Resources · 42
Accessible conversational assistants: multiple modes, recovery and human handoff
Design a support assistant that people can access, understand and leave for human help without starting their request again.
· 18 min
What this guide helps achieve
- Make input and output usable through several modes
- Let people understand and correct an answer
- Define when and how to reach a person
- Test the whole journey with users and difficult cases
Quick check
- Does the service work with a keyboard and relevant assistive technology?
- Can a person read a reply without depending on speech or a visual effect?
- Do errors explain how to recover?
- Does the answer distinguish verified information from a guess?
- Is human help discoverable before and after failure?
- Does the agent receive only necessary context, with the user’s choice?
Step-by-step method
- 01
Set tasks and explicit boundaries
Choose real requests: find information, fix an account issue, understand a charge and seek human help. State what the assistant may explain, suggest or pass on, and what it must never decide alone.
Deliverable: task map, necessary data and handoff triggers.
- 02
Offer several ways to interact
Check text, keyboard use, focus, reading order, contrast, zoom and announcements of new messages. If voice is offered, keep a complete text route and controls to slow, reread or interrupt it.
Deliverable: journeys tested on representative devices and input modes.
- 03
Make answers and errors recoverable
An answer should show where material information comes from when a decision needs evidence. Avoid claiming certainty without support. If the request is ambiguous, clarify the need, keep useful input and give a clear next action.
Deliverable: correct, uncertain, wrong and unanswerable response scenarios.
- 04
Design the transfer to a person
Define handoff triggers: explicit request, repeated failure, sensitive situation, disagreement or an action outside scope. Explain the channel, availability and information to be shared; let the user amend or remove that context.
Deliverable: handoff protocol, reviewable summary and receiving owner.
- 05
Test continuity end to end
Replay every task with varied abilities, devices and environments, then follow what happens after transfer. Track resolution and abandonment, review serious errors and repeat after a model or content change.
Deliverable: dated test set, prioritised findings and release decision.
Worked example
Illustrative situation
A customer asks why an expected refund is absent. The assistant first gives general information, then states it cannot verify the customer’s case. Using a keyboard and screen reader, the customer can reach the same handoff control.
Decision and expected evidence
Before transfer, the customer reviews and corrects a summary that omits payment details. The human agent receives the issue, steps already tried and unresolved question; the customer does not have to repeat the journey. The test records continuity and resolution, not only automation volume.
Management indicators
| Indicator | What it measures | First action |
|---|---|---|
| Task completion | Share of scenarios resolved without misleading information by access mode | Fix the journey or narrow scope |
| Effective handoff | Requests reaching a person with useful context | Repair availability, summary or routing |
| Recoverable error | Failures where the user can correct and continue without re-entering everything | Revise messages and retained state |
| Answer quality | Answers checked against a dated set including unanswerable cases | Update sources and refusal thresholds |
Common pitfalls
- Treating voice as the only accessible channel
- Moving focus for every new message without control
- Giving confident answers when evidence is absent
- Hiding human contact behind repeated refusals
- Passing the full conversation without need or choice
- Measuring only deflection from support
Frequently asked questions
Does an accessible chatbot make the whole service accessible?
No. Pages, forms, transfer channels and responses after handoff are part of the same task.
Must handoff wait for repeated failures?
No. People should be able to request it directly; repeated failure or sensitive context may also trigger it.
Is the W3C natural-language interface note a conformance standard?
No. It describes user needs and is a work in progress. Applicable WCAG criteria for the web service must be checked separately.
Should the entire conversation be retained for the agent?
Pass only the context needed for the request, tell the user and allow correction or removal before transfer.
Official references
Continue exploring
Related analysis
