Design the system around a bounded task
AI is most dependable when the objective, available context, permitted actions, and success criteria are explicit. Start with one workflow step where assistance is valuable and failure can be detected.
Keep authoritative records, permissions, and business rules outside the model. The model may interpret or propose; the surrounding application should enforce.
Evaluate the workflow, not the demonstration
Build a representative evaluation set from real task shapes, including ambiguous inputs and failure cases. Measure correctness, unsupported claims, escalation behavior, latency, cost, and the effort users spend reviewing output.
- Approved data boundary
- Versioned prompts and models
- Representative evaluation set
- Auditable inputs and outputs
- Rate and cost controls
- Human escalation path
Make fallback a normal state
A durable workflow can continue when the model is unavailable, uncertain, or outside policy. Communicate confidence carefully, preserve the original evidence, and make human correction easy. Use those corrections to improve evaluation before expanding autonomy.