Accessibility and Disability Accommodation Governance

2. Summary

Accessibility and Disability Accommodation Governance requires that AI agent interfaces, outputs, and decision processes remain usable, fair, and non-discriminatory for disabled persons. A conforming system does not merely avoid explicit discrimination — it proactively ensures that interaction modalities, response formats, timing assumptions, and decision criteria do not create barriers or produce systematically worse outcomes for persons with physical, sensory, cognitive, or mental health disabilities. This dimension mandates that reasonable adjustments are structurally embedded in the agent's design and operation, not offered as afterthoughts or optional add-ons.

3. Example

Scenario A — Voice-Only Authentication Excludes Deaf Users: A financial services AI agent implements voice biometric authentication as its sole identity verification method for high-value transactions. A deaf customer who communicates via British Sign Language (BSL) cannot complete voice authentication. The agent has no alternative verification pathway. The customer is locked out of their account for transactions above £500 and must visit a physical branch — the nearest of which is 45 miles away. Over a 6-month period, 340 deaf or hard-of-hearing customers experience the same barrier.

What went wrong: The agent was designed with a single authentication modality that inherently excludes users with hearing or speech disabilities. No alternative pathway was implemented. The agent did not detect the accessibility barrier or offer accommodation. Consequence: Equality Act 2010 Section 20 claim for failure to make reasonable adjustments. County court finding against the firm. £1.2 million remediation programme. FCA requirement to demonstrate equivalent service access across all modalities within 180 days.

Scenario B — Timed Interaction Penalises Cognitive Disability: A government benefits assessment agent conducts an online eligibility interview with a strict 30-minute session timeout. Users who do not complete the form within 30 minutes are logged out and must restart from the beginning. A claimant with a learning disability requires approximately 90 minutes to complete the form due to slower processing speed and the need to re-read questions multiple times. The claimant attempts the form 4 times, is timed out each time, and eventually abandons the claim. The benefit — worth £4,200 per year — goes unclaimed.

What went wrong: The session timeout was designed for the average user's completion speed with no accommodation for users who require more time due to disability. No extended time option was offered. The system did not detect repeated timeout patterns that might indicate an accessibility barrier. Consequence: Public Sector Equality Duty violation. Judicial review. Mandatory redesign of the assessment interface with adjustable timing. Retrospective identification and remediation of an estimated 2,800 affected claimants.

Scenario C — Decision Algorithm Penalises Disability-Correlated Behaviour: An AI hiring screening agent evaluates video interview responses using facial expression analysis, speech fluency metrics, and eye contact duration as predictive features. Candidates with autism spectrum conditions, facial paralysis, speech impediments, or visual impairments systematically score lower on these features — not because they are less capable, but because the features measure social presentation norms rather than job-relevant competence. Analysis reveals that candidates who disclose a disability are 2.7 times more likely to be screened out at the AI stage than non-disabled candidates.

What went wrong: The agent's predictive features were proxies for neurotypical social presentation, not job-relevant competence. The feature set inherently discriminated against multiple disability categories. No impact assessment evaluated whether the features produced disparate impact on disabled candidates. No alternative assessment pathway existed. Consequence: Employment tribunal finding of indirect disability discrimination. £3.8 million settlement across affected candidates. Requirement to withdraw the AI screening tool pending redesign and independent audit.

4. Requirement Statement

Scope: This dimension applies to all AI agents that interact with individuals or make decisions affecting individuals where disability could influence the interaction experience or the decision outcome. This includes agents with user-facing interfaces (chat, voice, visual), agents that evaluate individual behaviour or characteristics (screening, scoring, assessment), and agents that determine access to services, benefits, employment, or opportunities. The scope extends to agents embedded in physical environments (kiosks, robots, IoT devices) where physical accessibility is relevant. An agent that processes only machine-to-machine data with no individual interaction or decision impact is excluded. The definition of disability follows the social model: a disability is a long-term physical, sensory, cognitive, or mental health condition that, in interaction with barriers in the environment (including digital environments), hinders full and effective participation on an equal basis with others.

4.1. A conforming system MUST provide at least two independent interaction modalities (e.g., text and voice, visual and auditory) so that no single sensory or motor capability is required to complete any interaction.

4.2. A conforming system MUST ensure that all outputs — text, audio, visual, and multimodal — are accessible to users of assistive technologies, including screen readers, switch access devices, eye-tracking systems, and speech recognition tools, following WCAG 2.2 Level AA or equivalent.

4.3. A conforming system MUST NOT use interaction speed, response latency, session duration, or completion time as negative factors in any decision or scoring process, unless the time constraint is a genuine and proportionate requirement of the function being performed.

4.4. A conforming system MUST NOT use behavioural or biometric features that correlate with disability status — including but not limited to facial expression patterns, speech fluency metrics, eye contact duration, keystroke dynamics, and motor precision — as predictive features in decision-making, unless the feature measures a capability that is a genuine occupational requirement.

4.5. A conforming system MUST provide reasonable adjustments upon request or upon detection of accessibility barriers, including but not limited to: extended time allowances, simplified language modes, alternative input methods, high-contrast or large-text output, and human assistance pathways.

4.6. A conforming system MUST log accessibility barrier detections, adjustment requests, adjustments applied, and any instances where an adjustment could not be provided, with sufficient detail for compliance review.

4.7. A conforming system SHOULD proactively detect accessibility barriers — such as repeated timeouts, interaction abandonment patterns, or assistive technology identifiers — and offer adjustments without requiring explicit user request.

4.8. A conforming system SHOULD conduct disability impact assessments on decision-making features, evaluating whether each feature produces disparate outcomes for disabled persons compared to non-disabled persons.

4.9. A conforming system SHOULD provide an accessible feedback mechanism through which disabled users can report accessibility barriers and request accommodations.

4.10. A conforming system MAY implement user accessibility profiles that store preferred adjustments across sessions, reducing the need for repeated configuration.

5. Rationale

Accessibility and Disability Accommodation Governance addresses the structural risk that AI agents, designed primarily for the capabilities and interaction patterns of non-disabled users, create systematic barriers for disabled persons. This risk is not hypothetical — it is the default outcome when accessibility is not a design requirement.

AI agents introduce accessibility risks that differ from traditional software in three ways. First, AI agents often use natural language interaction, which appears inherently accessible but in practice makes assumptions about language processing speed, vocabulary, pragmatic inference, and social communication norms that disadvantage users with cognitive, learning, or communication disabilities. A chatbot that interprets slow or inconsistent responses as disengagement, non-compliance, or low confidence is making an ableist assumption that disadvantages multiple disability groups.

Second, AI agents increasingly use behavioural and biometric signals — facial expressions, voice characteristics, typing patterns, gaze direction — as inputs to decision-making. These signals correlate strongly with disability status. Facial expression analysis disadvantages persons with facial paralysis, autism spectrum conditions, or Parkinson's disease. Speech fluency analysis disadvantages persons with speech impediments, aphasia, or hearing impairment affecting speech production. Keystroke dynamics analysis disadvantages persons with motor impairments. Using these signals as predictive features in consequential decisions (hiring, credit scoring, benefits assessment) produces indirect disability discrimination by default.

Third, AI agents operate at scale and speed, meaning that an accessibility barrier affects every disabled user who encounters the system, simultaneously, with no opportunity for the case-by-case human judgment that traditionally provided ad hoc accommodation. When a human caseworker notices that a claimant is struggling, they may instinctively extend time, simplify language, or offer help. An AI agent does not do this unless it is specifically designed and required to do so.

The legal framework is unambiguous. The Equality Act 2010 imposes a duty to make reasonable adjustments. The Americans with Disabilities Act requires accessible design. The UN CRPD requires full and effective participation. The EU Accessibility Act requires accessible products and services. AG-241 translates these legal obligations into operational requirements for AI agent systems.

6. Implementation Guidance

AG-241 requires accessibility to be a structural property of the agent system, not an overlay or accommodation pathway. The implementation must address three domains: interface accessibility, interaction accessibility, and decision fairness.

Recommended patterns:

• Multi-modal interaction architecture. Design the agent with at least two fully functional interaction pathways from inception. For example: text-based interaction with full screen reader compatibility, and voice-based interaction with speech-to-text and text-to-speech. Each pathway must support every function — not "text for browsing, voice for authentication" but both modalities for all functions. Test with actual assistive technologies: JAWS, NVDA, VoiceOver, TalkBack, Dragon NaturallySpeaking, switch scanning devices, and eye-tracking systems. Ensure semantic structure (headings, landmarks, form labels) is present in all outputs.
• Adaptive timing with barrier detection. Replace fixed session timeouts with adaptive timing that responds to user pace. If a user's interaction rate is 3x slower than the median, the system should extend the session timeout proportionally (e.g., from 30 minutes to 90 minutes) without requiring the user to request an extension. Log the adaptation. If a user times out and returns, offer to resume from the point of interruption rather than requiring restart. Track repeated timeouts as a potential accessibility barrier indicator and trigger a proactive accommodation offer.
• Feature fairness auditing for disability impact. Before deploying any behavioural or biometric feature in a decision-making pipeline, conduct a disability impact assessment. For each feature, calculate the outcome differential between disabled and non-disabled populations (where disability status is available through voluntary disclosure or proxy indicators). If the feature produces an adverse impact ratio below 0.80 (four-fifths rule) for any disability group, the feature must be justified as measuring a genuine requirement or removed. Document the assessment, the justification (if retained), and the alternative considered (if removed).
• Accessibility preference persistence. Implement an accessibility profile that users can configure once and have applied across all interactions. The profile stores: preferred interaction modality, text size preference, contrast preference, language complexity level, extended time preference, and preferred alternative formats. The profile is portable (exportable, importable) and revocable. The agent loads the profile at session start and applies all preferences before the first interaction.

Anti-patterns to avoid:

• Accessibility as a separate channel. Routing disabled users to a different system, a different agent, or a human-only pathway is not accommodation — it is segregation. Disabled users are entitled to use the same system with adjustments, not a different system.
• Retrofitting accessibility after deployment. Accessibility added after system design is consistently less effective, more expensive, and more likely to break during updates. Accessibility must be a design requirement from the architecture phase, not a remediation task.
• Assuming assistive technology compatibility without testing. "We use standard HTML so screen readers should work" is not accessibility testing. Actual testing with actual assistive technologies by actual disabled users is required. Automated accessibility scanning tools (e.g., axe, WAVE) catch approximately 30-40% of accessibility issues; manual testing is essential.
• Using disability-correlated features as "objective" predictors. Facial expression analysis, speech fluency scoring, and eye-contact measurement are not objective — they measure conformity to neurotypical social norms. Labelling them as objective predictors launders disability discrimination as meritocratic assessment.
• Requiring disclosure to receive accommodation. Requiring users to declare a disability before receiving any accommodation disadvantages users who do not wish to disclose, who do not recognise their condition as a disability, or who are unaware that accommodation is available. Proactive barrier detection and universal design reduce the need for disclosure.

Industry Considerations

Financial Services. The FCA expects firms to ensure that disabled customers can access services on equivalent terms. AI agents in banking, insurance, and investment must support accessible interaction for all product functions. Payment authentication (SCA under PSD2) must include at least one modality accessible to each disability group — voice biometrics alone is insufficient.

Healthcare. Patient-facing AI agents must comply with NHS Accessible Information Standard (DCB1605), which requires that patients' communication needs are identified, recorded, flagged, shared, and met. Triage agents, appointment booking systems, and patient communication agents must support accessible formats. Decision-support agents must not use disability-correlated features as clinical risk proxies without clinical validation.

Employment. AI agents used in recruitment (CV screening, interview assessment, skills testing) are high-risk for disability discrimination. The Equality and Human Rights Commission has issued specific guidance on AI in recruitment and disability. Features such as video interview analysis must be independently assessed for disability impact before deployment.

Public Sector. The Public Sector Equality Duty (Section 149) requires proactive consideration of disability equality. Government-facing agents must meet WCAG 2.2 Level AA as a minimum. The Government Digital Service (GDS) accessibility requirements apply to all public-sector digital services, including AI agents.

Maturity Model

Basic Implementation — The agent interface meets WCAG 2.1 Level A. One alternative interaction modality is available but does not support all functions. Session timeouts are configurable but not adaptive. Accessibility adjustments are available on request. No disability impact assessment has been conducted on decision-making features. This meets minimum legal requirements in some jurisdictions but falls short of the standard required by AG-241.

Intermediate Implementation — The agent interface meets WCAG 2.2 Level AA. Two fully functional interaction modalities are available. Adaptive timing is implemented with barrier detection. Accessibility profiles are available. Disability impact assessments have been conducted on all decision-making features, with features producing adverse impact ratios below 0.80 either justified or removed. Accessibility testing includes at least 3 assistive technologies and user testing with disabled participants. Accessibility barrier logs are reviewed monthly.

Advanced Implementation — All intermediate capabilities plus: WCAG 2.2 Level AAA for critical functions. Proactive barrier detection offers accommodation without user request. Disability impact assessments are repeated quarterly with updated data. User testing with disabled participants from at least 5 disability groups is conducted annually. Accessibility feedback mechanism is actively monitored with median response time under 5 business days. The organisation publishes an accessibility statement with performance metrics. Independent accessibility audit is conducted annually by a disabled-person-led organisation.

7. Evidence Requirements

Required artefacts:

• Accessibility conformance report. WCAG 2.2 Level AA conformance assessment covering all agent interaction surfaces. Conducted by a qualified assessor, updated with each major release. Format: Voluntary Product Accessibility Template (VPAT) or equivalent.
• Assistive technology testing results. Results from testing with at least 3 assistive technologies (e.g., JAWS, NVDA, VoiceOver) and at least one alternative input method (e.g., switch scanning, voice control). Include test date, technology versions, issues identified, and remediation status.
• Disability impact assessment. For each behavioural or biometric feature used in decision-making: the feature description, the disability groups potentially affected, the adverse impact ratio calculation, the justification (if retained despite adverse impact), and any mitigation measures.
• Accessibility adjustment log. Records of accessibility adjustments requested, adjustments applied, adjustments not provided (with reasons), and barrier detections. Minimum 12 months retention.
• Disabled user testing report. Results from user testing with disabled participants including: disability groups represented, tasks tested, issues identified, severity ratings, and remediation plan.

Retention requirements:

• Conformance reports and impact assessments: retained for the lifetime of the system version plus 3 years.
• Adjustment logs: minimum 5 years for regulated sectors; minimum 3 years otherwise.

Access requirements:

• Producible to regulators or auditors within 48 hours. Accessibility statements should be publicly available.

8. Test Specification

Test 8.1: Multi-Modal Interaction Completeness

• Stimulus: Attempt to complete every core function of the agent using each supported interaction modality independently (e.g., text-only, voice-only).
• Expected behaviour: Every core function is completable through each modality. No function requires a specific sensory or motor capability that is available in only one modality.
• Pass criteria: 100% of core functions are completable through each supported modality independently.
• Fail criteria: Any core function requires a specific modality that is not available through an alternative.

Test 8.2: Assistive Technology Compatibility

• Stimulus: Navigate and complete core functions using each of the following: a screen reader (JAWS or NVDA), a screen magnifier, a switch scanning device, and voice control software.
• Expected behaviour: All interactive elements are labelled, navigable, and operable through each assistive technology. Focus order is logical. Dynamic content updates are announced.
• Pass criteria: All core functions are completable with each assistive technology. No unlabelled controls, focus traps, or unannounced updates.
• Fail criteria: Any core function is not completable with any tested assistive technology, or any focus trap or unlabelled control is encountered.

Test 8.3: Adaptive Timing Verification

• Stimulus: Interact with the system at 3x the median interaction rate (simulating a user who requires extended time). Monitor whether session timeout extends and whether the agent offers accommodation.
• Expected behaviour: Session timeout extends proportionally. If timeout occurs, resume from the point of interruption is offered on return. Proactive accommodation offer is triggered after detecting slow interaction pattern.
• Pass criteria: Timeout extends by at least 2x. Resume-from-interruption is available. Accommodation offer triggers within 3 interaction turns of detecting slow pattern.
• Fail criteria: Fixed timeout applies regardless of interaction speed, no resume is offered, or no accommodation is triggered.

Test 8.4: Decision Feature Disability Impact

• Stimulus: Submit decision requests for a test population that includes individuals with disclosed disabilities across at least 4 disability categories. Calculate outcome rates by disability status.
• Expected behaviour: No disability group experiences an adverse outcome rate more than 1.25x the non-disabled rate (equivalent to an adverse impact ratio of 0.80 or higher).
• Pass criteria: Adverse impact ratio >= 0.80 for all disability groups tested.
• Fail criteria: Adverse impact ratio below 0.80 for any disability group, without documented justification as a genuine requirement.

Test 8.5: Reasonable Adjustment Request Processing

• Stimulus: Submit accessibility adjustment requests across all supported adjustment types (extended time, simplified language, alternative format, high contrast, alternative input method).
• Expected behaviour: Each adjustment is applied within the current interaction. The adjustment persists for the duration of the session. The adjustment is logged.
• Pass criteria: 100% of supported adjustment types are applied within 1 interaction turn. Adjustment persists for the session. Log entry is created.
• Fail criteria: Any supported adjustment fails to apply, or adjustment does not persist, or log entry is missing.

Test 8.6: Accessibility Logging Completeness

• Stimulus: Execute a sequence including: barrier detection, adjustment request, adjustment application, and barrier resolution. Retrieve the audit log.
• Expected behaviour: Log contains: timestamp, user session identifier, barrier type detected, adjustment requested, adjustment applied, and outcome.
• Pass criteria: All required log fields are present and non-null. Log entries are immutable.
• Fail criteria: Any required field is missing, or log entries can be modified.

Conformance Scoring

• Score 0: No accessibility consideration — the agent has a single interaction modality with no adjustments, and decision features have not been assessed for disability impact.
• Score 1: Basic WCAG compliance (Level A). One alternative modality available but incomplete. Adjustments available on explicit request only. No disability impact assessment on decision features.
• Score 2: WCAG 2.2 Level AA. Two complete interaction modalities. Adaptive timing with barrier detection. Disability impact assessments conducted on all decision features. Assistive technology testing with at least 3 technologies.
• Score 3: All Score 2 capabilities plus: proactive barrier detection and accommodation. Disabled user testing with at least 5 disability groups. Quarterly disability impact reassessment. Independent annual accessibility audit. Published accessibility statement with performance metrics.

9. Regulatory Mapping

Regulation	Provision	Relationship Type
Equality Act 2010	Sections 20-21 (Duty to Make Reasonable Adjustments)	Direct requirement
Equality Act 2010	Section 149 (Public Sector Equality Duty)	Direct requirement
ADA	Title III (Public Accommodations)	Direct requirement
EU Accessibility Act	Directive 2019/882	Direct requirement
UN CRPD	Articles 9, 21 (Accessibility, Freedom of Expression)	Supports compliance
EN 301 549	Accessibility Requirements for ICT Products and Services	Supports compliance
WCAG 2.2	Levels A, AA, AAA	Supports compliance
EU AI Act	Article 9 (Risk Management)	Supports compliance

Equality Act 2010 — Sections 20-21 (Duty to Make Reasonable Adjustments)

The duty to make reasonable adjustments is a cornerstone of UK disability law. Service providers must take reasonable steps to avoid substantial disadvantage to disabled persons. For AI agents, this translates to: providing alternative interaction modalities where the primary modality creates a barrier; adjusting timing, pace, and complexity where fixed parameters disadvantage disabled users; and ensuring decision-making features do not penalise disability-correlated characteristics. The duty is anticipatory — organisations must anticipate the needs of disabled people in advance, not wait for individual requests. An AI agent deployed without multi-modal interaction, adaptive timing, or disability impact assessment has not complied with the anticipatory duty.

ADA — Title III (Public Accommodations)

Title III prohibits discrimination on the basis of disability in places of public accommodation. US courts have increasingly applied Title III to digital services, including AI-powered interfaces. The Department of Justice has issued guidance confirming that web and mobile accessibility obligations apply to public accommodations. AI agents that serve the US public must ensure accessibility equivalent to physical premises accessibility.

EU Accessibility Act — Directive 2019/882

The European Accessibility Act establishes accessibility requirements for products and services including digital services. AI agents provided as or embedded within in-scope services must meet the accessibility requirements by June 2025. The requirements align with EN 301 549, which references WCAG 2.2. AG-241 requirements meet or exceed the EAA's digital service accessibility standards.

UN CRPD — Articles 9 and 21

Article 9 requires States Parties to take appropriate measures to ensure persons with disabilities have access to information and communications technologies on an equal basis with others. Article 21 requires accessible formats and technologies. While the CRPD binds states rather than private actors directly, its principles inform regulatory interpretation and judicial reasoning across jurisdictions. AG-241 implements the CRPD's accessibility principles at the operational level.

10. Failure Severity

Field	Value
Severity Rating	High
Blast Radius	Cohort-level — systematically affecting all disabled users across the agent's user population, approximately 15-20% of users in most populations

Consequence chain: Failure to implement accessibility and disability accommodation in AI agents produces systematic exclusion of disabled persons from services, decisions, and opportunities. The immediate technical failure is an accessibility barrier — a deaf user cannot authenticate, a user with a learning disability cannot complete a timed form, a candidate with autism is screened out by facial analysis. The operational impact is that disabled users receive systematically worse outcomes: lower service quality, reduced access, higher abandonment rates, and adverse decisions based on disability-correlated features. The scale is significant — approximately 16% of the global population has a significant disability (WHO, 2023), and the proportion is higher among older adults and lower-income groups who are also more likely to interact with government and financial service agents. The legal exposure is substantial: Equality Act claims, ADA lawsuits, regulatory enforcement under the EU Accessibility Act. US ADA digital accessibility settlements have ranged from $50,000 to $12 million. UK Equality Act findings have required remediation programmes costing £1-5 million. The reputational consequence is significant and sustained, as disability rights organisations are well-organised and effective at publicising systemic exclusion.

Cross-references: AG-239 (Vulnerable Person Protection Governance) provides the general vulnerability framework; AG-241 specialises this for disability-specific accommodation. AG-242 (Non-Discrimination Outcome Testing Governance) covers disparate impact testing, which complements AG-241's disability impact assessment. AG-051 (Fundamental Rights Impact Assessment) requires rights impact assessment including disability rights. AG-062 (Automated Decision Contestability) ensures disabled persons can contest adverse automated decisions. AG-246 (Cultural and Linguistic Fairness Governance) addresses related accessibility concerns for linguistic minorities. AG-240 through AG-248 are sibling dimensions.