8 ideas to share

Why most AI agents fail after the demo

Eight ideas from the book, one slide each. Share the whole set or the one your team needs to hear. Every share includes a link to buy the book.

Idea 1 of 8 · The opening story
A demo is not a Monday.
Production asks something else: can you rely on it when evidence conflicts, systems fail and the consequences belong to someone?
Idea 2 of 8 · Plain English
Delegate like you would to a new colleague.
Ask instead: which responsibility could I safely delegate today, with the evidence, permissions and review I have?
Idea 3 of 8 · Human Twin Design
Map the work, not the role.
Turn each move into something you can build, control and test. Copy the judgement, not the habits and workarounds.
Idea 4 of 8 · The OPT model
Zoom until the work is small enough to own.
Zoom down until the owner, evidence and “done” are clear. Boundaries are discovered, not declared.
Idea 5 of 8 · Choosing the work
Not every task deserves an agent.
If judgement adds nothing, use rules. If you cannot verify it, you cannot evaluate it. Fail a gate, fix that first.
Idea 6 of 8 · Autonomy
The consequence sets the autonomy.
The cost of a mistake does. Choose the minimum useful autonomy per action, not per agent.
Idea 7 of 8 · Evaluation
97% accurate is not a release decision.
A useful eval changes a decision. Measure at five layers, and turn every incident into a test.
Idea 8 of 8 · Cost
The model was 4p. The outcome cost 61p.
Count everything one run consumes, divide by the runs actually accepted, then manage that number.