How to choose your first AI project

There is a pattern we see repeatedly. A business decides to "do something with AI", picks the most visible and ambitious idea in the room, spends four months on it, and ends up with something that half works and nobody trusts. The organisation concludes AI is overhyped and shelves the whole subject.

The idea was rarely the problem. The sequencing was.

The four questions

Before committing to anything, put the candidate project through these four questions. A good first project answers yes to all four.

Is it frequent?

Automating something that happens weekly is a hobby. Automating something that happens fifty times a day is a business case. Frequency is what turns a small time saving into a meaningful one, and it is also what gives you enough examples to tell whether the thing is working.

Is the failure mode survivable?

Ask what happens on the day it gets something wrong, because that day will come. If the answer is "a customer receives a slightly awkward email", fine. If the answer is "we send the wrong medication information", absolutely not. Start where mistakes are cheap and visible.

Can you tell whether it worked?

You need a number you can look at before and after. Calls answered. Hours spent. Quotes issued within 24 hours. Enquiries converted. If the only available measure is a vague sense that things feel smoother, you will never be able to justify the second project.

Does somebody actively want it?

Find the person who currently does the tedious task and hates it. Their enthusiasm will carry the project through the awkward middle weeks. A project sponsored only by the person who read an article about AI on a flight tends to stall.

The traps

Starting with the hardest problem. Your most complex process is the worst possible place to learn. You will be debugging the technology and the process simultaneously, and you will not know which is at fault.

Boiling the ocean. "AI across the business" is not a project, it is an aspiration. Pick one process. Finish it. Then pick the next.

Buying a platform before you have a problem. Tool-first thinking produces expensive shelfware. Understand the process you want to change, then choose the technology.

Ignoring the people. If the team believes this is the first step towards replacing them, it will fail, quietly and thoroughly. Be direct about intent from day one. In smaller businesses the honest answer is almost always about capacity rather than headcount, and people can tell when you mean it.

What a good first project looks like

In our experience the sweet spot has these characteristics: it takes six to ten weeks, it costs less than the annual salary of the person currently doing the work, it touches one clear process, it produces a number that visibly improves, and at least one person in the business is genuinely delighted by it.

That last criterion sounds soft. It is not. Internal enthusiasm is the fuel for everything that follows, and you only get it by making somebody's working week noticeably better.

The order that works

Automate the boring thing. Measure it. Tell everybody. Then take on something more ambitious with a team that now believes it will work.

It is slower for the first three months and dramatically faster for the next three years.