Most AI companies selling to local business want you to buy a chatbot. It's the easiest thing in their catalogue to demo — looks impressive on a call, bolts onto any website in an afternoon, works for a plumber or a dentist or a wholesaler with the same script swapped in. That ease of selling is exactly the problem. It's built to be sellable, not to fix anything specific to your business.
I know, because I built one. For my own business, Markus IT Support. Here's what actually happened.
Strip the branding off and a website chatbot is, functionally, a glorified search bar wearing a friendly avatar. Everything it can tell a visitor is already sitting on the pages around it — opening hours, services, pricing, the contact form. It isn't adding information. It's adding a step: now someone has to type a question and wait for an answer, instead of scanning a page they could read in five seconds.
That's the whole case against it, before a single customer ever interacts with one. Then real customers did, and it got worse.
I built it expecting it to handle the obvious questions and free up my time. Instead, three things happened, consistently, that no demo ever shows you.
People tried to break it. Not maliciously — out of habit. The moment someone realises they're talking to a bot, a lot of them start testing it: odd phrasing, trick questions, seeing how fast it falls over. It fell over often, because it only knew what was on the site, and real questions rarely arrive in the shape a script expects.
It read as condescending. When it worked, it was just repeating back information the visitor had almost certainly already seen — which reads less like "helpful assistant" and more like "we assumed you hadn't looked." When it didn't work, it read as worse: a business that couldn't be bothered to answer properly, hiding behind a script.
It failed more than it helped. Genuinely useful resolutions were the exception, not the rule. And every failed exchange didn't just waste the visitor's time — it cost trust I'd otherwise have kept, from a customer who might have just phoned instead and gotten a real, if imperfect, human answer.
Net result: a fair amount of build effort, and a small but real amount of trust lost, for a feature that never once produced a lead I could point to and say "the chatbot got us that."
Across every enquiry I've ever taken — Markus IT Support, OptimisedLead, anything — has a single person ever said, unprompted, "I really liked chatting to your bot"? No. Not once. People do say they liked getting me on the phone, or a plain-English reply that actually answered their question. Nobody has ever praised the chatbot experience, because there isn't one to praise — there's just a wait to get to a real answer, or a real person, faster.
If a feature has never once been the thing a customer thanked you for, that's worth taking seriously before you build another one.
When I read an enquiry myself, I'm not just answering the literal question in it. I'm reading how specific it is, how urgent, whether it's actually a fit for what I do — I'm gauging the lead, not just responding to it. A script can't do that. It can only match a question to a canned answer and hope the match is close enough. It has no judgment, because judgment was never in the training data for "what should I say to this specific person, in their specific situation."
That gap — between answering and judging — is most of what "agents, not chatbots" actually means on this site. Agents can draft, check, and prepare. A human still decides. A chatbot skips straight past the deciding and calls the guessing "help."
Because it's a product, not a diagnosis. A chatbot is the same pitch regardless of what's actually slowing your business down — which makes it easy to sell to anyone and cheap to build once. That's a good business model for whoever's selling it. It's rarely the right fix for whoever's buying it, because it was never built around your actual bottleneck in the first place.
I've made this argument in general terms before, in the plain-English explainer on agents versus chatbots. This is the version with my own money and my own customers' patience spent finding out it was true.
I'll give the honest counter-argument, because I'd rather do that than pretend this is one-sided. A chatbot answers at 3am when you're asleep and nobody's picking up the phone. That's a real use case — some enquiries genuinely land out of hours, and a human can't be there for all of them.
One more scoping note, to be fair about what I'm actually arguing against. Not every chatbot is the same thing. Some go a lot further than mine did — reasoning past what's written on the site, pulling in live data, actually doing something rather than just answering — and those can add real value. They also cost more to build properly, more to keep correct, and still need the same human-checking-the-work discipline as anything else. What I'm talking about here is the generic chatbot product — the FAQ-matching kind — because that's overwhelmingly what actually gets sold to small businesses, not the deeper kind.
Even setting the deeper kind aside, I just don't think the generic version is worth what it costs. Weigh it up: a handful of out-of-hours enquiries, most of which can wait until morning anyway, against the trust you lose every time the bot gets it wrong, sounds condescending, or gets caught out by someone testing it. In my own experience, that trade wasn't close. If you've actually got a chatbot working for you without those problems, genuinely tell me — I'd rather be corrected by a real result than be proven right by default.
This is exactly why the Working Diagnosis starts with your actual bottleneck, not a product. If the honest answer is "you don't need an agent, you need a better FAQ page," I'll tell you that — same as I'd have told myself, a bit earlier, about the chatbot.
See how it works