Start with a bounded, observable process
The first automation should have stable inputs, a clear owner and a result that someone can verify. A recurring handoff or reconciliation task is often a better candidate than an ambiguous decision involving several teams.
Document the current process before estimating savings. Measure frequency, handling time, exception rate and the consequence of a wrong result. Estimates based on guesses should remain labelled assumptions.
Compare candidates using a simple record
| Candidate property | Favourable signal | Reason to investigate first |
|---|---|---|
| Rules | Explicit and stable | People disagree about the correct outcome |
| Input | Structured and accessible | Missing access or inconsistent records |
| Failure | Recoverable and visible | Silent irreversible consequences |
| Ownership | Named operator | No team owns exceptions |
| Measurement | Before/after task evidence | Only a promised percentage improvement |
A small deterministic integration may be enough. For example, saving a validated enquiry and notifying its owner does not inherently require a language model. Reserve AI for tasks where interpretation is genuinely needed and review can catch errors.
Account for operation, not just build effort
Include provider fees, retries, monitoring, maintenance and human review. Time released by automation is not automatically cash saved; the organisation still needs a way to use that capacity.
Pilot on a small, representative set. Keep the previous path available while testing. Decide how the operator pauses the automation, resolves duplicates and reprocesses failed work.
Write the stop condition
Before launch, specify what error pattern should halt processing and who responds. A process that keeps producing questionable output without an owner is not operationally complete.
Use an automation review to separate deterministic work, AI-assisted work and decisions that should remain human. Expand only after observing the first workflow under realistic conditions.
A practical question
How should we estimate automation savings?
Measure task frequency and handling time, then record assumptions about exceptions, review and ongoing provider costs. Time released is capacity; it becomes financial saving only under a defensible business model.