The short version: your AI chatbot is probably pretending to be human, and customers clock it within one exchange, which does more damage to trust than the mistake it's trying to hide. Tell people they're talking to AI, give it a clear "I don't know, let me get a person" exit, and check the transcripts every week. That single change fixes most of the reputation damage I see small businesses doing to themselves.
The chatbot mistake nobody talks about
Everyone writes about which AI chatbot to install. Nobody writes about the fact that most small businesses set theirs up to lie, and it's costing them reviews.
I'm not talking about the chatbot getting facts wrong, though that happens too. I'm talking about the deliberate choice, made by a founder or a marketing agency, to give the bot a human name, a chirpy tone, and instructions to never admit it's software. "Hi, I'm Amy, how can I help today?" Amy does not exist. And the customer typing to Amy usually works it out within about three messages, because Amy answers too fast, phrases things too smoothly, and then falls apart the second someone asks something slightly off-script.
That fall apart moment, the one where "Amy" starts looping the same canned line, is the moment you lose the sale. Not because the bot couldn't help. Because the customer now feels tricked on top of unhelped.
A real example from a client's site
Last year I worked with a small homeware business, family run, turnover around £900,000, based in the Cotswolds. They'd installed a chatbot on their product pages to handle delivery and returns questions, because their two-person customer service team was drowning in the same twelve questions on repeat. Smart move in principle.
The problem was how it was set up. The bot introduced itself as "Sophie from customer care," used exclamation marks constantly, and had no instruction for what to do when it didn't know the answer. So when a customer asked about a delayed order during a known courier issue, the bot cheerfully repeated the standard three-to-five-day delivery line four times in a row, ignoring what the customer had said. The customer left a one-star review that said, word for word, "spoke to Sophie who clearly isn't a real person and clearly wasn't listening either." That review sat on their page for six weeks before they even noticed it, because nobody was checking the transcripts.
We fixed three things. First, the bot introduced itself honestly: "You're chatting with our AI assistant, here to help with quick questions." Second, we gave it a hard rule: after two failed attempts to resolve something, it had to say "I think this needs a real person, let me get Claire or Tom to pick this up" and route it into their shared inbox with the full transcript attached. Third, someone read every transcript once a week, fifteen minutes, Friday afternoon. Within two months their chatbot-related complaints dropped from roughly one a week to almost none, and two customers praised the handover in reviews, which nobody expected.
The uncomfortable bit most people skip
Here's the part that makes marketing people uncomfortable: customers do not mind talking to AI. They mind being lied to about it. There's a difference between "this is a bot, and it's fast and useful" and "we tried to make you think this was a person, and it wasn't good enough to pull that off." The first is fine. Most people talk to AI dozens of times a week now without a second thought. The second is what generates the angry reviews, because it adds a layer of feeling foolish on top of not getting help.
The other uncomfortable truth is that most small businesses install a chatbot to look bigger than they are, not to serve customers better. That's not always a bad reason to do something, plenty of good tools start from a slightly vain motive. But if that's your real reason, at least be honest with yourself about it, because it changes how you should set the thing up. If you want to look established, invest in the escalation path and the tone, not the pretend name.
The three-question test before you switch a chatbot on
Before you launch or relaunch any AI assistant on your site, answer these three questions, on paper, not in your head:
- What happens when it doesn't know the answer? If your honest answer is "it makes something up" or "it just repeats itself," don't launch it yet.
- Who reads the transcripts, and how often? If nobody's name is next to that answer, nobody will do it.
- Does the bot ever say it's a bot? If the answer is no, change that this week, it costs nothing and it's the single biggest trust lever you have.
If you can answer all three cleanly, you're in better shape than most of the small businesses I audit.
Step by step: setting up the escalation
This is the part that takes the work, and it's worth doing rather than trusting the default settings in whatever tool you're using, whether that's Intercom Fin, Tidio, Zendesk AI, or a custom GPT wired into your site.
- Set a hard fail limit. Two unresolved attempts, three at most. After that, automatic handover, no exceptions written into the prompt.
- Write the handover line yourself. Don't let the platform default handle it. Something like: "I don't want to guess on this one, I'm looping in a real person from our team who'll reply within a few hours." Specific, honest, sets a time expectation.
- Route the full transcript, not a summary. Summaries lose the exact phrasing that upset someone. Your human team needs to see precisely what the bot said.
- Build the fifteen-minute Friday habit. One person, same day each week, reads every transcript from that week. This is the step almost everyone skips, and it's the one that catches problems before they become reviews.
- Tell the bot its own name is optional but its nature isn't. Call it Sophie if you like the branding, fine, but the very first message needs the word "AI" or "assistant" in it somewhere, in plain language.
None of this needs a developer. Most of it is prompt and settings work you can do in an afternoon. If you don't have the afternoon, or the confidence to write the escalation logic yourself, this is exactly the kind of small, contained project an AI consultant for a small business can set up in a single session rather than you guessing at it for a month.
What this costs you if you ignore it
I'd put a rough number on it based on what I see across client accounts: a badly configured chatbot that gets caught pretending to be human tends to generate one clearly negative, chatbot-referencing review for roughly every 150 to 250 conversations where something goes even slightly wrong. That doesn't sound like much until you remember that one specific, quotable one-star review, especially one that names the bot by its fake human name, gets read by far more prospective customers than the ninety-nine fine conversations that happened quietly around it. Reviews about being deceived get shared, screenshotted, and remembered in a way that "the chatbot was a bit slow" never does.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Compare that to the cost of fixing it: maybe two hours of setup, plus fifteen minutes a week ongoing. That's an absurd return on effort once you sit down and look at it side by side.
Where this connects to your wider marketing
The honesty principle here isn't really about chatbots. It's the same lesson Dale Carnegie was writing about long before anyone had heard of AI: people forgive a limitation stated plainly far more readily than they forgive a limitation they had to discover themselves. Brands that get this right build it into everything, not just customer service. Look at how Duolingo built an entire tone of voice around being upfront and a bit self-deprecating rather than pretending to be something it isn't, and it's a huge part of why people trust the brand enough to laugh at its notifications instead of blocking them.
The same thinking applies to any tool on your site that's doing the talking for you, not just chat widgets. If you've got an interactive calculator or quiz converting visitors on your site, the same rule holds: tell people plainly what it does and doesn't know, and they'll trust the result more, not less. And if you're tracking how people behave once they hit your chatbot or your forms, tools built the way Hotjar approaches behavioural data will show you the exact moment people bail out of a conversation, which is usually the moment the bot starts pretending.
A word on scale
One objection I hear a lot: "we're too small for a proper AI chatbot setup, we're not Revolut." Fair, you're probably not running support for millions of customers. But the principle scales down just as well as it scales up. Revolut's whole customer support model leans heavily on in-app AI and self-serve, and it works partly because they're transparent about what's automated and what isn't, and partly because the escalation to a human is fast when it needs to be. You can copy that exact philosophy with a two-person team and a £30 a month chatbot plan. The size of the business doesn't change the psychology of the customer on the other end of the chat window.
Even brands built entirely on trust and calm, like Calm, understand this instinctively: the moment a customer senses they're being managed rather than helped, the whole relationship shifts. An AI chatbot that's honest about being AI, and quick to hand off when it's stuck, protects that relationship. One that pretends does the opposite, and it does it fast.
What I'd do this week if I were you
Open your own chatbot right now, in an incognito window, and ask it something slightly awkward, something a real customer might ask that isn't in your FAQ. Watch what it does. If it makes something up, or loops, or never once mentions it's AI, you've found your task for this week. Fix the introduction line first, it takes five minutes. Then build the escalation rule. Then put fifteen minutes in your Friday calendar to read transcripts. That's the whole project, and it's the single highest-use AI change most small businesses can make this month, cheaper than almost anything else on your marketing to-do list and more likely to protect your star rating than another round of ad spend.
Frequently asked questions
Should my small business chatbot pretend to be a real person?
No. Give it a name if you like, but the first message should make clear it's an AI assistant. Customers generally don't mind talking to AI, they mind being deliberately misled about it, and that discovery is what turns a minor hiccup into a one-star review.
How often should someone check chatbot transcripts?
Once a week is the minimum that works in practice. Fifteen minutes, same day each week, one named person responsible. Checking "when we remember" means it never happens, and problems sit live on your site for weeks before anyone notices.
What should an AI chatbot say when it doesn't know the answer?
Something honest and specific, not a vague apology loop. For example: "I don't want to guess on this one, I'm getting a real person to reply within a few hours," sent along with the full transcript to a human inbox. Two failed attempts is a sensible hard limit before automatic handover.
Is it worth paying for help setting up an AI chatbot ?
If you don't have an afternoon to write the escalation logic and honesty rules yourself, yes, a short paid session with someone who does this regularly will cost far less than the reviews and lost sales a badly configured bot generates over a few months.
Related reading: Why Your AI Still Writes Like Everyone Else's (And the 45-Minute Fix That Changes That) and Why All Your Competitors' AI Content Sounds Exactly Like Yours.
