AI
Using AI to triage service requests without losing the human
How classification models can route 311-style requests, and the guardrails that keep residents from falling through.
Service request systems receive reports about potholes, streetlights, graffiti, noise, and everything else, in whatever words the resident chose. Someone reads each one and routes it. A classification model can do most of that routing in seconds, if it is deployed with the right guardrails.
The design
The model reads the request text, and any photo, and predicts the category and department with a confidence score. High-confidence requests route automatically. Low-confidence ones go to a person. Every automated decision is logged and can be corrected, and corrections feed evaluation.
Guardrails that matter
Confidence thresholds set from data. Choose the threshold by measuring accuracy on past requests, not by guessing.
Safety categories always reviewed. Anything that might be an emergency, a hazard, or a report involving a person's welfare goes to a human regardless of confidence.
Resident visibility. The resident sees the category and can correct it.
Weekly review. Staff look at misrouted requests and at the model's confusion patterns.
Fair treatment. Check that routing accuracy does not differ by neighborhood or language.
What to expect
Most requests are routine and the model handles them well. The hard cases are ambiguous, multi-issue, or emotionally charged reports, which is exactly what human staff should be spending their time on.
Multilingual
Modern models handle multiple languages well, but evaluate per language; accuracy is not uniform.
Integration
The model sits behind the service request platform's API. It should not require replacing the platform.
For procurement
Whether buying a feature of a platform or a separate service, ask for accuracy on your historical requests, how thresholds are set, what is always reviewed, and how corrections are captured.