The licence looks affordable. The expensive part may be making the finished workflow usable by the people expected to rely on it.

Buy a usable workflow, not a persuasive demonstration

The price on an AI proposal usually describes the licence, the integration and perhaps a little training. It rarely tells you what happens if a colleague cannot operate the finished workflow with their usual assistive technology. That omission can turn an apparently economical purchase into another project: redesigning controls, rebuilding generated documents and finding someone to complete work the new system was meant to simplify.

In its 24 August procurement guidance, AbilityNet describes this precise buying problem. AI services combine interfaces, models, third-party components and data sources. A reassuring demonstration does not establish whether those parts work together for disabled users, or who will fix a barrier once the contract is signed.

The purchasing unit should therefore be a completed business task. For an internal proposal assistant, that could mean finding a customer brief, requesting a draft, checking its sources, changing an assumption and exporting an accessible document. The test must include the actual application and the actual output, not simply an impressive answer in the supplier's preferred chat window.

What this means in practice is straightforward: write the task before inviting the demonstration. Ask the supplier to show the same task with keyboard navigation and the assistive technologies relevant to your users. Ask what remains untested. Treat a candid limitation as information you can price and manage, rather than rewarding a vague assurance that everything is accessible.

This is not an argument for rejecting AI. An assistant may remove substantial barriers to reading, writing or handling information. It is an argument for buying that benefit deliberately. If the commercial promise is less work, a workaround that transfers work to someone else belongs in the evaluation.

The purchasing unit should be a completed business task.

Put the missing work into the cost comparison

Compare suppliers on the cost of delivering an accepted outcome, not just the cost of access. Accessibility belongs beside security, integration and support in that calculation. Keep the categories separate so a low licence fee cannot conceal uncertainty elsewhere. You do not need an invented industry average or a complicated financial model to make this useful.

Start with supplier charges. Then record the work your own team must provide: checking generated content, correcting headings or tables, supporting people who cannot use a control, and maintaining an alternative process. Add external specialist assessment where needed. Separate one-off setup from recurring effort, because a document-formatting problem that happens on every proposal is not merely an implementation snag.

Use a small pilot to measure those activities. Record the task, the observed barrier, who resolved it and the time spent. Do not record unnecessary medical details about participants. A cost estimate can use agreed internal hourly rates, but label assumptions and show the range. An uncertain remediation quotation should remain uncertain rather than disappearing inside a confident return-on-investment percentage.

AbilityNet's 6 October review of AI and accessibility warns that faster AI-assisted development can repeat inaccessible patterns at scale. The commercial implication is not that every generated screen is defective. It is that producing more screens or content does not make the assurance and correction work free.

What this means in practice: ask two shortlisted suppliers to complete the same bounded workflow, then compare purchase price, assessment effort, recurring correction and unresolved limitations. Keep critical usability requirements as acceptance conditions, not optional benefits that can be traded away for a discount. A cheaper service can still win. It simply has to be cheaper after the necessary work is counted.

A workaround that recurs on every task is an operating cost.

Ask for evidence that matches the product you will use

An accessibility statement is a starting point, not a substitute for evidence. Ask which product version, interface, output types and assistive technologies were assessed. A report about the supplier's marketing website does not establish the accessibility of its internal AI editor. A report about an editor does not necessarily cover generated PDFs, embedded tables or your custom sign-in journey.

The GOV.UK guide to WCAG 2.2 provides a useful technical vocabulary: four principles, supported by 13 guidelines. Content should be perceivable, operable, understandable and robust. Government services must achieve level AA under their accessibility requirements. Do not turn that public-sector statement into a claim that every private business has exactly the same statutory regime.

For procurement, translate the principles into observable tasks. Can someone reach every necessary control without a mouse? Does the interface announce that a new answer has arrived without repeatedly interrupting a screen reader? Can a user identify which source supports a claim? Do headings and table relationships survive export? Can someone correct an error without losing the entire conversation?

Tools such as NVDA and VoiceOver can support testing, but installing a screen reader does not make a team competent to interpret every result. Use knowledgeable testers and involve people who use relevant assistive technologies in their normal work. Match the browser and device combinations to the proposed deployment rather than collecting a long list of irrelevant test environments.

Request the known-issues list as well as the successful test results. Ask who owns each issue, what workaround exists and what evidence will demonstrate a fix. Avoid a binary questionnaire where every answer is simply yes. A specific limitation with an agreed remedy is much more commercially useful than an unqualified claim with no supporting record.

Evidence about the marketing website is not evidence about your AI workflow.

Use AI testing to widen coverage, not to invent user approval

The obvious counterargument is cost. Small businesses cannot run an extensive research programme for every tool, and an AI-assisted accessibility checker may look like a practical alternative. There is a sensible case for automated help. It can find repeatable problems early, check more pages and reduce the time specialists spend on defects a machine could have flagged.

The boundary is what the result proves. In its 13 October discussion of synthetic accessibility testing, AbilityNet distinguishes simulated constraints from lived experience. It names Collectivly as an example of research-informed AI personas, while noting that its example reports warn of inaccuracies and concerns requiring tests with real people. That is a qualified example, not an endorsement or a performance comparison.

A simulated participant completing a journey does not demonstrate that a real person can complete it confidently. Uncertainty, fatigue, personal settings and recovery strategies matter. An automated tool may find that a button has an accessible name while missing that its wording leaves someone unsure whether they are saving a draft or sending an email.

For a proportionate approach, layer the methods. Run automated checks and expert review first. Correct obvious failures before inviting disabled users to spend time on the workflow. Then use focused human sessions on the tasks where mistakes or exclusion would matter most. Record which findings came from software, specialists and direct user observation so the evidence cannot quietly change meaning later.

Do not demand that a handful of participants represent every disability. Explain the coverage and gaps honestly. If the proposed purchase depends on a task that remains untested, limit the pilot or retain a usable alternative until that uncertainty is resolved. Automation should make genuine user involvement more valuable, not make it disappear from the budget.

A synthetic user completing a task is not human approval.

Contract for change, not just the day of purchase

An AI service can change after acceptance. A supplier may replace its editor, alter generated document formats or introduce a new streaming response interface. Even when the model produces better answers, the surrounding interaction can become harder to use. Initial evidence therefore needs an explicit scope and a plan for reassessment when something material changes.

Ask the supplier to identify changes it controls, changes controlled by upstream providers and the notice it can realistically offer. Agree which changes trigger retesting, how barriers are reported and who coordinates a response. Avoid promises that a supplier cannot enforce across its own supply chain. A contract should make uncertainty visible rather than pretending that a hosted service can be frozen indefinitely.

Set out the practical remedies. These might include fixing the problem within an agreed period, preserving an alternative accessible route, assisting with export or allowing termination under defined conditions. Have appropriate commercial or legal advice on the wording. None of those clauses removes the buyer's responsibility to operate the service sensibly or meet applicable obligations.

For workplace deployments, GOV.UK guidance on reasonable adjustments states that employers must make reasonable adjustments so disabled workers are not substantially disadvantaged. The applicable framework and circumstances matter, including the separate arrangements in Northern Ireland. Buying an AI subscription does not make that responsibility the software supplier's alone.

Before signing, produce a short decision record: the critical task, tested configuration, known limitations, expected running costs, owner and reassessment triggers. If the evidence is incomplete, authorise a bounded evaluation rather than pretending you have approved a business-wide rollout. The best-value contract is the one that lets people do useful work reliably, with a clear route to resolve failures. Its price should reflect that complete service, not just the attractive number on the first page.

The best-value contract prices a complete usable service, not just a licence.

Key Takeaways

  • Define a completed business task before comparing AI supplier demonstrations.
  • Separate licence price, setup assessment, recurring correction and unresolved remediation costs.
  • Request evidence for the deployed interface and generated outputs, not only the supplier website.
  • Combine automated checks, specialist review and focused testing with disabled users.
  • Agree material-change retesting, named ownership and practical remedies before signing.