← All perspectives

Perspective / UX engineering

Accessibility engineering: from semantic tokens to tested customer journeys

Accessibility depends on component behaviour, content and complete workflows. Connect semantic design decisions to automated regression checks, manual evaluation and journey-level acceptance evidence through an illustrative engineering model.

10 min readReviewed

Hands using a refreshable braille display beside a laptop.
Conceptual editorial image.

Accessibility engineering makes inclusive use a testable part of delivery across components, content and complete customer journeys. Shared tokens and automated checks can catch important classes of defect, while human evaluation is needed to understand behaviour and context. The executive task is to fund a coherent quality system with clear scope, evidence and ownership. That system should explain what has been assessed, which users and workflows it covers, and where further work remains, without presenting a scan result as proof of universal accessibility.

The executive view

  • Set an assessment scope covering complete customer processes and the integrations on which they depend.
  • Combine token and automated checks with component-state testing, manual evaluation and relevant user evidence.
  • Maintain traceable acceptance evidence and regression coverage as components, content and integrations change.
01

Define the scope of the quality commitment

Identify the service being evaluated, its critical journeys, supported technology context and the accessibility standard used for assessment. Include authenticated workflows, downloads and embedded services where they form part of the experience. A convenient sample of public pages may omit the activity customers actually need to complete.

WCAG 2.2 provides testable success criteria and conformance requirements. The current Recommendation is dated December 2024. Its scope includes complete pages and complete processes, which makes journey boundaries important when planning evidence.

Translate that scope into delivery responsibilities. The component team may own interaction patterns; editors own published content; another supplier may operate an embedded service. Assign responsibility for integration and end-to-end evaluation as well. A collection of separately assessed parts does not by itself establish the behaviour of the assembled journey.

Separate the standard being assessed from a claim that the entire organisation is accessible. A product release, a document library and an external identity journey can have different evidence and unresolved barriers. Name the release and process in scope, identify the intended conformance level and retain the evidence used. Procurement answers should describe that scope directly rather than generalise from a supplier's component badge.

Include the states a customer reaches when something goes wrong. An expired session, an unavailable upload or an account-permission change may expose different controls and messages from the primary journey. These states are often essential to complete the process. They also reveal shared responsibility: an identity provider may render a sign-in page, while the application must preserve context and make a failed return understandable. Assign the boundary owner who can coordinate evaluation and remediation across suppliers. Without that ownership, each component can appear acceptable in isolation while the assembled service leaves a customer unable to proceed.

02

Write component contracts for behaviour and state

A component contract should describe purpose, semantics, interaction and state changes. For a dialog, define how it is identified, where focus moves, how it closes and what happens when it disappears. For a form field, define its label, required state, instructions, error relationship and disabled behaviour.

Represent states explicitly in the component catalogue. Default, focus, loading, invalid, success and unavailable states can expose different problems. Testing only the visual default leaves much of the actual interaction unexamined.

Keep semantic responsibilities visible in the interface between component and consumer. If an application must provide an accessible name or meaningful error message, document that requirement and validate what can be checked. Prefer APIs that make correct use straightforward. Escape hatches may be necessary, but their effect and review requirements should be clear rather than becoming undocumented ways around the shared behaviour.

Take a confirmation dialog for changing an account's payment terms. An implementation needs an accessible name, coherent content order, appropriate initial focus and a route to close or cancel. The W3C modal-dialog pattern describes expected focus and keyboard behaviour. Use it as implementation guidance, then evaluate the actual component with its real content and surrounding page.

The consumer contract should identify responsibilities that cannot be supplied by a generic primitive. The application owns the business-specific warning, the action label and the destination of focus when the trigger disappears after success. An automated component test can verify that the dialog is named and that focus remains within it. It cannot establish that the warning makes the financial consequence understandable. Keep that distinction in the acceptance record. Native controls can reduce custom interaction work where they fit, but a native element still requires correct labelling, content and integration. The objective is predictable behaviour throughout the actual task, not a particular implementation technique.

03

Validate tokens within their rendered context

Semantic tokens help maintain repeated decisions such as foreground and background pairings, focus appearance and spacing. Add checks for the supported pairings and themes, and make a change to a shared value trigger review of affected components.

A token cannot fully describe the rendered result. Text can sit over imagery; transparency changes effective colour; user settings and content length affect layout. A valid token value can therefore produce a problematic interface when combined with other elements.

Use resolved component examples containing realistic content and supported states. Evaluate zoom, reflow, long labels and error messages alongside the intended presentation. Keep exceptions traceable to their owners. A product-specific override should identify which assumptions of the shared component it changes and what additional evaluation it needs. This connects the efficiency of a shared system to evidence about the interfaces people actually encounter.

Token checks should resolve complete combinations. A muted text colour can pass against one background and fail against another; an overlay can change both. Document supported pairings for body copy, disabled states, errors, selected items and focus treatment. Make unsupported combinations visible in the design system rather than assuming that individually approved colours are universally interchangeable.

Control geometry also belongs in the rendered contract. WCAG's Target Size Minimum explanation describes the 24 by 24 CSS-pixel AA criterion and its conditions and exceptions. Treat the actual requirement carefully rather than turning it into an unexplained universal button size. A dense data interface needs evaluation of adjacent targets, spacing and equivalent controls. Test long commercial labels, translated content and validation messages at supported reflow conditions. A neatly aligned default screenshot can hide clipped controls or an error message that displaces the only route forward. These are integration behaviours that need representative content, not only a token-value assertion.

04

Combine automation with skilled human evaluation

WebAIM’s 2026 Million report found automatically detectable WCAG failures on 95.9% of its sampled homepages. The evaluation is a useful signal of persistent defects, but it covers detectable issues within a defined sample. It does not establish a legal finding about that proportion of businesses or prove accessibility where errors were absent.

W3C’s evaluation overview, updated in August 2026, makes the complementary role of tools and knowledgeable human assessment explicit. Use automation for repeatable coverage, then manual evaluation for behaviour, meaning and context.

Build the two into one evidence workflow. A failing automated check should identify the affected component or page and an owner. Manual findings should contain reproducible steps, environment, observed behaviour and the relevant requirement. Preserve enough detail for another evaluator to understand the result without relying on the original tester’s memory.

05

Evaluate a complete illustrative procurement journey

Consider a hypothetical service where a customer requests a quotation and uploads supporting documents. The central path is only part of the assessment. Test the journey across realistic content, validation and completion states.

Journey stageEngineering questionEvidence to retain
Select a serviceCan the choice be understood and operated without a pointer?Navigation and control evaluation
Enter requirementsAre names, instructions and required fields coherent?Form semantics and task observation
Correct an errorCan the user find the problem and retain valid input?Error announcement and recovery evidence
Upload a documentAre status, constraints and failures understandable?Supported upload-state evaluation
Review the requestCan the user inspect and change information?Reading order and edit-path results
Complete the processIs the outcome clear through the supported interface?Confirmation-state and focus evaluation

This is an illustrative framework, not a claim that any particular service has passed assessment. Each row should become concrete test cases with defined environments and expected behaviour.

For document upload, model the asynchronous states explicitly: file selected, validating, uploading, accepted, rejected and interrupted. Decide what is announced, what remains available for inspection and whether users can cancel or retry. Repeating a generic live-region announcement on every progress update can become disruptive, while announcing nothing may leave the result unclear. Evaluate the message timing and interaction in the supported assistive-technology context.

At final submission, separate an in-progress request from a confirmed business outcome. If the service times out after accepting the request, the customer needs a recovery route that avoids creating a duplicate. Retain valid input where appropriate and make the current status inspectable. Test moving backwards to correct a field as well as progressing forwards. A user who can technically focus every control may still be unable to understand where an error occurred or which information survived. The end-to-end evidence should therefore include the final stored or displayed state, not only the ability to navigate the form.

06

Prioritise defects across the dependency graph

A shared component defect can affect many journeys, while a local content problem may block one critical task. Prioritisation needs both the number of affected instances and the consequence for the person using the service. Counted errors alone do not capture that difference.

Record the user impact, affected tasks, recoverability, scope and evidence quality. Identify whether the cause belongs to a shared primitive, component implementation, consumer integration or content process. Fixing the earliest appropriate dependency can address repeated defects, but it also requires regression evaluation across consumers.

Include people with relevant disabilities in evaluation where appropriate, with suitable recruitment and research arrangements. Their task experience can expose barriers beyond a mechanical reading of a checklist. Keep that evidence distinct from formal conformance assessment: useful user feedback and standards-based evaluation answer related but different questions. Neither should be used to imply more coverage than it actually provides.

Use the dependency graph to estimate the reach of a repair, while keeping blocked tasks visible. Suppose an illustrative audit finds one broken shared selector used on forty screens and one local confirmation error on a high-value payment journey. The selector may offer the largest reuse benefit, but the payment defect could carry the more urgent consequence. Do not decide by counting either screens or violations alone. Record severity, affected tasks, workaround quality and confidence in the evidence.

A shared fix also needs a bounded rollout plan. Identify consumers with custom behaviour, test representative integrations and monitor for regressions in focus, layout and validation. If a temporary workaround is necessary, give it a named owner and an expiry or review trigger. Keep known barriers in the operational support information so a customer can obtain useful assistance while remediation proceeds. Support cannot replace an accessible product, but it should understand the actual obstacle rather than ask the customer to repeat a failing interaction.

07

Keep acceptance evidence current after launch

Define the release evidence required for consequential changes. New interaction patterns, a replacement component, a third-party integration or substantial content changes can alter the accessibility of a previously assessed journey. Reuse existing tests where their assumptions remain valid and add coverage where the behaviour changes.

Maintain an acceptance matrix linking requirements, components, journeys, results and unresolved work. Record assessment dates and environment details. A statement about a previous release should not silently become a claim about every future version.

The operating decision is how to maintain this evidence within ordinary delivery. Assign ownership, reserve remediation capacity and review recurrent causes. Accessibility becomes more dependable when teams can trace a design decision through implementation to evaluated task behaviour, then repeat the relevant assessment as the service evolves. That approach supports clear procurement evidence while preserving an honest account of its scope and limitations.

Evidence maintenance can be proportional to change. A corrected spelling error does not normally need the same assessment as a new complex combobox, while a content edit that removes an instruction may change task comprehension materially. Define the triggers and use them in pull-request or release review. Store a compact evidence record containing the assessed version, environments, test method and result, with links to reproducible findings.

Use metrics to improve the delivery process without turning them into a conformance score. Escaped high-severity defects, time to remediate a blocked task and recurrence in shared components can identify where investment is useful. A falling automated error count can indicate progress but cannot establish that every required journey is usable. Review the evidence with product, engineering, content and supplier owners. That shared view supports a more accurate procurement response and a more useful remediation plan, because it connects stated commitments to specific behaviours that have actually been evaluated.

Further questions

Questions this raises

Can automated accessibility testing prove WCAG conformance?

No. Automated tools detect some issues, while skilled human evaluation is needed for other requirements and contextual behaviour. An absence of detected errors does not establish full conformance.

Do accessible design tokens make every component accessible?

No. Tokens can support consistent values and checks, but semantics, interaction, content, state changes and rendered combinations still need evaluation.

What should accessibility acceptance evidence include?

Record scope, applicable requirements, journeys, component states, environments, test methods, findings and unresolved work. Link the evidence to the release assessed and update it when consequential behaviour changes.

References

Sources & further reading

  1. W3C Web Content Accessibility Guidelines 2.2 12 December 2024
  2. WebAIM Million: 2026 report February 2026 evaluation
  3. W3C WAI: Evaluating Web Accessibility Overview Updated 12 August 2026
  4. W3C ARIA Authoring Practices: Modal Dialog Pattern
  5. W3C: Understanding Target Size Minimum

Connected capabilities

Brand & digital experienceSoftware & CRMStrategy & digital transformation

From thinking to doing

What could this mean
for your business?

Your next chapter

What could your
business do better?

Tell us where you want to go, what’s getting in the way, or what you need to deliver. We’ll start there.

Start a conversation

Start a conversation

Tell us what
you have in mind.

Interactive preview. Your information stays in this page and is not sent.