Why customers ask and what a wrong answer costs
Customers ask about AI for ordinary reasons. Some want to know whether they can reach a person. Some are judging whether an image, a review or a message can be believed. Some are worried about their data, often after reading about AI tools in the news. And some are testing whether the business is straight with them. The question is rarely hostile, but the answer is remembered.
The expectation-disconfirmation model explains why a wrong answer costs more than the question seems to justify. Satisfaction is the experience judged against what the customer expected, and the business's own words set much of that expectation. A customer told "no AI here" who later sees an automated message, or told "a person wrote this" who later learns otherwise, is not judging the AI use; they are judging the gap between what they were told and what was true. The disclosure and data privacy guide makes the same point in its own terms: the reaction is proportionate to the surprise, not to the harm.
There are firmer reasons too. The site's guide to AI claims and consumer protection, drawing on the FTC's business guidance, records that the standard for deception does not change with the technology: an assistant presented as a person, a fabricated review or a misleading answer is judged like any other deceptive practice. An untrue answer to a direct question is a statement by the business, whoever gave it.
The questions customers ask about AI
Most questions fall into a handful of kinds. The table maps each to what the customer usually wants, where the true answer comes from and who should give it. It is general practice; the specific wording comes later.
| Question | What the customer usually wants | Where the true answer comes from | Who answers |
|---|---|---|---|
| Am I talking to a person or a bot? | To know who they are dealing with and whether they can reach a person | The assistant's setup and handover route | The assistant itself, immediately, and any agent |
| Did a person write this message or post? | To know whether the words reflect a person's judgement | The register row for the channel and its review route | The team that owns the channel |
| Is this photo or video real? | To know whether the product, place or person is as shown | The creative record and the register row for image tools | Marketing, from the asset's record |
| Are these reviews real? | To know whether the praise came from real customers | The review collection process | Marketing or service; the answer must be yes, because generated reviews are never acceptable |
| Did my details go into an AI tool? | Reassurance about privacy, sometimes a data rights request | The register's data column and vendor terms | Service for the facts; the privacy owner for requests |
| Did my data train an AI model? | To know whether their information left the business's control | The vendor terms recorded in the register | The privacy owner or the register owner |
| Can I opt out of AI? | Usually a person instead of an assistant, or no automated processing of their data | The handover route; the data protection owner for processing | Service for handover; the privacy owner for data |
| Do you use AI at all? | A general sense of the business's approach | A short public statement agreed from the register | Anyone, from the agreed wording |
Two of these have fixed answers whatever the register says. The assistant must say it is automated when asked, and the business's reviews must be real. The AI claims guide is clear that a generated review is a fake review whatever prompt produced it; if the honest answer to "are these reviews real" is anything other than yes, the problem is the reviews, not the reply.
Finding the true answer in the register
The AI use register is the single source of the facts. It records each use as a row, with the tool and feature, the data it sees, what it can do, the review route and the disclosure. Nearly every customer question maps to one or two of those columns: "was this written by a person" is the review route and what the tool can do; "did my details go into an AI tool" is the data column and the vendor terms; "is this image real" is the use row for image tools and the record kept with the asset.
That only works if the register is complete. The AI use policy tutorial notes that registers routinely find more built-in uses than teams expected: features inside the email platform, the booking system or the ad platform that draft copy or process customer data. Those are exactly the uses staff forget when they answer "do you use AI". Before writing any answer, check that the register covers every customer-facing channel and every tool that handles customer data, including those run by agencies.
The rule that follows is simple: no answer denies a use the register records. If the register shows that appointment reminders are drafted by a feature in the booking system, an answer saying "our messages are all written by our team" is untrue, however it was meant. Where the register is unclear, the answer is "I will check", and the question goes to the register's owner.
Writing answers people can trust
Prepared answers belong in the same library as every other prepared answer. The tutorial on building a library of saved replies builds a Saved ReplyA pre-written answer a support agent inserts and then edits for the customer in front of them, keeping facts and tone consistent across the team. for each question customers actually ask, with an owner, a review date and the personal parts marked; questions about AI join it as contact reasons like any other. Write one reply per question, and give the same facts to the assistant, the service team and the social team, so a customer who asks twice gets the same answer.
Good answers share a few habits.
- Answer the question asked, first. "Yes, you are chatting with our automated assistant. You can reach a person by choosing 'talk to the team'." The detail can follow.
- Be specific about what AI does and does not do. "Our assistant answers from our help pages; bookings and refunds are handled by our staff" is more reassuring and more accurate than "we use AI responsibly".
- Name what a person does. Where a person reviews or approves, say so, but only where it is true for every case the answer covers.
- Point to the privacy route for data questions. Give the facts the register records and say how to make a request, rather than answering a rights request in a chat.
- Avoid absolute promises. "We never use AI" and "your data is never shared" are rarely true for long; describe what happens now and keep the reply's review date short.
- Keep the replies in step with the register. A saved reply's owner should be told when the register row behind it changes, the same way the saved replies tutorial ties replies to policy changes.
The assistant needs special care, because it will be asked about itself. The AI in customer service guide gives the rule the disclosure guide sets: say it is an automated assistant at the start and whenever asked, and offer a route to a person. Both guides place that rule in the European transparency direction, as the Commission's overview describes it, and in consumer protection guidance on impersonation, without drawing legal conclusions. Make sure the assistant's answers about its own nature, and about what happens to what customers type, come from the same approved wording as the staff replies.
Questions that are really data rights requests
Some questions about AI are requests about Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. wearing different words. "What do you hold about me in your AI tools?", "Delete what I told your chatbot", "I don't want my information used for AI" and "Stop using my data to target me" each ask for something the GDPRThe EU regulation setting the rules for processing personal data, including its principles, lawful bases, individuals' rights and breach duties. gives people rights over, as the site's privacy pages describe it. These need the business's existing privacy route, not a saved reply.
The marketing opt-out procedure shows the pattern. A person does not have to use the right words; an objection to marketing in a social message is an opt-out, handled across every tool, and the site's consent guides record, following the UK regulator's direct marketing guidance and the GDPR, that the right to object to direct marketing is absolute. A request to delete data goes to the data protection owner, who decides what is deleted and what is kept. Questions about whether there is a Lawful BasisOne of the six grounds the GDPR accepts for processing personal data, such as consent, contract or legitimate interests, chosen for each purpose. for a particular use of customer data in an AI tool are also for that owner, and the privacy and consent guide explains how consent shapes what tracking collects in the first place.
Train staff and configure the assistant to recognise these questions and hand them on with the customer's words recorded, rather than reassuring and closing the conversation. The saved reply for such questions confirms the request has been passed on and says who will reply.
When a question reveals a problem
Sometimes the question is the first sign that something is wrong. A customer asks whether a product photo was generated because the product does not look like the picture. A customer asks why the chat assistant claimed to be a person. A customer asks why an AI tool knew their order history. Each points to an undisclosed use, a misleading asset or data where it should not be, and each needs more than a reply.
Route these questions to the register's owner as incidents, not contacts. The guide to handling an AI mistake sets the response: contain the source, correct what customers saw where they saw it, tell the people who need to know, trace the cause and log it so the system changes. If the customer is dissatisfied, the question is also a complaint, and the complaints procedure tutorial applies: log it, investigate and respond as with any other complaint.
When the question is public, on social media or in reviews, the answer is visible to everyone who reads the thread. If it gathers attention, the social media crisis plan sets the route, including a Holding StatementA brief first public response in a crisis that acknowledges the issue, tells affected people what to do now and says when more will follow. that acknowledges the question and says when a fuller answer will follow, without guessing at what happened. The fuller answer must be true once the trace is done, which is another reason never to deny a use before checking the register.
Finally, close the loop. If staff have been giving an answer that turns out to be untrue, correct the saved reply, tell the staff, and consider whether customers who were told the wrong thing need to hear the correction. The AI governance hub places this routine alongside the policy, the register and the other governance guides it draws on.
An opticians chain prepares its answers (illustrative)
Feldquin Opticians collects its customers' AI questions and corrects one answer
Feldquin Opticians, an invented chain of high-street opticians with 9 branches, asks its branch staff, its social media manager and its service team to log every customer question about AI for a quarter. They collect 46 questions.
| Kind of question | Count | Prepared answer |
|---|---|---|
| Is the website chat a person? | 17 | Yes, it is automated; how to reach the branch team |
| Did my details go into an AI tool? | 11 | What the register says; how to make a data request |
| Are frame photos edited or generated? | 8 | Photos are of real frames; backgrounds are edited |
| Do you use AI to write messages to me? | 6 | Which messages are drafted with AI and who checks them |
| Please delete what I told the chat | 4 | Passed to the privacy lead as a data request |
Checking each answer against the register turns up a problem. Branch staff had been telling customers that appointment reminders were written by the branch team, but the register shows that a feature in the booking system has drafted reminder wording for the past 7 months, with a member of the central team approving each template.
The service lead corrects the saved reply, which now explains that reminder wording is drafted with an AI feature and approved by a person before it is used. The 4 deletion requests and 3 of the 11 data questions, which asked for a copy of what was held, go to the privacy lead through the existing request route. The register owner logs the wrong answer as an incident, briefs every branch and sets the saved replies' review to follow any change in the register.
Common mistakes
- Answering from memory. Staff forget the AI features built into everyday tools; the register is the source of the facts.
- Denying a use to reassure. An untrue "no AI here" is a statement by the business, and customers find out.
- Letting the assistant dodge the question. When asked, it says it is automated and offers a person.
- Treating data requests as chat. Questions about what is held, deleting it or stopping its use go through the privacy route, with the customer's words recorded.
- Vague reassurance. "We use AI responsibly" answers nothing; say what the tool does and what a person does.
- Separate answers in separate channels. The assistant, the service team and the social team give the same facts from the same saved replies.
- Closing a question that revealed a problem. An undisclosed use, a misleading image or misplaced data is an incident, not a contact.
