cx-ai-disclosure

Use to design and audit how customers are told they are interacting with AI, or that AI was involved in a decision about them, and to evidence that it happened. Trigger for "do we tell customers it's

SKILL.md

GitHub

Telling customers about AI

Three distinct obligations get collapsed into one conversation, and they have different triggers, different audiences and different evidence:

  1. You are talking to a machine. Transparency about the interaction itself.
  2. AI was involved in a decision about you. Transparency about a process, which may carry a right to an explanation or to human review.
  3. Your conversation was processed by AI. A data-processing question — QA scoring, sentiment analysis, summarisation — that the customer may never see any output from.

The third is the one most often missed, because nothing about it is visible to the customer. A support operation running AI QA over every conversation is processing personal data with AI, whether or not a bot ever replied.

What is legally required varies by jurisdiction, sector and use, and it is changing. Nothing here states an obligation. This gives you the design and the evidence; the determination is legal's.

Design the interaction disclosure

Up front, not on request. A bot that identifies itself only when asked has left every customer who did not think to ask uninformed. Disclose at the start of the interaction.

In the conversation, not only in a policy. A line in the privacy notice is not a disclosure to someone in a chat window.

Answer the direct question honestly, always. "Am I talking to a human?" must get a straight answer. A bot deflecting that question is the single worst failure available here, it is trivially screenshotable, and it destroys trust in everything else you say about AI. Test it explicitly — ask your own bot, in every language.

Give a route to a human, and make it work. Disclosure without an escape is worse than none: the customer now knows they are stuck with a machine. Measure whether the route completes, not whether it exists.

Name it consistently. A bot with a human-sounding name and no other signal is misleading by design. If it has a persona, the disclosure has to be clearer to compensate.

Handle the handover in both directions. Tell the customer when they move from bot to human, and when a human hands back to automation. Mid-conversation transitions are where customers most often end up misled, because they reasonably assume continuity.

Cover the channels people forget — voice bots, proactive outbound messages, and automated replies that look like a person wrote them. An automated acknowledgement in a human agent's name is a disclosure failure hiding as a template.

AI-assisted decisions

Where AI contributed to a decision affecting the person — a claim, a refusal, a redress amount, an account action, or an agent's performance assessment:

  • Say that it was involved, at the point the decision is communicated.
  • Explain the basis in terms the person can act on. "The model scored it 62" is not an explanation. What factors mattered, and what would change the outcome.
  • Offer human review where that applies, and make it real — a review that re-runs the same model is not human review, and a human who only sees the model's conclusion is rubber-stamping.
  • Record the provenance: model, version, prompt, retrieved sources, raw output before post-processing. Without this the explanation cannot be reconstructed later, which is a separate problem when the person asks again in six months.

This applies internally too. An agent assessed with AI QA is a person subject to AI-assisted decision-making about them, and the same transparency logic applies — plus employment-law considerations that vary by jurisdiction.

Auditing what actually happens

Design is the easy half. Test the operation:

  • Disclosure presence rate, per channel, per path, per language. Automated paths built after the policy are the routine gap.
  • Position in the conversation — was it first, or after several turns?
  • The direct-question test. Sample conversations where the customer asked whether they were talking to a human, and check what happened. Report the honest-answer rate as its own number.
  • Handover disclosure, both directions.
  • Human-route completion, not availability.
  • Decision-explanation presence, where AI-assisted decisions were communicated.
  • Consistency after a change. A model or prompt change can alter behaviour without anyone updating the disclosure; re-test after every deployment.

The failure modes worth naming

  • Disclosure on request only. Covered above and worth repeating: it is the most common design.
  • A persona that undercuts the disclosure. Disclosed once, then forty turns of human-styled conversation with a first name.
  • Disclosure in a language the customer is not using.
  • A human-sounding automated reply in a named agent's voice.
  • Silent escalation to AI. A human conversation quietly continued by automation.
  • Invisible processing undisclosed. QA scoring, sentiment analysis and summarisation of conversations, with nothing anywhere telling the customer.
  • Recording notification updated for AI and not the reverse — a firm that discloses AI processing but has quietly started retaining transcripts longer to feed it.

Guardrails

  • Do not state what the law requires. Design and evidence; the obligation is legal's determination, and it differs by jurisdiction and is moving.
  • Never instruct a bot to deny being a bot, or to deflect the question. If a system currently does either, that is an urgent finding, not an audit note.
  • Do not treat a disclosure as a licence. Telling the customer it is AI does not make an otherwise-inappropriate automated interaction appropriate — a bot handling a vulnerability disclosure is a problem whether or not it introduced itself.
  • Do not remove a human route to improve containment. If that has happened, report it as a conduct finding.
  • A disclosure failure affecting decisions already communicated goes to compliance and legal, because the remedy may involve re-notifying customers.
  • Cite ids; quote only the disclosure line.

Present results to the user

  1. The three obligations, separated, with which of them are in scope here and what is currently done for each — the invisible-processing one named explicitly.
  2. Design review against the points above, per channel and per path.
  3. Presence and position rates, per channel, path and language, in counts as well as rates.
  4. The direct-question test result, as its own headline number.
  5. Handover disclosure, both directions, and human-route completion rather than availability.
  6. AI-assisted decision explanations — presence, and whether the explanation is actionable rather than a score.
  7. Provenance recording, so an explanation can be reconstructed later.
  8. Internal application — agents assessed by AI, and whether they were told.
  9. What needs a legal determination, and anything requiring immediate change rather than a recommendation.