Method / Widen, Score, Defend
Watch the work. Allocate the tasks. Encode the result.
AI features fail because the model does not know how the work actually happens. Neither does the spec, or the prompt, that somebody wrote for it. The people doing the work cannot articulate the tacit part on demand, so it never gets written down, and the system gets built around a guess.
Widen
Watch the work liveNaturalistic observation, stimulated recall, probing. I watch people do the actual job with the tools they actually have, then ask the follow-up questions that surface what they could not have told you in a meeting. This produces two things: the context nobody thought to say out loud, and a complete list of every task the system touches.
Score
Decide who does whatTask by task: human or machine, and if machine, how much of the decision. That second question has a rigorous answer. Thomas Sheridan's levels of automation, first set out for supervisory control of complex systems, describe a scale from the machine offering no assistance, through the machine suggesting one option, to the machine acting and informing the human afterwards only if asked.
The choice of level is scored against explicit criteria: consequence if it is wrong, reversibility, volume, and where liability sits. This is function allocation, and it is the work I have been doing since a Navy destroyer program in 2002, where the question was which jobs on the ship the machines should take and which had to stay with the crew.
Defend
Make it usable, and make it surviveThe allocation is only worth something if the team can act on it and if it holds up when challenged. So it ships as prompts and specs the team uses this week, with the reasoning attached to each line so it can be defended to whoever asks. Where the result turns out to be novel, it gets a shortlist your patent counsel can act on.