The dull work is the work
Every AI pilot I have watched stall began with the most interesting problem in the business. The ones that stick begin with the job nobody wants on a Friday afternoon.
Ask a leadership team where AI should start and you will usually hear about the most interesting problem they have. Demand forecasting. A customer-facing assistant. Something that produces a chart the board has not seen before. These are interesting problems, and that is precisely what makes them the wrong place to begin.
The interesting problem is interesting because it is hard. It leans on data nobody collected properly, judgement that lives in two people's heads, and a definition of success that has never been written down. Six weeks in, the pilot is still arguing about what the number means. The sponsor stops attending. The conclusion everyone draws is that AI does not work here, when what actually happened is that the first attempt was aimed at the hardest target in the building.
Meanwhile, somewhere in the same business, there is a person who spends Monday morning opening two spreadsheet exports from third-party delivery aggregators. They paste them side by side and try to work out which orders were late, which were never made at all, and whether the money that arrived matches the sales the aggregator says happened. Across 78 stores. They cannot finish it, so they sample it. Nobody challenges the sample, because there is nothing to challenge it against.
That is the work worth automating. It is dull, which means it is well defined. The inputs already exist as files, which means there is something to build on. And the person doing it can tell you in one sentence what a correct answer looks like, which means you can tell whether the thing you built works.
There is a second reason to start there, and it matters more than the first. The dull job is usually sitting on top of the data everything else needs. You cannot forecast demand for a group whose order-level history is trapped in weekly PDF exports. Reconcile the exports and you have built the forecasting foundation as a side effect, with a result in week four rather than a promise in month six.
The pattern repeats. Stock counts done on paper and typed up later become stock counts that reconcile themselves, and then the variance report somebody wanted all along. Maintenance requests shouted down a phone become tickets, and then the first honest answer to which equipment keeps failing. Each of these started as administration and ended as a number the business did not previously have.
The objection to starting dull is that it is not exciting enough to get funded. In my experience the opposite is true. A sponsor who watches one person get their Monday back will fund the next thing without a business case, because they have seen it work on something they understood. A sponsor who sat through six weeks of a forecasting pilot will not.
So the first question is not what AI could do for the business. It is which job in this building is done badly because a person is doing something a machine should do, and who would notice within a week if it stopped.