Look for repetition plus a clear source of truth
The strongest first projects are repetitive, frequent and already governed by rules your staff can explain. Examples include classifying incoming requests, extracting fields from recurring document types, preparing daily operational reports, routing approvals and following up on known events.
Avoid the glamorous first project
A general “ask our company anything” assistant sounds impressive but often has weak economics and difficult permissions. Automating one painful back-office loop can create more value and gives you the integration layer needed for larger AI work later.
Score candidate workflows
Give each candidate a rough score before building.
- Frequency: does it happen every day or week?
- Labour: how much skilled time does it consume?
- Delay: does waiting create lost revenue or poor service?
- Rules: can a good employee explain how the decision is made?
- Data access: can the required systems be integrated safely?
- Failure tolerance: can uncertain cases be routed to a human?
Start narrow, instrument everything
Launch on a bounded set of cases and measure what the system actually handles. Keep the exceptions visible. A useful automation project should make the next business decision easier: expand, change the process, or stop.