What to automate first.
Start with a repeated, well-understood workflow that has reliable inputs and a clear owner. Agree how to detect mistakes, handle exceptions and measure the result before automating it.
Start a projectChoose repeated work
List tasks people repeat and the time spent checking or repairing them. Consider enquiry acknowledgements, moving validated records between tools or preparing a recurring report. These are possibilities to assess, not promises of a particular saving.
A good first candidate has a clear trigger, consistent inputs and an outcome someone can verify. A process that changes every time may need clarifying before it needs automation. Keep high-consequence decisions with an appropriately responsible person.
- The same task happens repeatedly
- The rules and inputs are understood
- A person owns the result
- A mistake can be detected and handled
Follow the exceptions
Write the normal path and then ask what happens when information is missing, a system is unavailable or the same event arrives twice. Decide whether work should stop, retry or wait for someone to review it.
Check authorised access to each tool and use only the data the workflow needs. An API being available does not mean every action or record should be accessible. Keep access revocable and document who receives failure notifications.
- Missing or invalid information
- Duplicate events and repeated actions
- Unavailable services or expired access
- A named exception and alert owner
Use AI where it fits
Fixed rules and predictable record movement do not automatically need AI. A conventional integration may be easier to test and explain. AI can be considered when the task involves interpreting variable text, but that introduces uncertainty that needs managing.
Define what an AI-assisted step may do, which information it may use and when it must ask a person. Test unclear and incorrect inputs as well as convenient ones. An answer that sounds confident is not evidence that an action should be carried out.
- A reason to use AI instead of a fixed rule
- Limits on data and permitted actions
- Human review or handover conditions
- Tests for uncertain or incorrect output
Measure a useful change
Record a baseline using a measure relevant to the work: manual handling time, missed follow-ups, failed runs or report preparation effort. Include the work spent reviewing exceptions so the comparison remains honest.
Pilot the workflow on a bounded set of representative tasks before expanding it. Confirm how to pause it and recover from an incorrect result. After launch, review both the business measure and the operational effort; a workflow is not finished just because its first run succeeded.
- An agreed baseline and success measure
- Representative tests and a bounded pilot
- A pause and recovery procedure
- Ongoing review and maintenance responsibility
Your next step
The best first automation is one your team can understand, check and own. A reliable small workflow creates a better foundation for the next improvement than an ambitious system nobody can safely operate.
The project brief helps you describe what exists today and what should improve. For scope and handover, read how delivery works and what affects project investment.
