There is a version of an AI project that starts with the technology and looks for a problem. It produces an impressive demo, a proof of concept nobody uses, and a quiet write-off eight months later. There is another version that starts with an expensive, repetitive decision the business already makes hundreds of times a week. The second version tends to work.
Start with a decision, not a capability
The useful framing is not what can AI do, but which decision are we making repeatedly, slowly, or inconsistently? Approving a claim. Routing a support ticket. Estimating demand for a SKU. Extracting figures from a supplier invoice. These are narrow, measurable, and high volume — exactly the conditions where automation compounds.
Define what correct means before you build
This is the step teams skip, and it is the one that decides the outcome. Before any model exists, you need an agreed definition of a right answer and an acceptable error rate. A document-extraction system at 92 per cent accuracy might be excellent or unusable depending entirely on what happens to the other 8 per cent, and who catches it.
You also need a baseline. If a human currently does this task with a 6 per cent error rate, that is the number to beat. Without it, every result is unfalsifiable.
Data is the actual project
Expect the majority of the effort to go into data: locating it, cleaning it, labelling it, and dealing with the fact that it reflects historical decisions that may themselves have been inconsistent. Teams routinely budget for modelling and are surprised by data engineering. Plan the reverse.
Design for the failure case
Any system that makes predictions will be wrong sometimes. The design question is what happens then. Good deployments route low-confidence cases to a human, log the disagreement, and feed corrections back. Bad deployments present a confident answer with no indication that it might be wrong, and quietly erode trust until people stop using it.
A reasonable first engagement
Pick one decision. Establish the baseline. Build the narrowest thing that could beat it. Put it in front of real users with a human in the loop. Measure for a month. If it holds, widen the scope; if it does not, you have spent weeks rather than a year, and you have learned something concrete about your data.
The organisations getting real value from AI are rarely the ones with the most ambitious projects. They are the ones that shipped something small, measured it honestly, and iterated.