Where AI does not help, and what to do instead.

Read this one sceptically

We are an AI consulting studio. Every situation below is one where we are telling you not to buy the thing we sell, which means you should weigh this chapter differently from the rest of the guide — a refusal from a seller is either genuine or a technique, and you cannot tell from tone.

So here is the test. Each of the six comes with what to do instead, and in five of them the alternative costs you either nothing or a subscription. If the “instead” kept turning out to be a smaller thing to buy from us, you would be reading a technique.

The six

1. The job is small

This is the most common one by a distance and it is pure arithmetic. If the job costs less per year than a project costs to do, no amount of cleverness rescues the deal. The previous chapter’s calculator exists to settle this in about ninety seconds.

Instead:buy a subscription. Most small businesses using AI are already doing exactly this — 63% of them spend between one and forty dollars a month, measured from real payments across 4.6 million businesses by the JPMorganChase Institute and published on 14 April 2026. That is not a beginner tier. For a large share of businesses it is the correct permanent answer.

2. The process is not written down anywhere, including in your head

If three people in your business do the same job three different ways and all three are sometimes right, there is nothing stable to automate. Anything built on top will encode one of the three versions and then argue with the other two for a year.

Instead: write the process down. Not a manual, one page. This costs an afternoon, needs no consultant, and is worth doing whether or not anything else follows. A surprising number of the problems that arrive at our door dissolve at this step, which is a poor business model for us and a good outcome for the business.

3. The real problem is that the work should not be happening at all

A report nobody reads. A form that exists because a customer complained in 2019. A double-entry step that survives because two systems were never joined. Automating one of these makes a pointless thing happen faster and more reliably, which is worse than the current situation because now it is invisible.

Instead: stop doing it for a month and see who complains. Free, immediate, and reversible. If nobody complains, you have just removed the whole cost rather than a fraction of it.

4. The thing that is hurting is judgement, not volume

The evidence here is unusually clean. Seven randomised field experiments at one online retailer, all measured the same way, found a pre-sale chatbot lifting outcomes 16.3% while AI-written advertising titles came out 4.5% down and statistically indistinguishable from zero (Fang and colleagues, September 2023 to June 2024). Same company, same technology, same method. One narrow high-volume task moved; one judgement-shaped task did not.

And it can go further than not helping. In a randomised trial with 640 small entrepreneurs given an AI business mentor, the average effect could not be distinguished from zero, the strongest performers gained about 15%, and the weakest performers came out about 8% worse (Otis and colleagues, Harvard Business School, 2023). If the underlying judgement is shaky, a confident assistant amplifies the shakiness.

Instead: the honest answer here is often a person, or training, or a decision that has been avoided. That is not a thing we sell.

5. Nobody is going to change how they work

Factories bought electric motors in the 1890s and bolted them to the drive shaft they already had. The productivity took about forty years to appear, and it appeared when the factory floor was rebuilt around the motors rather than the other way round. We wrote that argument out at length in a research paper of ours called Nobody will use it, and it is not the training.

The modern version is that the new way is installed beside the old way, the old way is never switched off, and optional always loses to habit. And it is not always passive: 29%of knowledge workers admitted actively undermining their employer’s AI strategy, rising to 44% among the youngest, in a survey of 2,400 workers reported by Fortune on 30 July 2026.

Instead: settle the question of who is switching the old way off, and what happens if they do not, before anything is built. If the answer is nobody, do not build it. A tool that half the team routes around is worse than the manual process, because now you are paying for both.

6. You cannot tell afterwards whether it worked

If there is no count of the thing before, there will be no count of the thing after, and you will be left deciding with a feeling. The METR trial in the previous chapter is the warning: expert developers were 19% slower with AI and believed they had been 20% faster. A 39-point gap between the feeling and the fact, among experts.

Instead: count it for two weeks first. By hand, on paper if necessary. Two weeks of tally marks is a better foundation than any dashboard bought later, and it costs nothing but attention.

The pattern underneath them

Five of the six are not really about AI. They are about whether there is a stable, countable, genuinely necessary job with someone accountable for changing how it is done. Where that exists, tools help. Where it does not, tools get bought and then quietly stop being used, and the story becomes “we tried AI and it did not work for us” when what was tried was something else.

Almost every AI project that dies was, underneath, a management question that got answered with a purchase.

When the answer is yes

For balance, the shape that does work, from the same body of evidence: a narrow task, done many times, currently done by someone who is not a specialist in it, where the output can be checked, and where one named person is accountable for switching the old way off.

That is the shape where customer support agents resolved 14% more issues per hour and the least experienced among them resolved 34% more (Brynjolfsson, Li and Raymond, across 5,179 agents). Narrow, repeated, checkable, with someone responsible.

If you read the six above and none of them landed, the next question is whether we are the right people, which is a different question and has an honest answer that is sometimes no. That is the next chapter.