Start from repeated work, not from the model
The best automation candidates are tasks a person repeats dozens of times a week with a consistent input and a checkable output — quoting, document extraction, triage, first-line support.
If you cannot describe the correct output, an AI system cannot be evaluated, and unevaluated automation quietly creates work instead of removing it.
Grounding beats cleverness
A retrieval-augmented setup answering from your own documents will outperform a larger model guessing from general knowledge, and it gives you citations to audit.
Set confidence thresholds and route low-confidence cases to a human with full context. Escalation is a feature, not a failure.
Measure before and after
Record handling time, error rate and volume for two weeks before launch. Without that baseline every claim about savings is a guess.