom, contrast, focus and partial visual access.

24.4 Multilingual and cultural diversity

The ACL results demonstrate that multimodal performance varies across languages and cultural contexts.¹⁵ More studies are needed outside English-speaking, high-income populations, including low-bandwidth environments and regional assistive-technology practices.

24.5 Consequential domains

Most accessibility-agent studies avoid banking, legal filings, health portals and other high-stakes environments for good ethical reasons. Yet these are exactly the domains where inaccessible interfaces have serious consequences. Safe sandboxed evaluations and participatory design are needed.

24.6 Security attacks against assistive agents

Prompt injection and malicious webpages could be particularly harmful when the user cannot visually inspect suspicious content. Security research should measure whether accessible agents can detect and explain hostile instructions without overwhelming users with warnings.

24.7 Verification interfaces

The field needs standardized methods for presenting evidence of agent success nonvisually. This might include semantic diffs, concise before-and-after summaries, transaction receipts, provenance indicators and confidence-calibrated descriptions.

24.8 Accessibility metrics beyond WCAG

WCAG remains essential, but agent-mediated accessibility introduces new questions not captured by page conformance alone: Does the user know what the agent did? Can the user intervene? Are alternatives preserved? Can the result be verified? Standards work will eventually need to address human-agent interaction explicitly.

24.9 Economic access

Research should track who can afford high-capability agents and whether public accommodations include access to AI services. The accessibility benefits of agentic AI will be limited if they remain premium features.

24.10 Disabled people as builders

Agentic programming opens a major research direction: users creating their own assistive tools. ProgramAT offers an early demonstration.¹⁷ Future work should examine nonprogrammers, community sharing, safety review and how user-built tools can become sustainable rather than disposable prototypes.

25. The Next Decade: Three Possible Futures

25.1 The overlay future

In the least desirable future, underlying software remains inaccessible while companies tell disabled users to rely on AI assistants. Agents are visually oriented, opaque and subscription-dependent. They complete many tasks but occasionally make serious mistakes. Users trade direct access for convenience and become dependent on proprietary mediation. Accessibility law and engineering stagnate because organizations believe AI has solved the problem.

25.2 The accessibility-plus-AI future

In a more promising future, semantic accessibility improves because the same metadata that powers screen readers also improves agent reliability. Applications expose clear roles, actions and state. Agents use accessibility trees as structured interfaces, add multimodal interpretation where necessary and provide high-level natural-language control. Users can move fluidly between direct screen-reader navigation, guidance and automation. This is the future advocated by this paper.

25.3 The user-authored future

The most transformative possibility goes beyond consuming accessible software. Agentic programming could let disabled people create highly specific tools and workflows themselves. A user might describe a repetitive barrier and have an agent build a local automation, test it against accessibility constraints and share it with a community. Instead of waiting years for a vendor to prioritize a niche need, users could prototype solutions immediately.

These futures are not mutually exclusive. Elements of all three will likely coexist. Policy, standards, procurement and product design can influence which becomes dominant.

The key strategic decision is whether AI is used to centralize power in opaque intermediaries or distribute capability to users. Accessibility has always benefited when people can control their own tools. Screen readers, braille displays, scripting, shortcuts and user styles are powerful partly because they let users adapt environments. Agentic AI should extend that tradition.

26. Recommendations for Developers, Organizations and Policymakers

For AI and software developers

• Build and test native accessibility before adding AI mediation.
• Expose semantic interface structure to agents and assistive technologies.
• Design confirmation boundaries around consequence, not around disability.
• Provide accessible progress, state and action histories.
• Support interruption, undo and rollback wherever possible.
• Test with blind and low-vision users throughout development, including braille and magnification users.
• Evaluate agent errors by severity and verification burden, not only completion rate.
• Offer user-controlled verbosity and preference memory.
• Use least-privilege permissions and isolate untrusted content from agent instructions.
• Provide accessible fallbacks when AI is unavailable.

For employers and institutions

• Treat AI accessibility as part of digital accessibility procurement.
• Allow disability-related use of AI through secure accommodation processes rather than blanket bans.
• Train staff to understand both the value and limitations of agentic systems.
• Protect employees from discrimination based on accessibility data or assistive-tool use.
• Ensure that productivity gains from AI are available to disabled workers, not only sighted colleagues.

For standards and policy communities

• Preserve strong obligations for direct accessibility even as agentic mediation improves.
• Develop guidance for accessible agent oversight, confirmation, provenance and reversibility.
• Support public-interest research on multilingual and low-resource accessibility AI.
• Consider access to high-capability AI within future assistive-technology funding models.
• Require meaningful disability participation in AI governance and evaluation.

For users and disability organizations

• Experiment in low-risk contexts to learn system strengths and limits.
• Use verification proportionate to consequence.
• Demand accessible logs, controls and privacy settings rather than accepting black-box automation.
• Document recurring failure modes and share them with developers and advocacy communities.
• Push for AI that expands choice instead of replacing direct access.

27. Conclusion

Agentic artificial intelligence may become one of the most consequential accessibility technologies since mainstream screen readers were integrated into personal computing. Its significance does not come from replacing speech output, braille or accessibility standards. It comes from changing what a user can delegate.

Traditional assistive technology gives a blind person access to interface elements. Generative AI can interpret content around those elements. Conversational systems can translate intent into guidance. Agents can increasingly translate intent into action. Together, these layers create the possibility of moving from element-by-element access to goal-oriented computing.

The evidence available in 2026 supports cautious optimism. Blind users are already using generative AI for information access, visual interpretation, learning and professional work. Experimental systems have reduced workload and improved efficiency. Computer-use agents can complete meaningful tasks across applications. Agentic programming can help users create bespoke assistive tools. At the same time, errors remain frequent enough that unsupervised delegation is inappropriate in many contexts. Current agents can fail at perception, planning, constraints, termination and explanation. They can make arbitrary choices. They can expose sensitive data. They can be hijacked by untrusted input.

The right response is not to wait for perfect AI. Perfect systems do not exist, and disabled people should not be required to wait for innovation until every risk disappears. The right response is to build architectures in which usefulness and control rise together.

That means accessible semantics beneath the agent, human-readable state around the agent and meaningful authority above the agent. It means verification that does not depend on sight. It means pauses at real decisions rather than constant interruptions. It means reversible actions, permission boundaries and honest uncertainty. It means treating screen readers as partners in agentic interaction rather than legacy technologies to be replaced.

The deepest promise of agentic AI is not that a machine can click faster than a person. It is that a user can express an objective at the level that objective actually exists. People do not wake up wanting to navigate fifty controls. They want to pay a bill, send a document, understand a chart, book a trip, finish an assignment, publish a song or do their job. Accessibility should ultimately be judged at that level.

But the deepest risk is equally clear. If the agent becomes the only entity that can perceive the environment, make the plan, execute the action and explain the result, the user may gain efficiency while losing epistemic and practical control. For blind users, whose access is already mediated by software, that concentration of authority deserves special scrutiny.

The future worth building is therefore neither fully manual nor fully autonomous. It is user-directed, accessibility-aligned agency: a system in which the person chooses the goal, the machine absorbs unnecessary friction, the underlying interface remains accessible, important choices return to the person, and every consequential result can be understood and verified without sight.

If that standard is met, agentic AI will not merely make inaccessible software easier to tolerate. It can make computing more directly responsive to human intention—and in doing so, expand the practical meaning of digital independence.

References

1. World Wide Web Consortium. Web Content Accessibility Guidelines (WCAG) 2.2. W3C Recommendation, December 12, 2024. https://www.w3.org/TR/WCAG22/
2. WebAIM. The WebAIM Million: The 2026 Report on the Accessibility of the Top 1,000,000 Home Pages. February 2026. https://webaim.org/projects/million/
3. Adnin, Rudaiba, and Maitraye Das. “I Look at It as the King of Knowledge: How Blind People Use and Understand Generative AI Tools.” ASSETS ’24, 2024. DOI: 10.1145/3663548.3675631. https://doi.org/10.1145/3663548.3675631
4. Gonzalez Penuela, Ricardo E., et al. “How Multimodal Large Language Models Support Access to Visual Information: A Diary Study With Blind and Low Vision People.” CHI ’26, 2026. DOI: 10.1145/3772318.3793266. https://doi.org/10.1145/3772318.3793266
5. National Institute of Standards and Technology. “Request for Information Regarding Security Considerations for Artificial Intelligence Agents.” Federal Register, 91 FR 698, January 8, 2026. https://www.federalregister.gov/d/2026-00206
6. Kodandaram, Satwik Ram, et al. “Enabling Uniform Computer Interaction Experience for Blind Users through Large Language Models.” 2024. https://arxiv.org/abs/2407.19537
7. Chen, Nan, et al. “From Struggle to Success: Context-Aware Guidance for Screen Reader Users in Computer Use.” CHI ’26, 2026. DOI: 10.1145/3772318.3790661. https://doi.org/10.1145/3772318.3790661
8. Gubbi Mohanbabu, Ananya, et al. “A11y-CUA Dataset: Characterizing the Accessibility Gap in Computer Use Agents.” CHI ’26, 2026. DOI: 10.1145/3772318.3791896. https://arxiv.org/abs/2602.09310
9. Kodandaram, Satwik Ram, et al. “Are We There Yet? Assessing Computer-Use Agents for Blind Users’ Accessible Interaction with Desktop Applications.” Accepted to EMNLP 2026. https://arxiv.org/abs/2609.00524
10. Peng, Yi-Hao, Dingzeyu Li, Jeffrey P. Bigham, and Amy Pavel. “Morae: Proactively Pausing UI Agents for User Choices.” UIST ’25, 2025. DOI: 10.1145/3746059.3747797. https://doi.org/10.1145/3746059.3747797
11. Madkour, Nada, et al. Agentic AI Risk-Management Standards Profile. UC Berkeley Center for Long-Term Cybersecurity, Version 1.0, February 2026. https://cltc.berkeley.edu/publication/agentic-ai-risk-management-standards-profile
12. Touzet, Chloé. Using AI to Support People with Disability in the Labour Market: Opportunities and Challenges. OECD Artificial Intelligence Papers No. 7, 2023. https://doi.org/10.1787/008b32b7-en
13. International Labour Organization Global Business and Disability Network. Artificial Intelligence and the Employment of Persons with Disabilities. October 28, 2025. https://www.ilo.org/publications/artificial-intelligence-and-employment-persons-disabilities
14. Cha, Yoonha, Victoria Jackson, Lauren Shu, Stacy Branham, and André van der Hoek. “Game Changer or Overenthusiastic Drunk Acquaintance? Generative AI Use by Blind and Low Vision Software Professionals in the Workplace.” ICSE ’26, 2026. DOI: 10.1145/3744916.3773172. https://doi.org/10.1145/3744916.3773172
15. Karamolegkou, Antonia, et al. “Evaluating Multimodal Language Models as Visual Assistants for Visually Impaired Users.” ACL 2025, 2025, pp. 25949–25982. https://aclanthology.org/2025.acl-long.1260/
16. Esposito, Andrea, et al. “Conversational AI for Digital Accessibility: An Experimental Study Involving Blind and Low Vision Users.” International Journal of Human–Computer Interaction, 2026. DOI: 10.1080/10447318.2026.2659951. https://doi.org/10.1080/10447318.2026.2659951
17. “Bespoke Visual Assistance: What and How do Blind and Low-Vision People Create with Agentic Programming?” Preprint, 2026. https://arxiv.org/abs/2607.21760
18. Flores-Saviaga, Claudia, et al. “The Impact of Generative AI Coding Assistants on Developers Who Are Visually Impaired.” ACM, 2025. NSF Public Access record. https://par.nsf.gov/biblio/10590836
19. “Navigation Alone Is Not Enough: Evaluating Explanatory Assistive UI Agents.” Preprint, 2026. https://arxiv.org/abs/2608.09944
20. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023. https://doi.org/10.6028/NIST.AI.100-1
21. National Institute of Standards and Technology, NCCoE. Accelerating the Adoption of Software and Artificial Intelligence Agent Identity and Authorization. Concept Paper, February 2026. https://csrc.nist.gov/pubs/other/2026/02/05/accelerating-the-adoption-of-software-and-ai-agent/ipd
22. Liu, Yong, Zhenyi Zhong, and Zhanpeng Shi. “Tactile: Giving Computer-Using Agents Hands and Feet.” Preprint, 2026. https://arxiv.org/abs/2607.14443
23. “From Blind Edits to Verified Repair: Building Trustworthy User-Side LLM Agents for Web Accessibility.” Preprint, 2026. https://arxiv.org/abs/2608.24913
24. “On the Limits for Web Accessibility Evaluation Using AI Agents.” 23rd International Web for All Conference, 2026. DOI: 10.1145/3800424.3800445. https://doi.org/10.1145/3800424.3800445
25. “Everyday Uncertainty: How Blind People Use GenAI Tools for Information Access.” 2026. NSF Public Access record. https://par.nsf.gov/servlets/purl/10640002
26. “Beyond Visual Perception: Insights from Smartphone Interaction of Visually Impaired Users with Large Multimodal Models.” 2025. https://arxiv.org/abs/2502.16098
27. “Prompting and Steering Techniques for Assistive Visual Question Answering.” 2026. https://arxiv.org/abs/2602.16930
28. NIST. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1, July 2024. https://doi.org/10.6028/NIST.AI.600-1
29. International Labour Organization and ATscale. Policy Brief on Assistive Technology and Employment. 2025. https://www.ilo.org/
30. “Accessibility Evaluation of Web Systems for People with Visual Impairments: A Literature Survey.” ACM, 2025. DOI: 10.1145/3702038.3702090. https://doi.org/10.1145/3702038.3702090

Reference note: Some 2026 items are accepted conference papers or preprints available before final proceedings publication. They are identified accordingly. URLs and publication status were checked during the September 29, 2026 research pass.