The POC Answered the Wrong Question

A POC proves the model works. It proves almost nothing about whether your organisation will let it run.

That is the most expensive distinction I see missed in AI programmes.

A POC runs on clean sample data, in a sandbox, with a hand-picked team, against criteria the vendor helped write. It was built to succeed. That was the brief.

The teams who know this use real data. They still use the easy half of it. The edge cases stay out, and there are always more of them than anyone expects. They are not only in the data. They are in the process. What happens when five of them land in the same hour. Who owns the exception. Where it goes when nobody does.

Production runs on live pipelines at real volume. It touches customer data under actual privacy constraints. It needs an owner after go-live, an escalation path for when it errs, and someone willing to put their name to its outputs.

None of that was tested. None of it is technical.

The POC answered "can this work". The question on the table is "should we run this"

And "should we" is not one question. Will we be able to. Will it get approved. Will the teams use it. Will the resources still be there in nine months. Will anyone change how they work because of it.

Not one of those is answered by the model working. That is almost never where it fails.

Companies make the production call using only the evidence the pilot produced, and call it de-risked because something worked once.

It is not de-risked. It is unexamined.

The distance between a POC and production is not a build. It is a set of questions nobody was assigned to ask.