Vendors demo the model. The demo answers a question, drafts a letter, fills a form. What it rarely shows is everything around the model that decides whether the answer can be trusted, logged, approved and repeated next week.
Where the model belongs
Language models are good at a narrow set of back-office tasks: pulling fields out of a document, drafting text for a person to review, and summarizing source material with citations. In each, language is the input and a person checks the output.
Where it doesn't
Thresholds, routing, permissions, calculations, deadlines and approvals are rules. Rules can be written down and tested. A model's answer about which approval a $180,000 purchase needs is a guess, however confident it sounds. The rule is a lookup.
If a step can be a rule, it is a rule, not a prompt.
What that buys you
- Testability. The rules have unit tests. The one model step has an evaluation set built from real cases, rerun on every change.
- Auditability. Every prompt, retrieval and output is logged and kept on the agency's records schedule.
- Replaceability. The model can be swapped without rewriting the system around it.
- Accountability. A named employee approves every consequential action. The model never signs.
What it looks like in practice
The determinations generator Brian Hadley built at CACI International for a Washington, DC contracting office has this shape. Vendor data comes from Dun & Bradstreet, DC Clean Hands and the SAM.gov exclusions list. The specialist verifies it, and only then is a signature-ready draft produced for the specialist to sign. Every Munivera pilot follows the same pattern: deterministic intake, approved sources, one bounded model step, validation, and a person on the signature.