Anne for Legal
Anne is an AI legal-review workspace that runs inside the firm's own environment. It finds record inconsistencies and missing evidence in title work; the attorney decides what they mean.

Some legal matters are hard to use with hosted AI — because transmitting the matter can itself create the problem. A sealed probate proceeding, a protective order, an unannounced acquisition, a client agreement that names approved subprocessors: none of these are solved by a better privacy policy, because the obligation attaches to the sending, not to what happens after.
Anne is an AI legal-review workspace that runs inside the firm’s own environment. It examines the record, surfaces where it doesn’t close, and refuses to make the legal judgment. The attorney does that part.
The whole product is five verbs
Assemble — documents become source-linked records. Compare — Anne surfaces record inconsistencies and missing evidence. Review — counsel inspects the underlying evidence. Decide — the attorney records the disposition and the evidence relied on. Release — only attorney-approved output can become final.
If one sentence is worth remembering: Anne finds questions. The attorney answers them.
Where the data lives
Everything the matter touches — documents, Anne, the models, the evidence store, the audit trail — sits inside a firm-controlled environment. The drafting model and the independent verification model both run on the firm’s own appliance. Cloud fallback is disabled and fails closed: if the local model is down, Anne stops working rather than quietly reaching out. Export without attorney release is blocked.
One arrow leaves the building: a public citation string, sent to a case-law lookup connector only when an attorney explicitly asks for research — never matter text. Turning that connector off takes the whole workflow offline, by design.
Shield: the verification and release gate
Shield is a function inside Anne — fourteen checks a draft has to clear before an attorney can release it, in four groups:
- Private execution boundary and provenance — confirms local routing and that every cited record resolves to a hashed source.
- Citation identity and authority-text verification — retrieved text is checked against the source; nothing is quoted from memory.
- Contradiction, entailment, and fidelity — a second, independent local model checks every claim against the record, catches quoted text that doesn’t match its source, verifies numbers and dates, and flags material facts present in the record but missing from the draft.
- Attorney release gate — export stays locked until a named attorney reviews and releases. A machine failure can’t be waived.
Every layer has an explicit state — passed, flagged, needs review, blocked, or unverified. A layer that doesn’t run is never converted into a pass.
New evidence reopens review, it never edits history
Anne’s audit trail is append-only. When new evidence arrives, it creates a new evidence snapshot; any decision made against the old snapshot is marked for re-review, and any already-released draft stays pinned to the snapshot it was released against. Nothing gets rewritten — a later correction is a new record, not an edit to an old one.
A synthetic case: Meridian Yard
Everything below is fabricated for demonstration — no real party, instrument, or registry reference. It’s the standard case Anne v6 opens with, and every number here can be reproduced by running the product.
An $18.4 million acquisition of a former shipyard parcel in Quincy, Massachusetts. Eighteen years — 1946 to 2026 — of assembled record: 16 instruments, 8 hashed sources, 7 exceptions, 6 of 7 configured searches not provided.
One exception, concretely: there’s a mortgage on record — Book 9012, Page 118. There’s also a discharge on record — but it references Book 9012, Page 181. A discharge exists. The reference doesn’t match.
Notice what Anne does not do here. It does not say the mortgage is outstanding, or that title is defective. It surfaces the mismatch and asks counsel what it means — because maybe there’s an explanation sitting in evidence nobody has assembled yet, and that’s exactly the kind of question a lawyer should hold, not a model.
Compared by architecture, not vendor claims
Four ways to work a file like this: a hosted legal copilot, a generic assistant, manual paralegal review, or Anne on the firm’s own appliance. The first two fail on the earliest architecture questions — where is the matter processed, can the work happen at all under a sealed order, does an unannounced deal stay contained — regardless of how good their models are. Manual review keeps everything inside the firm but has no automated comparison. Anne keeps the documents inside, compares the record automatically, and enforces attorney control at the release gate.
To be plain about the limits: Anne is not a title opinion, not a title search, and not a substitute for judgment. Missing evidence is reported as missing — never as a clean search.
Seeing it
The ask is small on purpose: bring the shape of a matter, not the matter itself. Twenty minutes, Meridian Yard end to end on a local box, every check included — including the ones that fail. No client documents, no commitment.
Get in touch to set up a time.
Frequently asked questions
- Does Anne issue a title opinion?
- No. Anne organizes evidence, detects record inconsistencies, tracks missing searches, and prepares the attorney’s work. It does not determine whether title is clean, defective, marketable, insurable, or fit to certify — a conclusion guard scans every draft for that kind of language and blocks export if it appears.
- Does Shield guarantee the draft is correct?
- No. Shield runs explicit checks and blocks release when a requirement isn’t met. It does not replace professional judgment, and a layer that fails to run is reported as unverified — never silently treated as passed.
- Why two local models instead of one?
- Drafting and independent verification of claims are kept separate, so the draft is never treated as its own check. The deterministic checks — citation identity, numbers, dates — sit apart from both models.
- Our vendor says they don't train on our data. Doesn't that solve this?
- That’s a claim about what happens to data after it’s received. The restrictions Anne is built around — a sealed probate matter, a protective order, an unannounced acquisition, a client agreement naming approved subprocessors — attach to the act of transmission itself, not to what a recipient does afterward. Encryption, retention policy and SOC 2 are controls on the recipient; they don’t undo the fact that the file left the building.
- Isn't the appliance a single point of failure?
- It’s hardware the firm controls, backed up on the firm’s own schedule, and it keeps working through a registry outage or a disconnected network. Availability is an ordinary IT question. Confidentiality is not recoverable after the fact — that’s the problem Anne is built to avoid creating in the first place.