Where AI actually earns its place in a programme

Most of the “AI opportunities” I get shown aren’t AI opportunities. They’re process problems wearing an AI costume — a form nobody redesigned, an approval chain with four unnecessary steps, a spreadsheet doing the job of a proper data model. You don’t need a model for any of that. You need someone to sit down, map the as-is, and fix it.

Thirteen years of business analysis before I touched a line of Python is, if anything, an advantage here. The habit I built on regulated programmes — government, nuclear, energy trading — was to interrogate the problem before proposing a solution. That habit doesn’t switch off just because the solution on offer now includes machine learning.

The test I actually use, before I let a client near a model:

Is the judgement genuinely hard to encode as rules? If a decision tree or a lookup table would get you 90% of the way there, that’s what you should ship. It’s cheaper, it’s explainable, and nobody has to argue with a stakeholder about why the model said no.

Is there enough of the right data? Not “a lot of data” — the right data, labelled, representative, and not quietly biased by whatever process generated it. I’ve watched this stall more AI ideas than any technical constraint. Readiness assessment is unglamorous work and it’s usually the actual bottleneck.

Does getting it wrong sometimes matter less than getting it fast? Machine learning is probabilistic. If the cost of an occasional wrong answer is low and the cost of a slow answer is high, that’s a good sign. If a wrong answer is expensive or unsafe, you want a narrower, more deterministic tool — or a human in the loop that isn’t just theatre.

If a use case clears those three, I’m interested. If it doesn’t, the honest answer is usually “fix the process first” — and that’s not the answer people expect from someone who now builds AI tools for a living. It’s the answer that’s earned trust with every client I’ve worked with, though, because it’s the one that holds up six months later.