Week one defines the KPI. After that only that one counts.

Most industrial AI programmes are measured at the end. By then the number that would have proven the case was never instrumented, and the review turns into a discussion about the discussion.

The closing meeting nobody plans for

I have sat in that meeting more often than I would like. The model works, the team is proud of it, and then the CFO asks what it changed. From that moment the room is looking for a number. Someone offers hours saved, estimated by the people who saved them. Someone else offers a comparison against a quarter that had a different product mix. Twenty minutes later the discussion is no longer about the AI, it is about whether the measurement is fair.

That discussion cannot be won, because it was lost in week one. The failure is not analytical, it is contractual. Nobody agreed, before the work started, what would count as proof.

Two numbers that describe the state of the industry

McKinsey's State of AI survey found that fewer than one in five organizations track well-defined KPIs for their gen AI solutions, and that more than eight in ten see no tangible effect on enterprise level EBIT.

I read those two findings as one sentence. The value is not missing because the technology is weak. It is missing because almost nobody agreed in advance what value would look like, and an effect that was never instrumented cannot appear in a P&L.

The rule: name the number the client already has

In week one, before a line of code, the engagement names one number, and it has to be one the client already tracks: scrap rate, unplanned downtime, first pass yield, order intake per seller, days from quote to order, cost per tonne. Not a metric invented for the project.

The test is simple. Can you name the report the number appears in? Can you name the person who has to explain it when it moves the wrong way? If either answer is missing, it is not a KPI, it is an intention.

A number that already has an owner, a history and a report brings three things a project metric never has: a baseline that predates us and cannot be argued away, a definition the organization has already fought about once, and an audience that reads it every month whether we are there or not.

What that single decision changes

It forces the use case to be specific enough to move a number. "Improve quality" survives any workshop. "Reduce scrap on line four by fifteen percent by March" does not survive a vague use case, and killing a weak case early is the cheapest thing that can happen to it.

It gives the adoption owner something to defend in their own operating review. That is the difference between a project that is tolerated and one that is protected.

And it makes the closing conversation short, because the number is either there or it is not.

The objection: the data is not clean enough

This is the most common one, and it is usually true. It is also the point. If the number cannot be measured today, the first milestone is to measure it, not to build a model on top of a blind spot.

From my own experience, more programmes have died on an unreliable baseline than on a weak model. Instrumentation work looks bad on a slide and decides everything that follows. More than once the baseline itself surfaced something before a model had touched the data at all.

If the number cannot be measured today, the first milestone is to measure it, not to build a model on top of a blind spot.

The second objection: one number is too crude

It is, and that is deliberate. One lead number, plus at most two guardrails: one for quality or safety, one for cost. Three lines, not a scorecard with fourteen.

I have never seen a scorecard with fourteen entries carry a decision. A board that size is what an organization builds when nobody wants to stand behind a single line. Every additional metric is one more place to hide.

What we actually do in week one

Four things, and they fit on one page.

The baseline is pulled from the client's own system, not from ours, and it is signed off by the person who owns that system.

The target is stated as a delta and a date, not as a direction.

The owner for adoption is named. A person, not a function.

The stop criterion is written down at the same time as the target. If the number has not moved by the agreed date, the programme stops or changes, and that is decided now, while nobody has anything to defend yet.

Clients hesitate at the stop criterion, and it is exactly what makes the rest credible. A programme that cannot fail cannot prove anything either.

The number is not a report. It is the agreement.

Everything else in an engagement is negotiable: scope, sequence, architecture, who does what. The number is not. It is the one thing agreed while everyone was still calm, and it is what the closing conversation is measured against.

That is why week one is the most expensive week to get wrong, and the cheapest to get right.

Andreas Geiss

Founder, Axiva Industrial. Operator notes from the shop floor and from P&L experience.