Should You Tell Customers When They're Talking to AI?

By Victor Hoang · Updated Oct 13, 2026

Every team that deploys an artificial intelligence (AI) assistant on its website, in WhatsApp or in a support inbox hits the same fork. Someone proposes giving the bot a human name and a friendly photo, and letting it chat without saying what it is. The reasoning is simple: people convert better when they think they're talking to a person, so why spoil it?

The honest answer is that the reasoning is partly right and still leads to the wrong decision. Hiding AI buys short-term conversion at the cost of trust when it's discovered, so disclosure plus an easy path to a human is the better commercial default, not just the compliant one. Disclosure does cost something in the moment. But the thing you're borrowing against is the relationship, and the bill arrives at the worst time: after the customer has told someone else.

This piece lays out the evidence on both sides, a way to think about the trade, what the main laws actually require (and what they don't), and a decision table you can hand to whoever owns your chat and support channels. It's written for the executive who has to approve the policy, not for the engineer who builds the bot. It's not legal advice, so have counsel check anything that touches your markets.

What Hiding AI Actually Buys You

Start with the strongest argument for concealment, because it's real. The best-known piece of research on this, a field experiment published in Marketing Science by Xueming Luo and colleagues, randomly assigned more than 6,200 customers to structured outbound sales calls from either chatbots or human workers. When customers didn't know, the bots sold about as well as proficient human salespeople and roughly four times better than inexperienced ones.

Then the authors told some customers up front that they were talking to a bot. Purchases fell by more than 79.7 percent, and calls got shorter. The authors attribute the drop to perception, not performance: people who learned the caller was a machine saw it as less knowledgeable and less empathetic, even though its results said otherwise. So the same words, from the same system, converted very differently depending on what the customer believed.

Key Facts

• A field experiment on more than 6,200 customers (Luo et al., Marketing Science, 2019) found that telling customers up front they were talking to a chatbot cut purchase rates by more than 79.7%. Disclosing later in the conversation reduced the damage.
• In a Five9 survey of 3,000 consumers reported in July 2026, 55% said they trust AI when offered a clear way to reach a human, against about one in four when no human option exists.
• Qualtrics reported in 2023 that 88% of consumers want to be told when they're dealing with AI, and 41% say they always want that disclosure.

Two cautions keep that result honest. It studied outbound sales calls at a single company, in a scripted setting, so it tells you about the direction and size of the first-contact penalty, not a number your inbound chat will reproduce. And the paper itself found the penalty shrank when disclosure came later in the call and among customers who had used AI before. That second finding matters more than the headline, and we'll come back to it.

So yes: if you only measure the first conversation, hiding wins. The question an executive should ask is what else is on the ledger.

The Ledger Nobody Reads: Borrowed Conversion

Here's a mental model for the trade. Call it borrowed conversion. When you conceal AI, the extra purchases you win aren't earned by a better offer or a better answer. They're borrowed against a belief the customer holds, "I'm talking to a person who understands me." Every borrowed sale carries a repayment condition: the customer must never find out, or must not mind when they do.

Three things determine how expensive the repayment gets.

Discovery probability. Customers find out, especially now that they expect an instant answer, as we describe in the death of the contact form. The bot repeats a phrase, answers at 3 a.m. with perfect grammar, can't remember what was said two messages ago, or the customer simply asks "Are you a bot?" Regulators have converged on the same point: a growing number of rules treat a direct question as a trigger for an honest answer. If your system is built to deny it, you've designed a lie into your product.

Stakes of the relationship. Owning the chat layer means owning its trust, a point we make in the CMO case for owning the chat layer. A one-time $20 checkout and a six-figure B2B contract have different repayment terms. In B2B, the buyer who feels deceived doesn't just leave. They tell their buying committee, their peers and sometimes the sales rep they were really hoping to reach. The longer the sales cycle, the more chances there are for discovery, and the more the original conversation gets re-read in hindsight.

What the customer does next. Concealment tends to work until the moment it fails, and at that moment the customer re-interprets everything the bot said. Claims, prices and promises that were fine a minute ago become suspect. That's the real cost of discovery: it doesn't just lose the deal in progress, it discounts the whole channel.

Now compare that to what disclosure costs. It costs a measurable first-contact penalty, which you can shrink with good design, as we'll see. It doesn't cost a repayment event. That asymmetry is the heart of the argument: a known, bounded, shrinkable cost against an unknown, lumpy, reputational one.

What Customers Say They Want

The survey evidence points one way, with the usual caveat that much of it comes from vendors who sell customer-service software. Qualtrics, a survey-software company, found the high level of demand for being told shown above. Five9, a contact-center vendor, found something subtler that matters more to the commercial case: 41 percent of consumers said they're less likely to use a company that uses AI for customer service, and that share rose to 53 percent when there was no option to reach a human. Four in five said they'd use AI-powered service when they knew a path to a live representative existed.

Read those together and the pattern is not "customers hate AI." It's "customers hate being trapped." The thing people punish isn't the machine. It's the absence of an exit. That's good news, because an exit is something you control, and it's cheap compared with the cost of lost trust.

It also reframes the Luo result. The penalty came when customers were told they faced a machine and had no sense of what they'd get in return. A disclosure that says "You're chatting with our AI assistant. It can answer pricing and setup questions right now, and you can ask for a person anytime" tells the customer what the machine is good for and gives them a way out. You've changed the story from "less empathetic stranger" to "fast tool with a safety net." Test that with your own audience, because no one has published a number for your exact wording.

The Disclose-Route-Prove Test

Here's a simple framework for a policy that survives scrutiny from customers, regulators and your own sales team. Three questions, in order.

1. Disclose: does the customer know, early and plainly? The disclosure should appear at or before the first message, in plain words, not in a footer or a terms page. "AI assistant" in the chat header counts only if it's visible and the bot doesn't contradict it. A human name and photo on a bot fails this test, even if the fine print says otherwise.

2. Route: can the customer reach a person without friction? One click or one sentence ("talk to a person") should trigger a handoff with the transcript attached, so the customer doesn't repeat themselves. State the real hours and the real wait: "A teammate replies within 15 minutes, 8am to 8pm." If your human coverage is thin, say so honestly. A truthful promise beats a clever one. This is the same discipline we described in why chat leads look weaker: the damage comes from how conversations are handled, and routing is where most handling fails. For inbound sales, the same rule decides whether an AI agent should take the first conversation at all.

3. Prove: can you show what the AI said and did? Keep transcripts, log the handoffs, and review a sample each month. If a customer or a regulator asks what the system told someone, you should be able to answer in minutes. Conversation records are also the richest source of insight you own, which we cover in customer conversations as underused data.

If you can answer yes to all three, you're not just compliant. You've built a channel customers have a reason to use twice.

What the Law Requires, and What It Doesn't

The legal picture varies by market and keeps moving, so treat this as orientation, not advice. The scopes below are narrow on purpose. It's easy to read them too broadly or too narrowly.

European Union. The EU AI Act's Article 50 requires providers of AI systems that interact directly with people to design them so the people are informed they're dealing with an AI system. There's an exception where it's obvious to a reasonably well-informed, observant person from the context, and a separate one for certain law-enforcement systems. The disclosure has to be clear and given at the latest at the first interaction. The European Commission's AI page says the transparency rules take effect in August 2026, and that guidelines on them were published in July 2026. The duty falls on the provider of the system, so if you buy a chatbot, your vendor has obligations. But you decide how it's configured and presented, so ask for the disclosure to be built in rather than assuming it is.

California. The state's bot disclosure law, in Business and Professions Code section 17941, makes it unlawful to use a bot to communicate with a person in California online while hiding that it's a bot, if the purpose is to mislead them about the communication in order to encourage a purchase or sale or to influence an election. A person isn't liable if they disclose it's a bot, and the disclosure must be clear, conspicuous and reasonably designed to inform. Note the scope: it targets deception tied to selling or elections, and it turns on disclosure being the safe route. A sales chatbot that poses as a person is squarely the situation it was written for.

Utah. Utah's AI disclosure rules, as amended in 2025, require disclosure when a consumer directly asks whether they're talking to AI, and in "high-risk" interactions such as those that collect sensitive personal information or involve financial, legal, medical or mental-health decisions. A supplier is protected if the AI clearly discloses at the start and throughout that it's not human. The provisions were extended to July 2027.

Notice what the three have in common: they all reward the same behavior. Say it clearly, say it early, answer honestly when asked. A company that builds that into the product once is compliant in all three places, and likely in the next ones too. We make a related argument about governance in AI compliance that survives regulation changes: build to the shared principle, not to each statute.

A Decision Table: Conversation Type, Disclosure, Human Route

Not every conversation needs the same treatment. Use this as a starting point and adjust to your risk appetite and markets.

A warning on the first row: don't give the bot a human persona and then rely on a footer. If the experience contradicts the label, the label won't protect you with customers, and may not protect you in front of a regulator.

Making Disclosure Cheaper

The first-contact penalty is real, so design against it. Lead with a result, not a label: "I'm the AI assistant. I can give you a price range and book a demo in two minutes" discloses and states a benefit in one line. Make the handoff feel like an upgrade, with the person arriving already knowing what was asked, answered and viewed. And measure honestly: split results by conversation type, reply time and whether the customer used the handoff, because a bot that resolves simple questions and passes hard ones to a person may convert fewer first chats and close more deals.

Disclosure matters most when the AI speaks directly to a customer, the customer might reasonably think it's a person, and a decision or a sale rides on the exchange. A package-tracking bot labeled as such isn't deceiving anyone. Concealment decisions also rarely get made by executives. They get made by a growth manager tuning a conversion rate, or a vendor default no one questioned, which is why this belongs in your governance policy. For where the gaps hide, see the governance gap in AI at work, and for a proportionate setup, how much AI governance a small company needs.

What to Do Next

1. Inventory every place AI talks to a customer. Website chat, messaging, email replies, voice, support. Include vendor tools where the AI is switched on by default.
2. Check what each one says about itself. Does it carry a label before the first message? Does it use a human name or photo? Does it answer honestly if asked "Are you a bot?" Fix any that deny it.
3. Write one disclosure line per channel. Plain words, a stated capability, and an offer of a person. Keep it short enough to read in a second.
4. Build the exit. One click or one sentence to reach a human, with real hours, transcript passed along, and a named owner for the queue. If you can't staff it, say when people will hear back.
5. Run a clean test. On a random half of traffic, run the disclosed version; compare meetings booked and deals closed, not just first-chat conversion, over at least a full sales cycle for your market.
6. Add a market check. Ask counsel which jurisdictions you sell into and whether a disclosure duty applies, particularly in the EU, California and Utah.
7. Review a sample of transcripts monthly. Look for denials, dead ends and customers who ask for a person and don't get one.

Customers don't need your AI to be invisible. They need it to be honest, useful and easy to step around. Companies that get that right will find the conversion cost of disclosure shrinks with design, while the cost of being caught never does. Choose the version of the trade where you only pay once.