Sample chapter
Start with the work queue, not the model
The first serious AI project usually begins in the wrong room. A vendor demonstrates a model, a leader asks where it could help, and a team searches for work that justifies the seat. Reverse the order. Find recurring work that already has an owner, a measurable wait, a repeatable decision, and a reviewer. Then ask whether a model can remove one clearly limited step.
A work queue is evidence. It tells you how much work arrives, how long it waits, how often it returns, and who absorbs mistakes. “We should use AI in finance” has none of that. “Four hundred vendor invoices wait 2.6 days for coding and a controller reviews every exception” does.
The work-queue test
- Does the work arrive often enough to measure?
- Can one person name the acceptable output?
- Is there a reviewer with authority to reject it?
- Can the team recover when the output is wrong?
- Will the result save a delay, reduce a retry, or improve a decision—not merely produce more text?
If the answer is no, the project is not ready. A better model does not repair work nobody owns.