How to scope an AI workflow

Define the task

Write down who does the work today, what starts it and what a usable result contains. Choose a bounded task rather than a broad instruction to automate a department. Include the cases that require judgment or escalation, and record the current time and cost as a baseline.

Agree data access

List the systems and information the workflow needs. Identify the owner of each source and obtain the necessary permissions before a pilot receives access. Use sample or redacted material where possible. Decide what may leave your environment and which actions must require approval.

Set the acceptance criteria

Build an evaluation set from representative cases. Include difficult inputs, missing information and failure scenarios. Define the errors that would make an output unusable and the level of review required. NIST’s AI Risk Management Framework provides a voluntary structure for considering risk through the design and use of an AI system.

Keep a named reviewer

Specify who reviews the output, what they check and how the workflow stops or falls back when it cannot complete the task. A polished demonstration is not enough to decide whether a process can operate without supervision. Test the handoff with the people who will use it.

Count the full operating cost

Compare the pilot with the current process using the same set of tasks. Include provider charges, maintenance, review time and rework. Record both gains and failures. The decision to expand should follow the evidence from that comparison, with a clear owner for ongoing review.

Sources and further reading

• NIST: AI Risk Management Framework

Sources support the external guidance cited above; the planning approach is this company’s editorial perspective.