A practical checklist before you add AI to operations

Most of the AI projects that disappoint do not fail because the model was weak. They fail because the work underneath was not ready to be automated, and nobody checked first. So before you wire a model into anything that runs your business, here is the checklist we actually walk through with clients. It is dull. It is also the difference between a tool that earns its keep and a clever experiment that quietly gets switched off.

The questions worth asking first

• Can a person describe the task without contradicting themselves? If two experienced people give you two different versions of how a job is done, the model has no chance. Fix the disagreement before you automate it.
• Where does the data live, and do you trust it? If the answer involves four spreadsheets and someone's memory, the model will inherit all of that. Clean inputs are not a nice to have. They are the whole game.
• What does a good outcome look like, in plain terms? If you cannot say what success is, you cannot tell whether the model is helping or just producing output.
• What happens when it is wrong? Every model is wrong sometimes. Decide now who catches it, how, and how fast. If the answer is "we would probably notice eventually," stop.
• Is anyone genuinely accountable? Not a committee. One person who owns the result and can overrule the machine without asking permission.
• Will the people doing the work actually use it? A tool the floor resents gets worked around within a fortnight. Their buy in is not a soft factor. It is whether the thing lives or dies.

If you can answer those cleanly, you are in good shape. If several of them make you wince, that wince is useful information. It is telling you the foundation needs attention before the AI does, and that is almost always cheaper to discover now than after launch.

This is most of what a Discovery Sprint is for. In two to four weeks we are not building the flashy part. We are answering these questions honestly, in your real environment, with your real data, and ending with a working prototype that proves the answers hold. Sometimes the prototype shows that AI is exactly right for a step. Sometimes it shows that a small fix to a process removes the problem entirely and no model is needed at all. Both are good outcomes. Both save you from spending months on the wrong thing.

The temptation, especially when everyone around you is announcing their AI initiative, is to skip the boring questions and get to the demo. Resist it. The demo is easy. The model that holds up on a busy Monday, with messy inputs and a tired team, is the one that came from doing this checklist first. Add AI to operations that are ready, and it compounds. Add it to operations that are not, and you have just made the mess faster.