The boundaries come first
A language model in a product is easy to demonstrate and hard to trust. The difference is entirely in what it is allowed to reach and what it is allowed to assert — so that is what we design first, before anything is generated at all.
- The model never touches the database. It receives a bundle already filtered by tenant, by record, and by the permissions of the person asking. There is no query it can write.
- Facts are looked up, not generated. Dates, amounts, names and reference numbers come from the record and are handed to the model as facts.
- Citations are verified. An answer points at the document it came from, and a citation the model invents is dropped before anyone sees it.
- Anything with consequence waits for a human. The system raises it with a severity, never a verdict, and it stays unconfirmed until a person confirms it.
- The model is swappable. Claude, GPT, Gemini, or a model hosted on your own infrastructure — changed by configuration rather than a rewrite.
Software that is confident about the wrong Tuesday does real harm, and no amount of fluency makes up for it.
What that is good for
Intake that arrives as a conversation rather than a form. Documents that need reading before they can be filed. A pile of records somebody has to ask a question of. Routine questions that can be answered from your own data — and the honesty to refuse the ones that cannot.
This one is in development. The approach above is live in LegalFlow, which runs end to end and has 367 automated tests behind it.