Most AI deployments survive their pilot phase and die at their first audit.
The cause is almost never the model. The model is doing what it was trained to do. The cause is the specification. The team that built it produced a forty-page solution architecture document, three Confluence pages of design notes, and a Slack thread where the critical decisions live buried inside it. When the auditor asks "what is this system supposed to do, and how do you know it is doing it?", nobody can produce a one-page answer.
The specification that survives audit is five lines per workflow.
- Input — exactly what the system receives, in what format, from which upstream source.
- Transformation — what the model does to that input, stated as the business outcome (not the prompt).
- Confidence threshold — the score below which the system stops and escalates instead of acting.
- Exception path — exactly where escalated items go, who owns them, and what the SLA is.
- Audit log location — where every decision is written, in what schema, retained for how long.
Five lines. One workflow. If you cannot write the spec in five lines, you have not yet specified the system — you have described an aspiration.
Across the production AI engagements I have shipped, this is the dividing line. A healthcare receivables automation that runs roughly fifty thousand invoices a month through a confidence-scored extraction pipeline had a five-line spec per transaction type. A multi-channel marketplace reconciliation that posts to a clearing-account architecture across Amazon, Shopify, and Walmart — five lines per channel. An eight-language voice agent that books field service appointments and routes them by drive-time efficiency — five lines per intent.
The five-line specification is not a documentation step. It is a build constraint. If the team cannot agree on those five lines before code is written, the build has not started. The "code first, document later" approach is how the majority of enterprise AI pilots never reach production.
If your AI roadmap document is longer than your AI specification, you have a strategy artifact, not a build plan.
Filed under
