I build AI tools for small businesses, which makes this an odd article to write. But I turn down more of this work than I take on, and the reasons are consistent enough to be worth writing down. If you're about to spend money on any of the following, read this first and then decide.
A generic chatbot on your website
Note the word generic, because this one needs a distinction the sales pitches never make.
The off-the-shelf widget — trained on nothing in particular, dropped into the corner of your site, capable of little beyond restating your homepage — is the most requested and least justified thing on this list. It usually exists because the website didn't answer the question in the first place.
Think about what actually happens. Someone lands on your site wanting to know whether you service their area and roughly what a job costs. If the site says so, they've got their answer in four seconds, and Google can read it too. If it doesn't, you're now asking them to have a conversation with a machine to extract information you could simply have written down. That version costs more, converts worse, and does nothing for your search visibility.
But an assistant that actually knows your business is a different animal, and I build those, because done properly they're one of the better things available right now. The difference is what it's reading from. A real one is connected to your own material — your services, your process, your policies, your stock down to the spec of an individual item — held in a searchable index that updates itself as that material changes, so it isn't quietly six months out of date by winter. It answers in your voice rather than a chirpy default. It says it doesn't know instead of inventing something. And it hands off cleanly, capturing the details and booking a callback rather than leaving someone stuck at 11pm on a Sunday.
That's a genuinely useful thing to own. It is also not what most people are being sold when they're quoted for a chatbot.
Instead: answer the ten questions you actually get asked, properly, on the relevant pages. Then, if you want an assistant on top of that, build the kind that knows something — sitting over a clear site rather than compensating for an unclear one.
Generating your website copy
This is the one I argue hardest about, and it costs me the occasional job.
Your website's job is to sound like the person the customer is going to deal with. That's most of what they're deciding when they read it — not whether you're competent, which they assume, but whether they want you in their house or handling their money. Generated copy is, by design, the statistical middle of everything already written on the subject. It sounds like every competitor's generated copy, because it came from the same place.
There's also a practical problem. Once a site is written that way, nobody in the business can maintain it, because there's no voice to match. It calcifies.
Instead: talk, and have someone shape it. Record yourself explaining what you do to a friend who doesn't work in your industry. The transcript of that is better raw material than any prompt, and shaping it is a job AI is genuinely good at.
Automating something two people do differently
A quieter one, and a more expensive mistake. If two people in your business do the same task in different ways and both think they're right, automating it doesn't resolve that disagreement. It picks one at random, makes it permanent, and buries it in software where nobody can see the decision was ever made.
I've been called in to fix the aftermath of this more than once. The tool works exactly as specified. Half the team quietly stopped using it in week three.
Instead: agree the process on paper first, with the people who do it. If you can't get agreement, that's the actual project, and no amount of software substitutes for it.
An AI strategy
For a business under about twenty people, a document titled "AI Strategy" is almost always a way of feeling productive without deciding anything. It'll be out of date before it's finished, and nobody will read it twice.
Instead: pick the single most annoying repeated task in your week and deal with that one. Then do it again in three months. That's the whole strategy, and it has the advantage of producing results between meetings.
Anything sold to you with a deadline
If the pitch involves a window closing, a competitor threat you hadn't noticed until they mentioned it, or a discount expiring on Friday, the urgency is the product. Genuinely useful tools don't need it, because the thing they fix was already annoying you before anyone rang.
Instead: wait a fortnight. If it still seems like a good idea when nobody's pressuring you, it might be.
Buying tools before you have somewhere to put the output
The subscription trap. A tool produces something — summaries, drafts, transcriptions, leads — and there's no agreed place it goes and no person whose job it is to use it. So it accumulates, and then it stops being opened, and eleven months later it's still on the credit card.
Instead: before buying anything, answer this: what does this produce, who looks at it, and what do they do differently as a result? If any of those three is vague, don't buy it yet.
What's left
Quite a lot, as it happens. Stopping information being typed twice. Turning photographed paperwork into records. Making your own files and past jobs searchable in plain English. Being findable now that a big share of searches get answered without a click. Those are unglamorous, specific, and they pay.
The pattern is that the things worth doing tend to be boring and narrow, and the things being marketed hardest tend to be broad and exciting. That's not a coincidence. Broad and exciting is easier to sell to someone who doesn't yet know what they need.
