AI & Automation
Generative AI in agency operations: where it helps today
The agency tasks where large language models are already useful, the ones where they are not yet, and how to tell the difference.
Large language models can read, summarize, draft, classify, and answer questions in plain language. For agencies that run on documents, forms, and correspondence, that is a meaningful capability. It is also easy to apply badly. This article describes where the technology fits agency work now.
Uses with a clear return
Summarizing and drafting. Staff reports, meeting minutes from transcripts, plain-language summaries of technical documents, first drafts of correspondence. A person reviews and edits; the model saves the blank-page time.
Answering questions over agency documents. An assistant that answers staff questions from the municipal code, policy manuals, or contract files, with citations to the source passages. This is retrieval-augmented generation, covered in its own article.
Classifying and routing. Sorting incoming service requests, emails, or public comments by topic and urgency, with confidence scores and human review of low-confidence items.
Extracting data from documents. Pulling fields from permit applications, invoices, and forms into structured records, replacing manual keying.
Accessibility and language access. Draft alternative text, plain-language rewrites, and first-pass translations, all reviewed by a person.
Uses that need more caution
Decisions that affect a person's benefits, enforcement, or safety. Anything that requires the model to be right without a human checking. Public-facing chat without strict grounding in agency content. Uses involving criminal justice, health, or other regulated data unless the platform is approved for it.
What agencies should require of any implementation
- Grounding in agency sources with citations, not open-ended generation
- A human responsible for every consequential output
- Logging of prompts and outputs for audit and public records
- Data handling that keeps agency information within an approved boundary
- Evaluation against a test set before deployment and monitoring after
- Plain-language documentation of what the system does and its limits
How HQT approaches it
We start with a defined task, measure how well the model does it against real examples, and build the review workflow before the model goes into use. We deploy within the agency's cloud boundary where data sensitivity requires it. Every system ships with documentation a council member could read.