Before You Recommend It

A logistics client once asked me to evaluate an AI tool that promised to cut their dispatch planning time in half. The demo was flawless. The sales engineer ran a sample route, the model reorganised it in eleven seconds, and everyone in the room — including me, for about ninety seconds — was ready to sign.

Then I asked to see it run on last Tuesday's actual data: the one with three trucks down for maintenance, a driver who had called in sick, and a customer who had moved their delivery window twice. The tool choked. Not dramatically — it just quietly produced a plan that assumed all three trucks were available, because nobody had told it otherwise, and nobody had built a way to tell it otherwise.

The gap was not in the model. It was in the twenty minutes of unglamorous setup nobody had budgeted for.

Every vendor pitch is optimised to answer one question: can the model do the thing? That is rarely the question that decides whether the tool helps.

The question that matters is whether the tool can do the thing inside this organisation — with its data quality, its handoffs between departments, and its tolerance for being wrong in public. After twenty-six years of technology consulting I have watched three or four hype cycles come and go, and the pattern repeats: a genuinely useful capability arrives, the market gets breathless, and a meaningful fraction of early adopters buy for reasons that have nothing to do with fit. AI is no different from ERP or cloud migration in this respect. It is just faster, and the mistakes are cheaper to make and more expensive to unwind.

So before I tell a client to move forward with anything, I run the same five checks. None of them are exotic. All of them get skipped constantly, usually because the sales cycle rewards speed and nobody in the room is paid to slow it down.

Below they are seven questions, because two of the five have a second half that is easy to answer confidently and wrongly. Answer them about one tool you are evaluating right now — not about AI in general, which nobody can evaluate. The verdict at the end is not a grade. It is a list of what you would still have to justify if someone senior asked you tomorrow why you signed.