Adding an agent to any vertical: the 4-week pattern

The most useful thing about agent deployments is how repetitive they are. The vertical changes — outreach, invoice matching, support triage, inventory reporting — but the four weeks barely do. This is an annotated build log for a logistics client's quote-request workflow.

Week 1: map and ingest

Sit with the person who does the work today. Capture the real workflow, including the exceptions they handle on autopilot — those unwritten rules are exactly what kills naive automation. In parallel, the brain ingestion starts: rules, the fact base for this workflow, five examples of a good output. By Friday the agent can answer questions about the domain but doesn't act yet.

Week 2: open agent

The workflow runs agent-first, human-steered. The operator approves or corrects each step in a thin review UI. Every correction is logged — this is the raw material for the eval suite. Expect embarrassment here; week-2 transcripts are where you find out the price list had three contradicting versions.

Week 3: write the checks

The corrections from week 2 become explicit checks. In this build, fourteen corrections compress into six checks — three factual lookups against the brain, two format rules, one escalation condition ("any request mentioning hazardous goods goes to a human, always"). The agent now runs the full loop with the human only approving final output.

Week 4: close it

Scheduled runs, unattended, with the eval suite as gatekeeper and an escalation path with full transcripts. The first solo run in this build processed eleven quote requests; two escalated, correctly. That ratio — most handled, edge cases surfaced rather than guessed — is what "done" looks like.

The pattern holds because the hard part was never the model. It is extracting how the work actually works, then encoding the judgment as checks. Verticals differ in content; the verbs are identical.