Choose a decision you can score
Start with a bounded task, such as assigning an incoming support request to a queue. Define the allowed queues, what evidence selects each one and when a person must decide. Keep the existing workflow running while you evaluate the router. The goal is to learn whether its mistakes and review workload are acceptable for this particular task.
Google’s Gemini Decisions API announcement makes this timely. Its dedicated model scores candidate choices and returns confidence values across multimodal inputs. The workflow below is an evaluation method, not a hands-on review or a claim that the model meets your requirements.
Write the labels before calling the model
Collect representative historical cases you are allowed to use. Include ordinary traffic, ambiguous requests, missing information and rare cases with expensive consequences. Have a knowledgeable reviewer assign the expected route and record a brief reason. Keep disputed examples in a separate review group until the policy is clear. A model cannot resolve a business rule that your team has never agreed on.
Reserve a held-out set for the final check. Use a different set to improve question wording and thresholds, so repeated tuning does not quietly turn the test into training material.
Count the mistakes that matter
Compare predictions with the reviewed labels. Track results separately for each queue, including how often urgent cases go to routine handling and how often routine cases trigger escalation. Review the actual disagreements with a domain expert. A high overall accuracy can hide a weak route when most requests belong to the easy class.
Record invalid responses, timeouts, latency and cost alongside classification errors. An unavailable router still needs a defined fallback. For attached files or images, include unreadable inputs and check that missing evidence leads to review rather than an invented answer.
Check confidence against outcomes
Group predictions into confidence bands and compare each band with the fraction of decisions that were correct. If high-confidence predictions often fail, treat the score as uncalibrated for your task. Pick thresholds using the development set and the cost of each error. Send low-confidence, unsupported and disputed cases to a human queue. The application should enforce this fallback even if the model always returns a choice.
Run beside the existing workflow
In shadow mode, log the proposed route without changing where requests go. Compare it with the route a reviewer actually chose, and inspect changes in traffic mix. Record the model version, question text and policy version so a later failure can be reproduced. Keep sensitive content out of logs unless your approved process requires it.
After the held-out check passes your pre-agreed criteria, allow a limited set of reversible routes. Keep consequential actions behind review, monitor overrides and provide a switch back to the original workflow. Re-run the evaluation when the model, policy or input distribution changes.
Read the November 12 AI briefing for today’s model, research and enterprise-agent updates.
Questions
- How do you test an AI decision router before giving it control of a workflow?
- Start with a bounded task, such as assigning an incoming support request to a queue.