The first month goes well. The demo runs, the boss nods, there is applause in the meeting room. Someone starts talking about rolling it out to other departments.

In the second month someone asks: so how does my workflow change?

That is where the real problem begins.

The first month’s success is an illusion

The demo went smoothly because it ran on a curated dataset, operated by one highly motivated person, and nothing in the existing process had to give way for it. All three conditions disappear in month two.

I watched a warehouse picking-route project go through this. In the demo phase the model cut picking distance by 23% — a number so good it needed no explanation. Four weeks after rollout, the actual saving was 4%.

The gap was not in the model. The gap was that the floor supervisor had a sequencing method built up over ten years. He knew which racks jammed, which pallets were awkward, and when the forklift traffic peaked. The model knew none of it, and nobody had been assigned to feed that knowledge in — or to persuade him to change.

What fails is not the technology but the unowned cost of change

Every process change has a cost, and that cost always lands on a specific person. Project proposals describe expected benefits in detail. They rarely describe whose job gets harder for the first three months.

The second is the variable that decides the outcome.

• If nobody is assigned to absorb that cost, it quietly falls on whoever has the least standing to refuse — and that person will find ways to make the new process fail.
• If someone is assigned but lacks the authority to retire the old process, the result is identical.
• Adoption only happens when the person absorbing the cost and the person with authority to change the process are the same person.

This sounds like an organisational problem rather than an AI problem. It is. Almost every failed AI project is an organisational failure wearing a technical costume.

A better first question

The first question most teams ask is: how accurate is the model?

The question that actually predicts the outcome is: once this works, whose job gets harder?

If the answer is “nobody”, that usually means the project changes no real process, and the benefit will be zero. If the answer is a specific person, the follow-up questions become obvious — does he know? Does he agree? Does he have the standing to say no?

I have never seen an AI project die because the model was not good enough. They die because nobody was willing to change how they worked for it.

What the second month is actually for

Do not expand scope in month two. Narrow it. Find the person whose job gets harder, and spend the entire month on him.

Concretely: write down everything he knows that the model does not, and ask why for each item. That list is usually worth more than the model, because it is the company’s real operating knowledge — and it has never been written down before.

If you hold that list at the end of month two, the project will probably survive. If you do not, no accuracy figure will carry it past month four.