AI Agents for Law Firms

At a 45-attorney firm doing corporate and transactional work, the vendor contract queue does not care that the two associates who normally handle first-pass redlines are staffed on a trial this month. Nineteen NDAs and MSAs sit untouched for six days while new client statements of work keep arriving. Every day they wait is a day a client asks about turnaround, and a day closer to the firm eating the review time as a write-off instead of billing it. Down the hall, a mid-level partner spends forty minutes on a vendor MSA reviewing the same limitation-of-liability language for the eleventh time this month, because someone still has to read every clause before it goes to signature, and that someone bills at partner rates. Meanwhile a paralegal is nine hours into pulling case law for a new employment matter, cross-referencing three jurisdictions by hand because the research subscription returns four hundred results with no way to rank them against the actual fact pattern. None of this is a staffing problem you solve by hiring. Associates and paralegals are expensive and slow to onboard, and a backlog does not wait for the next hire to clear the bar exam.

Why a generalist chatbot and an autocomplete tool both miss the job

Paste a vendor MSA into ChatGPT or Claude.ai and you get a competent-sounding summary of the clauses. What you do not get is your firm's actual risk tolerance. A generalist chatbot has no memory of how your firm redlined the last forty MSAs, no concept of a fallback position on indemnification caps, and no fixed playbook to apply consistently across matters. Ask it for case law and it will produce citations that read correctly and are sometimes wrong, a risk that is merely annoying in a marketing draft and career-ending in a brief filed with a court. A generalist tool optimizes for a plausible-sounding answer to the question you typed. It has no mechanism for applying a firm-wide negotiation stance, no way to flag that a clause deviates from the standard your partners agreed on, and no institutional memory of the last time this exact vendor pushed back on the same term.

Cursor and GitHub Copilot solve an entirely different problem for an entirely different profession. They are autocomplete layers built for source code inside an IDE, fast next-token suggestions for a developer writing functions. Attorneys do not work in an IDE. They work in Word, in PDF redlines, in research databases with jurisdiction filters. Some legal tech vendors have bolted an autocomplete-style suggestion engine onto a document management system and marketed it as AI for legal teams, but it behaves exactly like Copilot behaves: it completes the next clause based on pattern matching, it does not score the risk of the completed clause, and it does not know that your assignment clause fallback differs from your indemnification fallback. An autocomplete engine for law firms produces plausible boilerplate. It does not produce a redline with a risk score, a research memo with cited authority checked against your jurisdiction, or a contract draft that matches your firm's actual negotiating position.

Brief, Clause, and Cite: a legal team built for firms, not consumers

Brief is the primary agent for a firm running this workflow in Claude Code. Brief drafts and reviews the actual documents, NDAs, MSAs, employment agreements, SLAs, and vendor contracts, without producing forty pages of legalese nobody on the client side will read. The brief-recon skill reads your firm's existing template library and house style before drafting anything, so the output matches how your firm actually writes contracts rather than a generic template pulled from a training set. The brief-draft skill produces the first draft against that house style, and brief-review catches the kind of internal inconsistency that costs a partner an hour to find manually, a defined term used before it is defined, a notice period that contradicts itself between sections, a governing law clause that does not match the counterparty's jurisdiction.

Clause sits next to Brief and owns a narrower, deeper job: redlining every clause in an incoming contract and scoring the risk before a human reads it. The clause-recon skill ingests the incoming document and identifies clause by clause what it is dealing with, indemnification, limitation of liability, assignment, termination, IP ownership. The clause-analyze skill compares each clause against your firm's stored negotiating positions and flags deviations with a risk score, so a partner opens the document already knowing which three clauses actually need their judgment and which seventeen are within acceptable range. The clause-playbook skill goes one step further and drafts the counter-language your firm would typically propose, based on the negotiation patterns your attorneys have used before, so the first redline pass is not a blank Word document, it is a marked-up draft ready for a partner's ten-minute review instead of a fresh forty-minute read.

Tonone's Clause redlines every clause and scores the risk before a partner opens the document, turning a forty-minute first read into a ten-minute confirmation.

Cite covers the third leg: case law synthesis, statute analysis, and jurisdiction comparison. The cite-recon skill takes the fact pattern and matter type and identifies which jurisdictions and which bodies of law are actually relevant, cutting a four-hundred-result research dump down to the set that matters. The cite-research skill produces the first-pass memo, synthesizing the controlling case law with pin citations a paralegal or associate can verify rather than trust blindly. The cite-compare skill is what makes multi-jurisdiction work tractable, laying out how the same legal question resolves differently across two or three jurisdictions side by side, which is exactly the kind of work that eats a full day when done by hand and takes minutes when the comparison is already structured.

A worked example: clearing the vendor contract backlog

Take a mid-size firm, call it Halsted & Voss, 45 attorneys, a transactional practice group that handles vendor contracts for a handful of mid-market SaaS clients. On a Tuesday, the group has 22 vendor MSAs and NDAs sitting in the queue, the oldest six days old. The two associates who normally handle first-pass redlines are staffed on litigation discovery this week, and the managing partner does not want to tell a client their contract review is delayed another week because of internal staffing. Separately, a new employment matter just came in that requires case law across three states on non-compete enforceability post-employment, work that would normally take a paralegal a full business day of manual research across a legal database with weak relevance ranking.

The firm runs Clause against all 22 contracts first. Clause reads each one, applies the firm's stored negotiating positions on the recurring clause types, and produces a risk-scored redline for every contract in the batch. Instead of an associate spending 45 to 90 minutes per contract on a first pass, a supervising partner spends roughly ten minutes per contract confirming the flagged items and adjusting language on the two or three clauses that actually need judgment. In parallel, the firm runs Cite against the new employment matter's fact pattern. Cite narrows the jurisdiction scope to the three relevant states, pulls the controlling authority for each, and produces a comparison memo the paralegal reviews and verifies rather than builds from scratch.

Halsted & Voss, Vendor Contract Batch + Non-Compete Research

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Clause: 22 vendor MSAs / NDAs, batch redline
  Flagged for partner review:      3 contracts (unusual indemnity cap,
                                    missing mutual NDA carve-out,
                                    non-standard termination-for-
                                    convenience window)
  Within standard playbook range:  19 contracts, counter-language
                                    drafted and ready to send
  Est. partner review time:        ~10 min/contract (down from ~45-90 min
                                    first-pass associate time)
  Batch turnaround:                same afternoon vs. 6+ day backlog

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Cite: non-compete enforceability, post-employment, 3 jurisdictions
  Jurisdictions scoped:            CA, TX, NY (fact pattern matched)
  Controlling authority:           7 cases identified, pin-cited
  Cross-jurisdiction comparison:   CA (unenforceable, Bus. & Prof.
                                    Code 16600), TX (enforceable with
                                    reasonable scope), NY (enforceable,
                                    fact-specific, blue-pencil doctrine
                                    applies)
  Paralegal task:                  verify citations, confirm fact
                                    match, not build memo from zero
  Est. time:                       under 2 hours vs. full business day
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Next: partner signs off on flagged contracts, associate finalizes
non-compete memo for client call at 4pm same day.

The backlog clears in an afternoon instead of stretching into a second week, and the non-compete memo is ready for a same-day client call instead of tomorrow's. Nobody got hired. The partner still made every judgment call that actually required a partner, the three flagged contracts, the fact-pattern confirmation on the non-compete memo. What changed is that the volume of low-judgment, high-repetition work, the nineteen standard contracts, the four hundred irrelevant search results, got handled before it reached a billable human, freeing that human's hours for the three contracts and the client call that actually needed their judgment.

Tonone's Cite scopes a fact pattern to the relevant jurisdictions and produces pin-cited authority a paralegal verifies instead of a research memo built from a four-hundred-result dump.

Brief, Clause, and Cite vs the alternatives

Firms evaluating AI for legal work usually compare three categories: a generalist chatbot, an autocomplete layer bolted onto a document management system, and a purpose-built legal agent team. The comparison only makes sense once you separate what each category was actually built to do. A generalist chatbot was built to answer questions well in a single turn. An autocomplete layer was built to speed up next-token suggestions inside an editor or document tool. Neither was built to hold a firm's negotiating playbook, score contract risk against it, or scope legal research to the jurisdictions that actually apply to a fact pattern.

Tonone's Brief drafts NDAs, MSAs, and vendor contracts against a firm's own template library, not a generic legalese template pulled from a training set.

Install and try

Tonone is free and MIT-licensed. Install it once and Brief, Clause, and Cite, along with the rest of the Tonone team, are available in your Claude Code session. You pay only for the Claude Code token usage during the work itself.