What should you build. What should you ignore.

The pilot graveyard

AI initiatives rarely fail on the model. They fail on scope: the wrong problem, the wrong data, no named owner. None of that shows up at the demo. It shows up a year later, at budget review.

• Several proofs of concept. None of them carrying real traffic.
• Every vendor demo works — on the vendor’s data.
• Nobody can state what year one costs, or what it returns.

A decision you can defend in the boardroom

I talk to the people who do the work, read the systems you already run, and score every opportunity the same way. Then I tell you which candidates to start, which to drop, and what each survivor costs to run for a year.

• Opportunity ranking: every candidate scored on value, effort, and time to first result
• A build / buy / ignore call on each, with the reasoning written down
• Target architecture and integration requirements for the stack you already run
• Operating-cost estimate and risk register that survive contact with finance
• A 90-day execution plan whose first deliverable has acceptance criteria

1. 01

   Listen

   Interviews with the people who actually run the process, plus a direct read of your code, data, and constraints.

2. 02

   Map

   Every candidate scored on the same axes, so the comparison stops depending on who argues hardest.

3. 03

   Decide

   A working session where we kill the weak ideas together and commit to the few that pay.

4. 04

   Plan

   Architecture, cost model, and a 90-day sequence — ready to execute, by your team or with mine.

The judgment behind the ranking comes from delivery, not theory. That includes an analytics platform roughly 50× faster than the BigQuery it replaced, at less than a fifth of the infrastructure cost. Measured results live in the work section, with their baselines attached.

• We already have an internal AI team.

  Then you need a second opinion that isn’t attached to last quarter’s decision. The plan is your team’s to run, including the parts where they were already right.

• Can you have something ready for our board next week?

  No. A roadmap worth funding needs interviews and a real look at your systems. If the deadline is immovable, I’d rather decline than hand you a plan I can’t defend.

• What if the answer is ‘don’t use AI here’?

  Then that’s the answer, and it’s the cheapest one you’ll ever buy. Sometimes the fix is a database index and a clearer process.