Courts sanction lawyers for citations that do not exist.
In legal work the failure is found by an opponent, in public, with your name on the filing.
Building AI for law firms rather than practicing? Yours is sell into the enterprise.
Three rules a firm needs
| Risk | The rule |
|---|---|
| A fabricated citation | a rule requires a citation, and where you supply a lookup function it checks that the citation is found in the source you point it at. An answer that fails either check goes back for a rewrite, or to a person |
| Privilege | material from one matter does not enter a prompt about another. Each request names its matter, and the wall between matters is a rule that checks every request, not a folder convention |
| Conflicts | the assistant does not work on a matter the firm is conflicted on, checked on each request against the conflicts system you already run |
A rule can check both what goes into the model and what comes out: role and judge checks.
Prove what the AI did
When a client asks whether AI touched their matter, or a court asks how a document was produced, the answer has to be a record, not what someone remembers. Every decision record holds the rule that decided and the frozen version it ran on: audit and evidence.
Ready for the data-handling review and procurement
Swiftward runs on-prem, with SSO, layered access control, secrets management and an append-only audit trail: the controls that a cautious general counsel's data-handling review asks for. See security.