How to plan an AI workflow pilot.

By ETE Oversight · Published October 18, 2026

A useful AI pilot answers a business question: can this workflow produce acceptable results with less total effort, reasonable operating cost, and appropriate human control? A successful demonstration alone does not answer that question.

1. Choose a workflow with a clear owner.

Start with a repeatable task that has accessible, approved information and a person who can judge the result. Internal procedure search, employee onboarding questions, and drafting routine responses for human review are possible starting points. Suitability depends on the information involved and the consequences of mistakes.

Write a one-sentence scope: “Help the operations team find the current approved procedure, with a source reference, while a person remains responsible for applying it.” Name the users and explicitly identify actions the assistant is not allowed to take.

2. Record the current process before changing it.

Measure a representative sample of tasks using the existing process. Record completion time, useful outcomes, mistakes, and follow-up effort. Use comparable tasks for the pilot. Include difficult and ambiguous questions, not just examples that make a demonstration look good.

3. Prepare approved knowledge and access.

Identify document owners, current versions, conflicting instructions, and missing content. Confirm which sources the pilot may use and which users may access them. An assistant cannot resolve an organization’s policy conflicts simply by retrieving both documents.

Agree how obsolete material will be removed, how new documents will be reviewed, and who handles questions the source material cannot answer. Confirm tool suitability and data handling before uploading organizational information.

4. Define the pilot’s controls and exit criteria.

Assign a human owner and a limited user group. Document escalation, manual fallback, and who can stop the pilot. Agree what evidence would justify expansion, further work, or stopping. Set these criteria before reviewing the final results.

A 30-day plan can provide structure, but access approvals, data preparation, and integration complexity determine the actual schedule. A pilot proposal should state dependencies, responsibilities, deliverables, and excluded work.

5. Calculate value after review effort and costs.

A useful starting calculation is: net capacity released equals baseline human time minus pilot human time, including checking, corrections, and escalations. Then consider recurring tool and support costs, plus the initial delivery cost.

Illustrative example: if a task previously took 12 minutes and now takes 3 minutes to draft plus 5 minutes to verify, the apparent capacity gain is 4 minutes per task. At 100 comparable tasks per month, that is about 6.7 hours before maintenance and other costs. These are hypothetical numbers, not ETE client results.

Released capacity becomes a financial benefit only if the organization can use it productively or reduce a real expense. Do not present every saved minute as cash savings.

6. End with a decision and a handover.

Record the tested configuration, results, limitations, user guidance, owners, and support arrangements. Decide whether to expand, revise, or stop. Keep unresolved issues visible. Where independent assurance is needed, distinguish the implementation team’s testing from a separate reviewer’s assessment.

Questions teams often ask

Do we need to build our own AI model?

Not necessarily. Evaluate whether established tools can meet the workflow’s requirements before considering a custom model. The right choice depends on access, integration, performance, cost, and control requirements.

What should we prepare before a scoping call?

A non-confidential description of the workflow, its owner, current tools, intended users, and desired outcome is enough to start. Do not send sensitive records or credentials in an initial inquiry.

For the evaluation side, see ETE’s AI agent readiness checklist and Workflow Readiness Review.