Start with the problem, not the technology

The most common question in AI conversations is "how can we use AI?" It sounds sensible. It's actually the wrong way round.

Start from the technology and you get a predictable pattern: a pilot, a demo that impresses a room, and then very little change in how work gets done. Not because the model was weak, but because nobody had agreed what problem it was solving or how anyone would know it had worked.

Why it keeps happening

• Pressure to be seen doing AI. Activity is easier to show than outcomes.
• Capabilities are easier to describe than problems. "We could summarise documents" is crisp. "Our claims process loses a day at handover" requires someone to look closely.
• Ownership is unclear. A technology project has a technology owner. A problem has a business owner — the person who will actually change how they work.

Every article on The Thought Project follows the same sequence, because it works just as well inside a business:

1. What is the problem? Specifically: who has it, how often, what it costs.
2. Why does the existing approach fail? Usually it asks too much of people's attention or memory.
3. Where can AI change the equation? A capability — reading, classifying, predicting, generating, acting — that addresses that specific failure.
4. What would the solution look like? Concrete enough to sketch. Honest about what stays human.
5. How would someone implement it? First steps, not a programme.

If you can't answer the first two convincingly, stop. The third question will only produce a demo.

The deeper point is about process. Adding AI to an existing workflow — a chatbot in front of the same support queue, a summariser on top of the same meeting — usually yields small gains.

The bigger gains come from asking: now that this capability exists, how would we design this process from scratch? If continuous monitoring is cheap, maybe the hygiene checklist disappears. If drafting is cheap, maybe agents review instead of write. If forecasting is cheap, maybe the bake list is drafted overnight.

That's a different kind of project. It needs a problem owner, not just a technology owner. It needs a clear measure of success. And it needs to be small enough to try next month.