Legal Matter AI
Upload a matter's documents and every passage of every one of them is read — the AI does not choose what to look at and does not stop early. It then extracts the facts, surfaces potential legal issues, weighs the supporting AND contradictory evidence, and drafts a source-cited report. An AI critic reviews the analysis first, and nothing is written until a supervising attorney approves it.
Your data: Matter documents are kept for the life of the matter, encrypted with your firm's own key. They are deleted when you delete them — never on a timer.
Available to every firm the moment you sign up. Create your practice, open a matter, add documents — there is no waitlist and nothing to enable.
The six agents, and what each one actually does
Not one model asked to do everything. Six with narrow jobs, because a single agent that gathers, analyses, argues and writes does all four badly — and there is nothing left to check its work.
1. Researcher
The only agent that touches your documents. It decides for itself what to search for next, which document types to narrow to, and when it has gathered enough — then stops. Its search budget is a hard ceiling in code, and if it hits that ceiling the report says so rather than presenting a truncated search as a complete one.
Produces: Research notes and the passages actually retrieved — never conclusions.
Cannot: choose which firm or matter it is searching. The search tool is bound to one authorised index and takes no tenant argument, so there is no instruction that widens its reach.
2. Fact Analyst
Extraction only, with no legal conclusions: the parties and their roles, key dates, events, obligations and their deadlines, undisputed facts, disputed facts with what in the record conflicts, and the damages figures the documents state. It also lists the factual questions the documents do not answer, explicitly rather than silently.
Produces: A structured factual picture where every claim carries its citation.
3. Legal Issue Analyst
Surfaces potential legal issues, each phrased neutrally, with what in the documents raises it, the provisions relied on, the points an attorney would want to weigh, and a confidence rating on whether the record supports raising it at all. It separates out the questions that documents alone cannot settle and hands them to you.
Produces: Potential issues for attorney consideration — not conclusions, and not advice.
4. Evidence & Gap Analyst
Takes each issue and weighs the record on both sides: what supports it, what contradicts it, what is missing that would resolve it, and how strongly the documents actually support it — strong, moderate, weak, or unsupported. The contradictory column is the point. A tool that only builds your case is a tool that surprises you later.
Produces: Per-issue evidence with a support level, plus specific further research worth doing.
5. AI Critic
A skeptic by construction — approval is the exception, not the default. It rejects the analysis when a claim has no citation or cites a passage that does not support it, when an issue has no basis in the documents, when contradictory evidence was plainly available and went unaddressed, when an obvious question was never researched, when a possibility is stated as a conclusion, or when a stated gap is too vague to be useful. Rejections come back as specific, answerable questions that go straight to the researcher.
Produces: A verdict the router reads as a boolean, so no prose gets reinterpreted into approval.
Cannot: reject a gap the documents simply cannot fill. Once the record is exhausted it is told so, and an honestly reported gap becomes the correct outcome.
6. Writer
Runs only after an attorney has approved, and composes — it does not analyse. It writes from the approved analysis alone, adding no fact that is not already there and no fact without the citation it already carries, because inventing at the last step would put an uncited claim into a document a lawyer signs. Your review comments are passed to it and reflected in the draft.
Produces: The final report, in fixed sections, ready to render.
Nothing decides what runs next except code
The supervisor that routes between agents is not a language model. Has the critic approved? Is the iteration budget spent? What did the attorney decide? Every one of those is a value to read, not a judgment call — and putting the loop guard, the thing that stops runaway spending, inside a model would make it something you can talk out of stopping.
The critic can send the analysis back for more research once. At the ceiling the run goes to the attorney anyway, carrying the unresolved critique, because throwing away the work would help nobody. Two passes is the whole budget — an analysis that cannot be made sound in two is telling you something about the record, not asking for a third.
The attorney gate is a real stop
When the analysis reaches review, the run halts and is written to storage. Nothing polls, nothing holds a connection open, and no cost accrues while it waits — it can sit for days. Resuming it needs the ai.approve permission, checked server-side before the run is touched, so it is a control and not a hidden button.
The reviewer sees the analysis, the critic’s verdict, and whether the critic actually approved it or the run simply ran out of passes. Approve and the writer runs. Request changes and it goes back to research carrying your comments. An unrecognised answer counts as requesting changes, never as approval.
What you get
A report with fixed sections, in this order, every time — so you always know where to look and what is missing is impossible to hide:
- Executive Summary
- Key Facts
- Chronology — in date order
- Potential Legal Issues
- Supporting Evidence
- Contradictory Evidence
- Evidence Gaps
- Attorney Review Notes — what you must decide or verify
- Source References
Every factual statement carries an inline citation — [Document name, p.12, §8.2] — and dates appear exactly as the document states them, never normalised or inferred. Contradictory Evidence is never omitted; if none was found, the report says so in as many words.
Delivered as a document, not a chat log
Rendered to PDF and DOCX and stored beside the matter’s other documents, carrying:
- A provenance block — the matter, the run, the attorney who approved it and when. A PDF outlives the session that made it, so it has to explain itself on its own.
- A SHA-256 hash — recorded alongside the file, so the report can be shown later to be unaltered. The same integrity guarantee the evidence extractor gives; the newer tool does not get a weaker one.
- The disclaimer, in the document — AI-assisted attorney work product requiring professional review. Not legal advice.
What it will not do
Worth as much as the feature list. A tool that is vague here is asking you to find its limits in front of a judge.
It cannot look outside the matter
Each matter has its own index. The agents search that one and have no way to name another — not another matter of yours, and certainly not another firm's. It is a boundary in the plumbing, not an instruction in a prompt.
It does not research case law
There is no external legal corpus behind it. Everything it says comes from documents you put in the matter. Case-law research is a separate tool on the roadmap, and conflating the two would let it cite authority it has never read.
It does not give legal advice
The output is draft attorney work product requiring professional review. The prompts, the data structures and the report itself all use tentative language deliberately — 'potential issue', 'considerations', 'matters requiring attorney judgment'.
It cannot finish without you
There is no autonomous mode, no scheduled run that files itself, and no configuration that skips the attorney gate. If nobody with approval authority reviews it, no report is ever written.
What it costs
An analysis costs what it costs to run, so your firm holds a prepaid credit balance rather than a subscription. A typical matter analysis runs $1.45–$2.04. Before a run starts you see an estimated range and your available balance, and you approve it. The estimate is held while the run works; when it finishes, only what was actually used is charged and the rest is released immediately.
Starter
$25.00
2,500,000 credits
roughly 12–17 analyses
A good way to try it on a real matter.
Practice
$100.00
11,000,000 credits
includes 1,000,000 bonus credits
roughly 53–75 analyses
10% more credits included. Sized for a firm running analyses every week.
Firm
$400.00
50,000,000 credits
includes 10,000,000 bonus credits
roughly 245–344 analyses
25% more credits included. For a litigation practice running matters continuously.
Derived from the same estimator that quotes your run, for a 10-document, 50-page matter. The spread is real: an analysis the AI critic sends back for a second pass costs several times one it approves first time. A large matter with many more documents costs more again, which is why you are shown an estimated range and your balance before every analysis, and why anything the analysis does not use is returned to your balance immediately.
Credits never expire
They carry forward month to month for as long as your account is open. No monthly minimum, nothing forfeited at the end of a period. Purchases are non-refundable.
A rejected run is not a wasted one
Send an analysis back and it re-researches; the writer only ever runs once, after you approve.
Adding documents is priced too
Bringing a document into a matter costs $0.0035 a page — 35 cents per 100 pages — charged once, however many times you analyse it afterwards. A scanned exhibit also needs OCR at $0.0045 a page (45 cents per 100 pages). Both appear on the same ledger, quoted and confirmed before anything is processed.
Your firm owner sees every credit spent, itemised by agent and by run, with the rate that was applied — so any charge can be explained. Deleting a matter destroys its documents and analyses but keeps that billing record, marked as deleted, so an invoice never refers to something that has vanished.
Open your first matter today
Create your practice, add the documents, and run an analysis — you approve the estimated cost before anything runs, and an attorney approves the result before anything is written.