The three questions to ask before you spend a penny on AI

Most AI money is wasted before a single tool is chosen. It is wasted at the decision about what to build, because that decision gets made on excitement instead of on three simple questions. Here they are.

When I run an audit, I am not really looking for the cleverest thing AI could do in a business. I am looking for the thing worth doing. Those are very different, and the gap between them is where budgets disappear. Before I recommend spending anything, every candidate process has to pass the same three questions. You can run them yourself.

Question one: does this process happen often enough to matter?

AI earns its keep on repetition. The value of automating something is the time it saves multiplied by how often it happens. A task that takes half a day but happens twice a year is almost never worth building for. A task that takes ten minutes but happens two hundred times a week is often a goldmine.

People get this backwards constantly. They point at the big, painful, occasional job and say "automate that", because it is the one that hurts. But the pain is not the point. The frequency is the point. Look for the small, dull, constant things first. That is where the hours actually are.

The best thing to automate is rarely the thing that hurts most. It is the thing that happens most.

Question two: is the input consistent enough to trust?

AI is brilliant at handling variation within a pattern, and hopeless when there is no pattern at all. So the second question is about the input to the process. Does the same kind of thing come in each time, in roughly the same shape, or is every case a genuine one-off?

Customer enquiries, invoices, applications, standard documents: these vary, but within limits, and that is exactly the territory where AI does well. A process where every single instance is truly unique, with no shared structure to learn from, is a poor candidate no matter how tempting it looks. If you cannot describe what a typical case looks like, the AI will not be able to either.

Question three: what does a mistake cost?

This is the question people skip, and the one that matters most. Every automated process will get some things wrong. The question is not whether it makes mistakes, it is what happens when it does.

If a mistake means a slightly awkward draft that a human reviews before it goes out, the cost is trivial and you should be aggressive. If a mistake means the wrong figure in a legal filing, or an offensive message sent straight to a customer with no human in between, the cost is severe and you need to build very differently, with checks and human sign-off in the loop.

Good implementation is mostly about designing around this third question. The right answer is almost never "let the AI do it unsupervised" or "do not use AI at all". It is "let the AI do the work and put a human at the point where a mistake would be expensive". Getting that boundary right is most of the job.

Putting the three together

The processes worth your money are the ones that pass all three: they happen often, the input is consistent, and a mistake is cheap or easy to catch. When a process clears all three, it is usually obvious in hindsight that it should have been automated years ago.

When a process fails one of them, it does not mean AI is useless there. It means you build with more care, or you wait, or you pick a smaller slice of the problem that does pass. And when a process fails two or three, the honest answer is to leave it alone and spend the money somewhere it will actually work.

That is the whole filter. It is not sophisticated, but it is the difference between AI spend that changes the numbers and AI spend that just felt like progress. When I do an audit, this is the lens I hold every opportunity up to before it earns a place on the list.