Start with the process, not AI
The strongest first candidate is repetitive, time-consuming, and has a clear input and output. It does not need to look impressive. Data transfer, request classification, or preparing documents can outperform a flashy chatbot without ownership.
Measure a baseline first: case volume, time per case, error rate, and exception cost. Otherwise you cannot judge the result after launch.
A scoring matrix
Score each process from one to five across five dimensions. High frequency, substantial manual effort, stable rules, and available data raise priority; severe risk without a control point lowers it.
- frequency and volume
- manual effort and cost
- rule stability
- input-data quality
- error impact and approval options
Human approval is a feature
Payments, legal decisions, customer communication, and uncertain AI output need an approval point. Automation prepares the proposal and evidence; a person decides.
Production workflows also require logging, error handling, retries, cost controls, and an accountable owner. A successful demo path is not enough.
Keep the first workflow narrow
Choose one source, one primary flow, and one success criterion. Track external API and model costs separately. Expand only after real data proves the flow.
This creates a measurable result sooner and avoids turning an unclear process into an expensive automation mesh.
How to prepare a pilot
Select a representative case sample including exceptions, name the process owner, and define when a person must intervene. Run the pilot long enough to expose less frequent situations.
The result should be more than a success percentage. Track accuracy, handling time, interventions, operating cost, and team feedback.
- real data with unnecessary personal data removed
- a clear rollback and manual path
- logs and a reason for every failure
- a predefined go/no-go criterion