Test workflow fit
AI is most useful when a task contains variable language, judgment, or pattern recognition that rules alone cannot handle economically. It is less suitable when exactness, deterministic behavior, or a simple lookup solves the problem.
Does the task happen often enough for improvement to matter?
Does the input vary in ways that make fixed rules brittle?
Can a human or system tell the model when an answer is wrong?
Would a form, rule, search index, or integration solve it better?
Inspect the data
Map what the model receives, what may be retrieved, what can leave the organization, and what must be excluded. “Available” does not always mean permitted, representative, current, or useful.
- Source, owner, classification, retention, and permitted purpose.
- Coverage gaps, bias, duplicates, stale content, and contradictory records.
- Personal, confidential, regulated, copyrighted, or client-controlled information.
- Provider training, logging, residency, deletion, and subcontractor terms.
Bound the risk
| Impact if wrong | Recommended control | Example |
|---|---|---|
| Low and reversible | User review and easy correction | Drafting an internal summary |
| Moderate | Structured validation, citations, and approval | Routing a client request |
| High or regulated | Human decision authority and strict exclusion rules | Eligibility, legal, medical, or financial action |
Define evaluation
Choose evaluation examples before choosing prompts. A useful test set represents normal work, difficult edge cases, prohibited requests, missing information, and known failure patterns.
Quality
What makes an output correct, complete, grounded, relevant, and appropriately cautious?
Operations
What latency, cost, review time, retry rate, and escalation volume are acceptable?
Safety
Which behaviors must always be blocked, flagged, logged, or routed to a person?
Plan the rollout
Start in a bounded workflow with observable outcomes and a clear owner. Expand only after the team understands failure patterns, review burden, operating cost, and user behavior.
Make the readiness decision.
- The workflow is frequent, variable, and valuable enough.
- Data use is understood and permitted.
- Human authority is clear at each risk level.
- Evaluation covers real work and important edge cases.
- Monitoring, correction, and rollback have owners.
- A simpler non-AI option has been considered honestly.