Choosing your first ops workflow for AI automation
The most expensive mistake we see at kickoff is not technical. It is workflow selection. The team picks the most painful, ambiguous, organizationally sensitive workflow on their plate, ships a pilot at it, and watches the rollout collapse under the weight of edge cases nobody could have scoped in advance.
Pick a smaller workflow. The first agent you ship is not your moonshot. It is the one that proves the platform.
The scoring rubric
Score every candidate workflow on six dimensions. Anything below 18 / 30 is a no for a first pilot.
| Dimension | 1–5 | What "5" looks like | | --- | --- | --- | | Volume | per week | 500+ cases / week | | Inputs | structuredness | Mostly structured fields, some text | | Decisions | enumerable | Under 20 distinct decision types | | Reviewability | by a human | Under 2 minutes per case to QA | | Reversibility | of the action | One-click rollback if wrong | | Owner | clarity | One ops leader on the hook |
The four use cases on our use cases page all score 24+ on this rubric. They are not a coincidence, they are exactly the workflows that clear pilots.
What to avoid (for now)
- Free-form output. A workflow whose deliverable is a multi-paragraph email or a strategic memo. Save it for v2.
- Rare events. Anomaly detection, fraud, single-shot underwriting at high stakes. Volume is too low for the eval harness to converge in 30 days.
- Distributed ownership. A workflow that touches three teams and has no single accountable leader. The agent is fine; the rollout dies in coordination.
- Net-new processes. "We do not do this today, but we want the agent to do it." That is process design, not automation. Do the process design first, then automate.
The workflow you actually want
A boring, expensive workflow with a queue. That phrase is doing a lot of work.
- Boring because the decisions are repetitive and you have hundreds of historical examples to ground the agent in.
- Expensive because the team currently spends meaningful headcount on it and the savings are visible in the quarter.
- With a queue because that is where supervisor handoffs route, where confidence thresholds matter, and where the eval harness pulls labels.
If you are unsure which one of yours fits, ask us, we will run the rubric against your four most-cited workflows and tell you which one to start with. Usually it is not the one the executive sponsor picked.
The win condition for a first pilot is not "automated everything." It is "shipped one thing, clean curve, supervisor queue under control, ready to expand." Pick the workflow that gives you that, and the second one becomes much easier.