Agree the scope and the finish line.
Identify the business result, the process boundary, and the systems involved. Define observable completion criteria before building.
Discovery artifact: process map, exception inventory, completion rule.
Custom implementation starts with how the work actually happens.
Bring a real example. The process diagram is welcome too, but the real example tends to know where the exceptions live.
Build around the real constraints.
Connect the software, model the rules, and assign approvals and exceptions. Include the manual steps that still require a person.
Build artifact: connected system modules with human decision points.
Prove it with representative work.
Run ordinary cases and awkward ones. Check the final records and outcomes, not just whether the AI produced a plausible answer.
Live operation: event log, verified outcome, and a clear exception route.
Put it into operation.
Roll out with the responsible team. Review execution records and agree the maintenance and support scope as part of your implementation.
Maintenance: a changed policy produces a versioned adjustment to the operation.