AI use
Where a model is used, and what its output can reach
One rule governs this product: no model output may become released legal language. AI does the research: it finds sources and raises flags for a person. Every candidate is then checked before an attorney sees it; a second check by a different model family is not yet running. A model may never clear a flag, release a version, or write a clause body. This page says exactly where the line falls, surface by surface.
Last checked: 2026-09-30
The line is enforced in code, not by policy
The split below is not an editorial choice. src/lib/clause-field-roles.ts holds two lists: the fields a model could draft while the agreement tools were open, and the clause bodies it may never write. A test fails the build if the two lists ever overlap.
| Surface | Model used? | Can its output reach a document? |
|---|---|---|
| Clause language released by licensed counsel | Never | Not applicable — no model touches it at any point |
| Research for a clause, before any candidate exists | Yes — more than one model family | No — research reaches an attorney as notes and sources, never as a clause body |
| Checking every candidate before an attorney sees it | No — the checks are code; a second check by a different model family is not yet running | No — a check can flag or hold a candidate and never rewrites it |
| Agreement clause bodies: liability, force majeure, intellectual property, warranty, confidentiality, dispute resolution, state-law notice, e-consent, contractor status, general provisions, revisions, cancellation | No — fixed templates | No new document — the agreement tools are closed. The text was template text, never model output. |
| Agreement descriptive fields: title, scope of work, deliverables, payment terms, schedule | Closed — the agreement tools were retired on 2026-10-21 | No — nothing new is produced while they are closed |
| Spanish translation of a finalized agreement | Closed — the agreement tools were retired on 2026-10-21 | No — nothing new is produced while they are closed |
| Invoice generation and description enhancement | Closed — the agreement tools were retired on 2026-10-21 | No — nothing new is produced while they are closed |
| Draft dispute letters | Closed — the agreement tools were retired on 2026-10-21 | No — nothing new is produced while they are closed |
| Draft small-claims court letters | Not applicable — none is produced | No — not produced since 2026-10-21, in any state |
| Support chat | Yes | No |
| Scope clarity questions in the create wizard | Closed — the agreement tools were retired on 2026-10-21 | No — advisory only, and none is asked while the wizard is closed |
| Payment screenshot amount and date precheck | Closed — the agreement link was closed on 2026-10-21 | No — advisory only, and nothing is read while the link is closed |
| Legal research scanning and flagging | Yes | No — a model may raise a flag and can never clear one |
Which model, and what it retains
Every model call that carries what a user or a visitor types goes to Anthropic's API. Anthropic does not use API inputs to train models by default. Research on public legal sources also uses models from Google and open models served through Hugging Face; they receive public legal text and research notes, never what a user or a visitor types. Pacterms sends no clause language released by counsel to any model, at any time.