Seven minutes was the real problem
A review workflow required an operator to inspect a recorded session, understand what happened, and capture the relevant facts. The average review took about seven minutes. The obvious proposal was “use AI to summarize the video.”
But a summary is not an operational outcome. What information does the reviewer actually need to decide, and how should uncertainty appear in that workflow?
My first framing was too narrow
AI integration is a relay race. The model runs one leg. Capture quality, transcription, prompt structure, validation, interface design, and human judgment run the others.
Optimizing only the model can make one leg faster while the baton still falls at every handoff.
What the working system actually required
The workflow first transforms unstructured audio and video into a transcript, then extracts a constrained set of review signals. Those signals are presented in a form that lets the operator verify rather than rediscover the entire session.
The output contract is crucial. Free-form prose is difficult to validate and slow to scan. Structured fields, evidence references, and explicit uncertainty make the result operationally useful. The interface must keep the source available because the AI assists the decision; it does not erase accountability.
The successful measure was not model accuracy in isolation. It was the end-to-end review time and the operational effort removed. The redesigned flow brought the average review below two minutes, reducing manual effort by more than 75%.
Where the prompt-first approach broke down
The first mental model treated transcription and analysis as the product. They were capabilities. The product was a faster, safer decision inside an existing process.
It was also easy to assume that a better prompt would fix every weak output. Some failures began earlier—with poor source quality—or later, when the interface did not help a reviewer verify the extracted claim.
The rule I kept
Production AI should be evaluated at the workflow boundary. Ask whether the complete system changes time, effort, consistency, or decision quality.
The model matters, but the engineering value appears only when probabilistic output is shaped into an auditable human process.