Start with the boring workflow
Intake, routing, enrichment, approvals, reporting. These are the processes where hours disappear and where a mistake is cheap to catch. They are also where automation pays back fastest.
Nobody puts 'automated our lead routing' in a keynote. It is still the highest-return AI work most companies can do.
Deterministic first, model second
If a rule can do it, a rule should do it. Rules are testable, cheap and explainable. Reserve models for the genuinely ambiguous steps — classification, summarisation, extraction from messy input.
Stage the rollout behind human review
Every deployment runs in shadow mode first: the agent proposes, a human approves, and we measure agreement. When agreement holds above the target for two weeks, the human step comes out.
This is slower to launch and dramatically faster to trust.
Instrument everything
Log every decision, every input and every override. The override log is the training set for the next iteration and the audit trail when someone asks why the system did what it did.