Work through these gates in order, while involving technology and risk reviewers from the beginning.
| Gate | Question to answer | Proceed when… | Otherwise… |
|---|---|---|---|
| 1. Establish the problem | Which calls need a better outcome? | Eligible demand and current outcomes can be measured. | Fix reporting, routing or callback processes first. |
| 2. Define the workflow | What may the agent complete, and what must it refuse or escalate? | Scope, permissions and fallback ownership are explicit. | Narrow the task; do not automate unresolved operating policy. |
| 3. Test the economics | Is AI preferable to the realistic alternatives? | A fully loaded model supports a bounded experiment. | Retain or improve human handling. |
| 4. Approve access and controls | Can this workflow operate within approved legal, security and customer-service boundaries? | Reviewers approve the actual configuration and data flows. | Keep testing synthetic and disconnected from production. |
| 5. Run the trial | Does it complete work correctly on your calls? | Transaction evidence, handoffs and economics meet agreed criteria. | Repair and retest, extend an inconclusive trial, or stop. |
| 6. Expand carefully | Does performance persist under a wider workload? | Monitoring, ownership, support and rollback are operational. | Hold the existing scope or revert. |
This gate-based approach is consistent with NIST’s emphasis on documented responsibilities, testing before deployment, measurement during operation and mechanisms to disengage systems that perform outside their intended use. It is not a NIST certification. (airc.nist.gov)