Four questions before you approve an AI project

Every portfolio company has an AI proposal in front of it. Most are sincere, some are good, and a great many will produce a demonstration that impresses a board and then quietly stops being used.

The difference is rarely the model. It is almost always one of four things, and all four can be checked in a meeting.

The first question is what decision this changes.

Not what it predicts. What it changes. If a system forecasts demand more accurately and nobody reorders differently as a result, the forecast has cost money and saved none. Ask who acts on the output, what they do today, and what they would do instead. If nobody in the room can name the person and the action, the project is a science experiment with a budget.

Good answers sound specific and slightly boring. A buyer stops placing a weekly order by hand. A scheduler stops ringing three sites to find out what is running. Boring is the point.

The second question is where the data lives and who is allowed to move it.

An AI project is a data project in better clothes. Most stall not on the model but on access. The numbers are in a system the vendor cannot reach, or the export is manual, or the only clean copy is a spreadsheet on somebody's machine. All of this is knowable on day one and almost never asked on day one.

So ask to see the data before you approve the project. Not a sample the vendor prepared. The real thing, pulled the way the system would have to pull it, by somebody who has to live with the answer.

The third question is what happens when it is wrong.

It will be wrong. That is not a flaw in the technology, it is what a probabilistic system is. The question is whether being wrong is survivable and whether anybody will notice. A pricing suggestion that drifts gets caught by a human who thinks the number looks odd. A model quietly reclassifying invoices may not be caught for a quarter.

Ask what the failure looks like, who sees it, and how soon. If the answer involves nobody checking, the project needs a control before it needs a budget.

The fourth question is who owns it in eighteen months.

Every one of these is a thing that has to be maintained. Data changes shape. A supplier changes a file format. The person who built it takes another job. A model nobody owns degrades quietly, and a degrading model is worse than no model, because people are still acting on it.

Ask whose job this becomes once the project team has gone. If the honest answer is that it is nobody's, you are buying an asset that turns into a liability on a schedule you cannot see.

None of these four are technical questions. Any of them can be asked by somebody who has never trained a model, and asking them will stop a good number of the proposals in front of you before any money moves. That is not a reason to be gloomy about AI in a portfolio company. It is how the ones worth having get built.