When Not to Use a Chatbot (From Someone Who Sells Them)

I build AI chatbots and voice agents for a living. I also talk people out of them regularly, because the failed builds all look alike and I would rather not be the one holding the invoice when the client realizes it.

This is not false modesty to make the pitch land better. These are the specific conditions under which I either decline the work or scope it down to something much smaller. If you recognize your business in several of them, spend the money somewhere else.

When nobody will own it

This is the big one. A chatbot is not a website you launch and forget. It is closer to an employee than to a brochure. Someone has to read transcripts, notice when it starts giving a wrong answer, update it when your prices or hours change, and respond when it breaks.

If you cannot name the person who will spend an hour a month on it, do not build it. What happens instead is predictable: month one it is great, month four it is quoting last year's pricing, month seven a customer screenshots a bad answer, and it gets switched off. You paid for the build and got seven months of slowly increasing risk.

The cheapest version of this test: ask who reads the transcripts. If the answer is a shrug, that is your answer.

When your information does not exist yet

An AI agent can only be as good as what it knows. If your pricing lives in the owner's head, your service area is "wherever we feel like driving," and your policies vary by customer, there is nothing to build a knowledge base from.

You will spend the whole project trying to extract facts from someone who has never written them down, and the result will hedge on everything because you had to tell it to. That is a bot that answers every question with "let me have someone get back to you," which is a voicemail with extra steps and a monthly bill.

Write the FAQ first. If writing the FAQ turns out to be hard, that is the actual project, and finishing it will improve your website, your ads, and your staff training whether or not you ever add a bot.

When the stakes of a wrong answer are high

Models make things up. Not often, if you build carefully, but the rate is never zero, and confident wrongness is the failure mode.

I once had a receptionist agent tell a caller, in a completely convincing tone, that it had blocked their number. It had no such ability. Nobody had told it what it could not do, so it filled the gap. The fix was an explicit list of abilities it does not have, and I put one in every agent now. But the lesson generalizes: the model will always be willing to be helpful about something it cannot do.

So if a wrong answer means a medical decision, a legal position, a binding quote on a six-figure job, a dosage, a compliance filing, or a safety instruction, do not put a generative agent in front of it unsupervised. You can still use one for triage, intake, and routing, where the worst case is a human gets a slightly imperfect summary. Draw the line at anything the customer might act on directly. More on containment in stopping chatbot hallucinations.

When your volume is too low to matter

Run the arithmetic honestly. If you get eight inbound inquiries a week and you answer six of them, the bot's job is capturing two. Two leads a week may well justify it for a remodeler with a $20,000 average job. It does not justify it for a business with a $60 ticket.

I would rather sell that business missed-call text-back, which costs a fraction, requires no maintenance beyond reading replies, and addresses the actual gap. The full agent can come later if the volume ever arrives.

The related trap is building a full conversational agent for a channel nobody uses. If your website gets 200 visitors a month, a chat widget will produce roughly zero conversations, and you will conclude AI does not work when what actually happened is you had no traffic.

When a form or a phone tree would do

Some interactions are genuinely linear. Three fields and a submit button. If the entire job is "collect name, email, and what you need," a form does that faster, more reliably, at zero marginal cost, and without any risk of saying something strange.

The test I use: is there real branching, or ambiguity in what the person wants? If a caller might want any of six different things and cannot be expected to know your internal vocabulary, an agent beats a menu badly. If everyone wants the same thing, build the form.

Same on voice. If your call flow is genuinely "press 1 for sales, 2 for service," a plain IVR is cheaper, faster, and more predictable than an AI agent.

When your customers do not want it

Two versions of this. First, demographics: some customer bases simply will not text, and some will hang up on any automated voice regardless of quality. You know your customers. Believe them over the vendor demo.

Second, relationship businesses. If your differentiator is that a real person who knows the customer picks up the phone, putting a bot in front of that is spending your entire advantage to save a few hours a week. High-touch, high-value, repeat-relationship businesses often should keep the human on the phone and use AI behind the scenes instead: summarizing calls, drafting follow-ups, prepping the human before they dial.

When the integration does not exist

A lot of the value in these systems comes from actions, not conversation. Booking, lookups, order status, account changes. If the system that holds that data has no API, no webhook, and no export, the agent cannot do the useful part.

I have taken projects where the plan was "the bot will check order status" and discovered the order system was a spreadsheet on one person's desktop. You can still ship something, but be honest that it is an FAQ machine, and price and scope it that way. See build vs buy for how integration constraints should drive the platform choice.

When you are buying it to avoid a hiring decision

Sometimes the real problem is that the business needs one more person and the owner does not want to hire. A chatbot will absorb some of that, and then the underlying capacity problem surfaces somewhere else: jobs booked that nobody can service, leads captured that nobody calls back, a queue of qualified prospects going cold in an inbox.

Capturing more leads than you can serve makes your reviews worse, not better. If the constraint is delivery capacity rather than lead capture, automation makes the problem arrive faster.

When you cannot test it properly

Voice especially. There is a category of bug that only appears on real calls, because a simulator cannot inject real system state like caller ID. I test every agent by calling and texting my own cell before any rollout, and I still find things. If your situation makes real end-to-end testing impossible (no access to the phone system, no test environment, a client who will not let you touch the live line), you cannot responsibly launch, and you should say so rather than shipping and hoping. The full discipline is in testing voice agents.

What to do instead

If several of the above apply, here is where I actually point people, roughly cheapest first:

• Missed-call text-back. Lowest effort, clearest payback, minimal maintenance.
• Write the FAQ and fix the website. Half of what a chatbot answers is a question your site should have answered.
• Better call routing. Sometimes calls are missed because forwarding is wrong, not because you need AI. Check phone forwarding before buying anything.
• AI behind the counter. Call summaries, follow-up drafts, transcript search. Zero customer-facing risk.
• Live chat with humans. If your traffic is low enough that a human can handle it, they will do it better. See AI chatbot vs live chat.

Then revisit the agent in six months, when you have the FAQ written, the traffic to justify it, and someone whose job includes reading transcripts. The build will go better and cost less, because most of what makes these projects expensive is the missing groundwork, not the technology.

FAQ

Are you saying AI chatbots do not work?

No. I run them in production and they book real jobs, answer real calls, and save real hours. I am saying they are operational systems with ongoing costs and failure modes, not one-time purchases, and a meaningful share of businesses that want one are not set up to run one yet.

What is the single biggest predictor of failure?

Nobody assigned to read transcripts. Every dead deployment I have seen traces back to that. The technology holds up fine; the neglect is what kills it.

Can I start small and see if it works?

Yes, and that is the right move. Put a bot on one narrow job (after-hours calls only, or one page of your site, or missed-call replies) and measure it for 60 days. Narrow scope also means a wrong answer has a small blast radius while you are learning.

How do I know if my industry is a bad fit?

It is less about industry and more about whether wrong answers are dangerous and whether your information is written down. Clinics can use them well for intake and scheduling while keeping clinical questions strictly off limits. The line runs through the use case, not the vertical.

What does a realistic budget look like if I do go ahead?

Setup in the hundreds and a monthly cost from about $99 up to a few hundred for a small business, plus usage. If that does not clear the value of the leads you are currently missing, the answer is no. The cost guide breaks the numbers down properly.