Field Notes
Short, dated, unrevised. A field note is something we noticed more than once and are not yet ready to assert. Notes that keep surviving contact with reality get promoted into the canon, and the note stays where it is.
The evaluation harness is the artefact that survives
Three programmes this quarter changed their model, their vendor, or their retrieval layer, and in each case the thing that made the change survivable was an evaluation set somebody built early and nobody wanted to fund. The models are the replaceable part. The cases you agreed to measure against are the asset, and they are also the only thing that makes a vendor switch a decision rather than a leap.
Nobody asks what the assistant costs per answer
Programmes report monthly spend and seat counts. Almost none report cost per completed task, which is the number that decides whether the thing scales past the pilot population. It is usually computable from data already collected and it is usually uncomfortable, which is likely the same reason.
Procurement questionnaires are now a leading indicator
Enterprise buyers are adding AI-usage disclosure questions to standard vendor questionnaires. Several organisations have found out about their own supplier exposure this way, because answering the questionnaire required someone to go and look. If you want to know what your board will ask in six months, read the questionnaire your largest customer sent you last week.
The second corpus is where programmes stall
First deployments land on a curated, well-governed document set. Second deployments reach into a store nobody has audited since a reorganisation, and the access review that the first deployment avoided arrives all at once with a deadline attached. Enough repeats of this and it became Entry 004.