Most businesses being sold an AI assistant right now should not build one, and we say that as people who build them. Not because the technology disappoints, it is genuinely good, but because a chatbot is an answer to one specific problem, and it is being sold as an answer to every problem. The mismatch is expensive, and the way it fails is worse than wasted money: it fails in front of your customers.
Here is the honest version, including the part that costs us work.
First, the case for
It would be easy to knock down a weak version of the argument, so let me put the strong one. A good assistant answers instantly at two in the morning. It never gets tired on the fortieth identical question. It scales to a busy Monday without anyone being hired. It gives consistent answers, which humans genuinely do not. And it frees your best people from work that bores them, which is a real thing and not a euphemism for cutting staff.
All of that is true, and where it applies the case is strong. The problem is that every one of those benefits depends on the same condition: lots of people asking the same fairly simple questions, which have written answers somewhere. Take that away and the benefits go with it, while the costs stay exactly where they were.
Do not build a chatbot if the volume is not there
This is the most common case by a distance. If you field a handful of enquiries a day, an assistant cannot save you meaningful time, because there is not meaningful time being spent. What it can do is add a layer between you and the small number of people who wanted to talk to you.
There is a rough test. If answering every repeat question yourself for a week would not seriously disrupt your week, you do not have a volume problem. You have a small number of conversations, and those are worth having personally, because at low volume the personal reply is a competitive advantage rather than a cost.
Do not build one if every question is different
Assistants are good at questions that have been asked before. They are poor at questions that require judgement about a specific situation, and they are worst of all at questions where the right answer depends on something they cannot see.
If your enquiries look like can you do this unusual thing for my unusual circumstances, the honest answer is almost always that it depends, and an assistant that says it depends is an assistant nobody uses twice. Consultancies, bespoke manufacturers, anyone whose value is in working out what the customer actually needs: your questions are the work, not an obstacle to it.
Do not build one if there is nothing to ground it in
An assistant answers from something. If that something does not exist, you are not really commissioning a chatbot, you are commissioning the documentation you have been avoiding for three years, with a chat window attached to the end.
That is a legitimate project and often a good one, but it should be priced and planned as what it is. When people skip the honest framing and build the assistant anyway, it answers from the general knowledge of a language model rather than from anything true about your business, and that is where the confident wrong answer comes from. Write the answers first. You may find that publishing them well solves most of the problem on its own, which is the outcome we most often recommend and least often get paid for.
Do not build one if nobody will own it
An assistant is not a thing you install. It is a thing that needs someone to read what it said, notice where it was wrong, and correct the material behind it. Weekly at first, then monthly.
Without an owner it degrades quietly. Your prices change and it quotes the old ones. You drop a service and it keeps offering it. Nobody notices, because nobody is reading the transcripts, and the first sign of trouble is a customer who is annoyed for a reason you cannot reconstruct. If you cannot name the person who will own it, and they cannot name the half hour a week, do not start.
Do not build one where being wrong is expensive
There are questions where a confident wrong answer is not an inconvenience but a real problem: anything medical, legal or financial, anything about safety, anything where the answer creates an obligation, anything a customer will act on without checking.
These systems can be constrained, grounded in your own material and made to refuse. They still get things wrong sometimes, and anyone who tells you otherwise is selling. So the question is not whether it will ever be wrong, it is what happens on the day it is. If the answer to that is we would have a serious problem, keep a human in the loop and use the assistant to draft rather than to send.
Do not build one to avoid your customers
This is the one that does lasting damage, and it is rarely the stated intention. It emerges anyway, because an assistant put in front of a support queue as a shield is optimising for deflection rather than resolution, and those look identical on a dashboard.
You know the experience from the other side. Everyone does. The loop, the reworded question, the eventual hunt for a phone number that is deliberately hard to find. The cost is not the interaction, it is what the customer concludes about how much you want to hear from them. Make the route to a person obvious and quick, on every single turn, or do not put an assistant in that queue at all.
What to do instead
| If the real problem is | Try this first |
|---|---|
| The same five questions, endlessly | Answer them properly on the page where people ask, rather than burying them in a help centre. Often removes most of the volume for the cost of an afternoon. |
| Nobody is available out of hours | A clear promise and a form that sets expectations honestly. People forgive waiting far more readily than they forgive being handled. |
| Enquiries arrive with half the information missing | A better form. Ask the three things you always end up asking, and the back-and-forth disappears without any AI at all. |
| Your team retypes the same answers | Saved replies, and an assistant that drafts for a human to send. All of the speed, none of the risk, because a person still presses send. |
| Nobody can find anything on your site | Search and navigation. Genuinely, this is the fix more often than a chat window, and it helps every visitor rather than the few who open the widget. |
Every row here is cheaper than an assistant and none of them need us. If one of them solves it, take it.
When it is genuinely right
So that this is not merely contrarian: the case where an assistant is clearly worth building looks like this, and if you recognise yourself here, stop reading and go and build it.
- Real volume, enough that repetition is measurably eating someone's week.
- Repetitive questions, where a large share of them are versions of the same twenty.
- Written answers that already exist and are true: a policy, a manual, a catalogue, a decent help centre.
- An owner, a named person with time in their week to read transcripts and correct what is behind them.
- A cheap failure mode, where the worst outcome of a wrong answer is a mildly irritated customer who then reaches a human quickly.
- An obvious escape hatch to a person, on every turn, that you would be happy to use yourself.
Five or six of those and it will probably pay for itself. Two or three and you are buying a project rather than a solution. If you want the numbers behind that judgement, what an AI chatbot costs sets out the bands, and AI for small business, what actually works covers the alternatives we more often recommend.
Not for you if
The reason to be this blunt is simple. The failures we get called in to fix are almost never technical. They are assistants built for a problem the business did not have, and the tell is always the same: nobody ever counted the questions first.