Start with the decision
Describe the business problem in plain language. Who makes the decision today? What information do they use? Where does the workflow break down? A specific question is easier to evaluate than a broad ambition to “use AI.” Define what useful improvement would mean before considering a solution.
Check the evidence
Review the information available to support the use case. Consider quality, access rights, freshness, representativeness, and ownership. For a knowledge assistant, document permissions and source reliability are central. For a prediction system, historical coverage and label quality may determine feasibility.
Match risk to oversight
Different applications need different safeguards. A drafting assistant and a system influencing a financial or health decision do not carry the same consequences. Establish human review, escalation paths, evaluation criteria, and appropriate limits. Include security and privacy stakeholders early.
Set the baseline before the pilot
Measure the current workflow and define a representative test set. Compare the proposed approach with simpler alternatives. Evaluate failure modes alongside successful examples, and include operating cost and ongoing maintenance in the decision.
Choose a bounded next step
Select one useful, testable workflow. Agree on users, scope, data boundaries, and a stop-or-expand decision. A pilot should produce evidence for a decision—not simply an impressive demonstration.
These are general planning considerations, not a substitute for an assessment of your specific business, regulatory, and technical requirements.
Explore related consulting expertise