01
Describe the change, not the tool
Start with what should improve and for whom. “Use an agent” is not an outcome; reducing measured time, preventing an error or increasing a capability can be checked.
- ✓Current state and baseline.
- ✓Expected outcome and beneficiary.
- ✓Approximate frequency and volume.
- ✓What remains outside the first scope.
02
Map the work that already exists
Automation inherits process ambiguity. Document steps, owners, decisions, exceptions and tools before designing an architecture.
Inputs
Documents, messages, records, events or instructions that start the work.
Decisions
Stable rules, expert judgment and cases requiring escalation.
Outputs
Artifacts or actions produced and who uses them.
Exceptions
Rare cases, incomplete data, missing permissions and external failures.
03
Define evidence and boundaries
The first experiment should answer one concrete question on a representative sample. Write the success and stop conditions before building.
- ✓Real, anonymised examples.
- ✓Metric, comparison or rubric.
- ✓Maximum experiment and operating cost.
- ✓Human review proportional to risk.
- ✓Rollback, fallback and incident owner.