Citizen Services

A permit application stalls, and the resident calls the office at four in the afternoon. Whatever the assistant says next becomes the position of the agency. Program owners who put AI into that conversation need it governed before it starts.

A resident cannot shop around for a different benefits office. That is the whole reason a program owner treats an assistant in the intake queue as a control problem rather than a productivity tool. The person on the other side of the counter has no alternative provider.

Two redaction patterns match the material that intake work actually produces. One masks a case, docket or matter number. The other masks a name that follows a benefits recipient label. Three blocked phrases stop a request that mentions controlled unclassified information, a sealed filing or an investigative file.

After the model answers, three checks run in order. Grounding measurement flags a claim that the supplied context does not support. Bias monitoring works from the built-in signal, and the starter sets no threshold of its own. Output scoring then holds the answer to a combined 0.86, with relevance at 0.78 and accuracy at 0.80.

History capture is switched on and writes a redacted copy of the exchange. Blocked requests are captured too, which matters when a resident asks why a service refused to answer. Capture does not fail open, so a failure to record stops the request instead of hiding it.

Eligibility rules, notice duties and appeal rights sit outside this starter. Nothing here decides whether a resident qualifies for a service. A caseworker approval path is available to build, but the review stage ships empty and runs nothing yet. Audit marking notes an evaluated request without creating a record store.