AI in operations is a workflow problem

Most AI transformation briefs are undocumented operations. Someone wants an agent. Nobody can write the standard operating procedure the agent is supposed to follow. We write the SOP first, then decide which steps a model should own.

Agents fail at the handoff, not at the paragraph. If escalation rules live in a veteran's head, the model will improvise — and improvisation in operations is how refunds go missing.

Fix the process, then automate the edges

Humans keep judgment: exceptions, angry customers, legal grey. Models draft, classify, and route. The product is the handoff, with a transcript a supervisor can audit.

We measure deflection only when the customer is actually done. A closed ticket that reopens in an hour is not automation. It is delay with extra steps.

Start with one workflow

Password resets, invoice copies, delivery windows. Pick a flow with volume and a clear happy path. Publish the SOP in the same repo as the prompts. Then expand.

If the business cannot describe the work on a whiteboard, it is not ready for an agent. That sentence saves more budget than any model upgrade.