Start from the bottleneck
The best automations do not begin with a model name. They begin with a queue: tickets aging out of SLA, reconciliations that slip weekly, or onboarding steps that require four tools and a spreadsheet nobody trusts.
Pick one bottleneck with a clear owner, a measurable baseline, and a definition of “done” that finance will recognize.
Separate policy from execution
Agents fail in production when “what good looks like” lives only in someone’s head. Write the policy in plain language: allowed actions, required confirmations, escalation paths, and data sources of truth.
Execution—the agent—should be the cheap part. Policy is where risk is managed.
Instrument before you optimize
Ship with logs you will actually read: inputs, tool calls, confidence signals, and human overrides. If you cannot explain why an agent chose a path, you cannot improve it safely.
Roll out like a product
Pilot with a narrow slice of volume, compare against control, then widen. The goal is compounding throughput, not a launch-day headline.
If you want a second pair of eyes on a workflow map, send it over—we review a handful each month for teams serious about shipping.
