Should You Tell Customers It's AI? The Data Says Yes, and Says Something More Important

Disclosure, up front

We build AI answering and chat systems, so we benefit if you deploy one. This article argues that the deployment most likely to cost you customers is the one without a human escape route — which is a constraint on what we can sell you, and we would rather say it now than after. Survey figures here are third-party and, as section 03 shows, they contradict each other more than usual.

We have covered when the law requires you to disclose. This is the question owners actually ask, which is different and more nervous:

If we tell customers it's AI, will they mind?

The 2026 data gives an unusually clear answer, and it is not the one people expect. Disclosure is not what costs you customers. Trapping them is.

The short version

Roughly three quarters of customers say they would leave a brand that offers only AI, with no human alternative.
70% want a human for serious matters like fraud and security.
Comfort depends almost entirely on the task — around two thirds are uncomfortable with AI for medical or investment advice, while a majority are happy with it for ordering or returning an item.
General trust in AI is drifting down, from about 62% in 2023 to 59% in 2025, with those calling it "very untrustworthy" rising from 5% to 12%.
The disclosure figures themselves conflict, which is section 03.

01The finding that should shape the decision

Nearly three quarters of customers say they would leave a brand that uses only AI with no human alternative. That is a much larger number than anything in the disclosure research, and it points at a different design decision entirely.

The risk is not that customers discover they are talking to software. Most now assume it. The risk is that they need a person and cannot reach one. That is the experience that produces the review, the churn and the story told to other people.

Which reframes the deployment question. The thing to get right is not the wording of your disclosure — it is whether there is a visible, reachable path to a human, and how quickly it appears when someone wants it.

02Comfort is task-specific, not general

Asking whether customers accept AI is like asking whether they accept phone calls. It depends entirely what for.

Customer comfort with AI by task type
Task	Customer comfort	What to do
Ordering, booking, scheduling	High — around 65% comfortable	Automate freely. This is the win.
Returns and simple admin	High — around 59% comfortable	Automate, with an obvious exit.
Status, hours, availability	High	Automate. Nobody wants to wait on hold for this.
Fraud or security concerns	Low — 70% want a human	Route to a person immediately. Do not triage first.
Medical advice	Low — about 69% uncomfortable	Do not. Note this is also regulated in several places.
Financial or investment advice	Low — about 68% uncomfortable	Do not.
Complaints	Low	Acknowledge automatically; resolve with a person.

Comfort figures from 2026 third-party consumer surveys of AI in customer service; percentages vary between studies and are self-reported preference rather than observed behaviour. The right-hand column is our recommendation. Note the pattern: comfort falls as consequence rises, which is the same axis as our decision-rights framework.

The practical read: automate the transactional, route the consequential. Most small businesses have the split backwards, because the transactional queries are the ones staff find easy and the consequential ones are where a bot is cheapest to deploy.

03The disclosure numbers contradict each other

We looked for a clean figure on how many customers want AI disclosed and could not produce one honestly.

The same body of 2026 reporting contains both a claim that 87% of consumers say businesses should disclose AI use and a claim that only 54% want clear disclosure about whether they are talking to AI or a human. Those cannot both be describing the same thing. They are probably measuring different questions — an abstract principle versus a practical preference — but they are quoted interchangeably, and we could not resolve them to a primary source with a stated sample.

A third figure in the same material says 14% would lose trust in a business whose AI agent did not identify itself, which is a much smaller and more specific claim.

So we are not going to tell you that "87% of customers demand disclosure". What the evidence supports is narrower and more useful:

A minority actively lose trust when AI is concealed — small, but they are the ones who complain publicly.
Nobody is recorded leaving because a business was upfront. There is no measured downside to disclosure in any of this material.
The large, consistent number is about the human escape route, not about disclosure at all.

Why disclosure is cheap insurance

The downside of disclosing is unmeasured and appears to be near zero. The downside of being caught concealing is a specific, well-documented trust loss, and now in some jurisdictions a legal obligation as well. That is an asymmetric bet with an obvious side.

04What good looks like

Say it in the first message, plainly. "Hi, I'm an automated assistant — I can help with bookings and account questions, or connect you to someone." One sentence that discloses, sets scope, and offers the exit.
Make the human path visible from the start, not something you have to fight for after three failed attempts. The fighting is what generates the bad review.
Escalate on frustration, not just on keywords. Repetition, rising message length, or the same question twice are better signals than waiting for someone to type "agent".
Never let it imply it is human. No invented first name paired with a stock photograph. This is where a reasonable deployment tips into the thing people feel deceived by.
Set expectations about the human, honestly. "Someone will reply by tomorrow morning" beats a transfer to a queue that nobody is staffing.
Do not automate the apology. When something has gone wrong, the entire content of an apology is that a person took responsibility.

05The trend worth watching

General trust in AI is drifting downward — roughly 62% in 2023 to 59% in 2025 — and the share calling it "very untrustworthy" more than doubled from 5% to 12%. The middle is thinning while the sceptical end hardens.

For a small business that cuts a specific way. You are not a faceless platform, and the reason people use you is often that they can reach an actual person. If you deploy AI in a way that removes that, you are spending your main advantage to save money on the part customers valued most.

Used the other way round — AI handling the booking so a human is free for the conversation that matters — it strengthens exactly the thing you compete on.

06The honest summary

Tell them. The measured downside of disclosure is approximately nothing, a minority lose trust when it is concealed, and in some jurisdictions you are now required to anyway.

Then spend your attention on the decision that actually moves customers: roughly three quarters say they would leave a brand offering only AI with no human alternative. Comfort tracks consequence — high for bookings and returns, low for fraud, complaints, and anything medical or financial.

Automate the transactional, route the consequential, and make the way to a person obvious from the first message rather than a prize for persistence.

07Common questions

Should I tell customers they are talking to AI?

Yes. The measured downside of disclosing appears to be close to zero — no study we found records customers leaving because a business was upfront — while concealment carries a documented trust loss among a minority who are also the most likely to complain publicly. In some jurisdictions, including under the EU AI Act's transparency rules, disclosure is now a legal obligation as well. It is an asymmetric bet with an obvious side.

Will customers leave if we use AI?

Not for using it. Roughly three quarters say they would leave a brand that offers only AI with no human alternative, which is a much larger and more consistent finding than anything about disclosure. The risk is not that customers discover they are talking to software — most now assume it. The risk is that they need a person and cannot reach one, and that is the experience that produces the review and the churn.

Which customer service tasks are safe to automate?

Transactional ones. Around 65% of consumers are comfortable with AI for ordering and booking, and around 59% for returns and simple admin; status, hours and availability questions are similarly uncontroversial and nobody wants to wait on hold for them. Comfort falls sharply as consequence rises: 70% want a human for fraud or security concerns, and roughly two thirds are uncomfortable with AI for medical or investment advice.

How reliable are the statistics about AI disclosure?

Less than they look. The same body of 2026 reporting contains both a claim that 87% of consumers say businesses should disclose AI use and a claim that only 54% want clear disclosure about whether they are talking to AI or a human. Those cannot both describe the same measurement, they are quoted interchangeably, and we could not resolve them to a primary source with a stated sample. The figures about wanting a human alternative are much more consistent.

What should the AI say in its first message?

One sentence that does three jobs: discloses, sets scope, and offers the exit. Something like: I am an automated assistant, I can help with bookings and account questions, or I can connect you to someone. That satisfies disclosure obligations, manages expectations about what it can do, and puts the human path in view immediately rather than making it a prize for persistence.

When should an AI hand over to a human?

On frustration, not just on keywords. Repetition, rising message length, or the same question asked twice are better escalation signals than waiting for someone to type the word agent. Route immediately and without triage for fraud or security concerns, complaints, and anything medical or financial. And set honest expectations about the handover — someone will reply by tomorrow morning beats a transfer into a queue nobody is staffing.

Is it acceptable to give the AI a human name and photo?

No. This is the specific practice that tips a reasonable deployment into something customers feel deceived by, and it also undermines any reliance on the EU AI Act's exemption for cases where it is obvious the user is dealing with a machine. An assistant can have a name and a personality without implying it is a person.

Does using AI hurt a small business more than a large one?

It can cut either way, and the difference is design. Trust in AI is drifting downward — about 62% in 2023 to 59% in 2025, with those calling it very untrustworthy rising from 5% to 12% — and much of why customers choose a small business is that they can reach an actual person. Deploying AI so that it removes that access spends your main advantage. Deploying it so the AI handles bookings and a human is free for the conversation that matters strengthens it.

Send us your bot's first message

Paste what your chat widget or phone system says when someone first reaches it, and tell us how a customer gets to a human. We will tell you whether the disclosure is clear enough to rely on, whether the escape route is visible or buried, and which of your query types are on the wrong side of the automate-or-route line. Free, and the answer is sometimes that it is fine as it is.

Ask for a disclosure check

Sources, read 9 September 2026: 2026 third-party consumer surveys of AI in customer service, for the roughly three-quarters who would leave a brand offering only AI, the 70% preferring a human for fraud and security, the task-comfort figures, and the trust trend from about 62% in 2023 to 59% in 2025 with "very untrustworthy" rising from 5% to 12%. All are self-reported preference rather than observed behaviour, and percentages differ between studies. The conflicting disclosure figures discussed in section 03 are presented as we found them precisely because we could not reconcile them, and we have deliberately not quoted either as fact. The task table's recommendations and the design guidance are our own. Related: What an AI Receptionist Does When It Doesn't Know the Answer and The EU AI Act's August 2026 Rules.

Hero image from Unsplash, used under the Unsplash License.