The first AI process should be neither the most visible nor the most technically impressive. A good starting point combines sufficient volume, clear business value, available data and a failure impact that can be controlled through review and fallback. That combination rarely appears in a simple list of “possible use cases.”
Observe work, not just descriptions
Interviews reveal the intended process. Add observation and samples to find the actual one: which inputs arrive, how many variants exist, where cases wait, which information is gathered from email, documents and systems, and where people decide through experience rather than a fixed rule.
Record exceptions and returns. A process that takes two minutes normally may still spend most effort on a small number of difficult cases.
Establish a baseline
Before a pilot, estimate volume, handling time, lead time, errors, rework and business outcome. Without a baseline, the project can show that a model generated something but not that work improved.
Avoid false precision. Transparent ranges and a plan for better measurement are preferable when source data is weak.
Assess four dimensions
Value: Does it save meaningful effort, improve quality or enable work that was previously impractical?
Feasibility: Are inputs, knowledge, systems and accountable specialists available?
Risk: What happens after an incorrect extraction, recommendation or action? Can a person review in time?
Operability: Are there owners, support, monitoring, budget and a way to update rules or knowledge?
The NIST AI RMF recommends examining context, impact and risk before and throughout deployment. Discovery should identify limits and affected people as carefully as opportunities.
Bound the pilot
Choose a vertical workflow slice with a defined input, user group, small tool set and measurable outcome. Specify fixed evaluation cases, human approvals, fallback and stop conditions.
Good discovery may lead to three outcomes: start a pilot, stabilize data or process first, or deliberately avoid automation. That honesty prevents a successful demo from becoming an unplanned production system.