Begin with friction, not software
The best first automation is rarely the most impressive demo. It is usually a process your team already understands, repeats frequently, and quietly resents. Think about information copied between systems, predictable follow-ups, recurring reports, or documents that always pass through the same review steps.
Starting with the process keeps the conversation grounded. Instead of asking what an AI agent could do, ask where work waits, where errors appear, and which routine steps consume attention without benefiting from human judgment.
Score the opportunity
A useful candidate has enough volume to matter and enough consistency to describe. It should also have a clear owner who can explain exceptions and decide whether the result is correct.
- Frequency: How often does the process happen?
- Effort: How much staff time does each run consume?
- Consistency: Can the normal path and exceptions be described?
- Risk: What happens if the automation is wrong or unavailable?
- Access: Can the required systems and data be connected safely?
- Evidence: How will you know the new process is better?
Choose a contained pilot
A strong pilot improves one complete path without trying to redesign the entire company. It includes logging, error handling, and a human fallback from the beginning. If AI is used for an uncertain task such as interpreting a document, define what confidence is acceptable and when a person must review the result.
Measure the before and after: elapsed time, staff effort, error rate, rework, or response time. The goal of a pilot is not to prove that automation is exciting. It is to produce enough evidence for a responsible go, change, or stop decision.