The craft · AI-for-X · lawyersFor solo and small firmsPairs with /comparison/advisor-vs-lawyer

AI-for-X · lawyers

AI inside a law firm: where it pays, where it harms, where it never goes.

Three lawyers have already been sanctioned for citing cases that did not exist. The risk is not that AI is unusable. The risk is that it is fluent, confident, and wrong on exactly the surface lawyers are paid for. This manual draws the line, names the workflows, and gives you the second-brain stack a small firm can actually run.

AudienceSolo and small-firm lawyers
First deploy~2 weeks
Hard floorBar rules & privilege
See the scope diagram Founder-side legal craft
Lawyer running an AI-supported research workflow.

What this work actually is

AI inside a law firm is back-of-house throughput on a workflow with a hard professional-conduct floor.

AI for lawyers is the deliberate use of LLMs on the research, drafting, review, and document-automation work that surrounds the legal opinion. It does not write the opinion. It does not sign the brief. It does not appear in court.

The hard floor is set by the bar. Privilege, confidentiality, competence, and supervision are not negotiable, and no vendor terms shift the duty back to the model. Workflows that respect the floor compound. Workflows that do not produce sanctions.

An AI hallucinated case is the same misconduct as a lawyer-invented case. The bar does not care which keyboard typed it.

The scope diagram

Three columns. Memorize the third one first.

Human owns

Where the lawyer does the work and AI does not enter the room.

  • Final legal opinion to a client.
  • Court filings without independent citation verification.
  • Negotiation strategy on a live matter.
  • Conflicts decisions and intake judgment.
  • Settlement authority and client communication on a sensitive call.
  • Anything that would be unauthorized lawyering by a non-lawyer if delegated.
Non-delegable · signed by the lawyer
AI assists

Where AI is a tool that the lawyer reads, edits, and signs off on.

  • First-draft research memos, with mandatory citation verification.
  • Contract review against a checklist of standard clauses.
  • Deposition transcript summary for the lawyer's read.
  • Document automation for routine matters.
  • Plain-English client communication drafts.
  • Discovery first-pass review with second-pass by lawyer.
Drafted by AI · verified and signed by the lawyer
Never AI

Where the file does not enter a model under any vendor terms.

  • Privileged communications into a public consumer chat.
  • Client identifiers in any prompt to a non-enterprise model.
  • Settlement amounts and litigation strategy outside the firm tenant.
  • Citations sent to court without opening every case.
  • Legal advice generated for a non-client through a chatbot.
  • Anything that breaks duty of confidentiality without informed written consent.
Hard floor · rules of professional conduct
Figure 01 · scope diagram A drawn version of the three-column scope sits here in v2: HUMAN (sealed at top) · AI ASSISTS (read and signed) · NEVER (no terms shift this).

What changes when this is done well

~6h
Per matter, reclaimed
First-draft research and standard-clause review handled with verified AI assist.
0
Unverified citations
Every cited case opened and read by a human before the brief leaves the firm.
2 wks
Setup horizon
Tenant onboarding, voice file, citation verifier, intake disclosure.
1
Disclosure clause
Engagement letter names what AI touches and how the firm verifies it.

What you need before you start

Five prerequisites. Two of them are paperwork that has to exist before the first prompt.

01 · Enterprise-tier model with a no-train clause

Vendor terms that match your duty of confidentiality.

Consumer chats are out. Read the actual terms; default to vendors that contractually do not train on your inputs and offer a tenant your firm controls.

02 · A written firm policy

One page. What is in scope, what is out, what gets logged.

Signed by every fee earner, refreshed annually. The policy is what supervisory duty looks like in writing.

03 · A redaction discipline

Names, account numbers, and identifiers stripped before paste.

Even on the enterprise tier. Belt-and-braces. The discipline survives vendor changes; the vendor terms do not.

04 · A citation verifier in the loop

Open every case the model cites. Every time.

Use a verified-research provider, an internal checklist, or a paralegal. The verification is the workflow, not an afterthought.

05 · An engagement-letter clause

Clients informed in writing, in plain English.

What AI touches, how it is verified, the right to opt out. Drafted with risk counsel, versioned, archived.

The split workflow

Six legal workflows. Each one shows what stays human and what AI handles.

01

Human owns

Frame the legal question.

Jurisdiction, parties, posture, the actual issue under the facts. Done by the lawyer, in the file, before any prompt.

AI assists

First-draft research memo.

From a tightly-framed brief, the model produces a research outline and candidate authorities. The output is treated as a hypothesis, not a citation.

02

Human owns

Open and read every cited authority.

No exception. Every case the model named is pulled, read, and confirmed to say what the model claimed it says.

AI assists

Pull the parallel-cite and headnote.

Where the verified-research provider supports it, AI builds the cite block and pulls headnotes. Cuts the lookup cost without removing the lawyer's read.

03

Human owns

Contract review judgment.

What the deal actually is, what the client cares about, where to push, where to fold. Owned by the lawyer.

AI assists

Clause-by-clause comparison against your firm playbook.

Standard clauses flagged, deviations highlighted, missing clauses surfaced. The lawyer reads the diff and makes the call.

04

Human owns

Discovery strategy and privilege calls.

What is responsive, what is privileged, what gets logged. Lawyer-and-paralegal work.

AI assists

First-pass categorization and dedup.

Keyword and category sort across large doc sets, with the lawyer setting the rules and reviewing every privileged-flagged item.

05

Human owns

Client communication on consequential matters.

Bad-news letters, settlement recommendations, fee disputes. Drafted and sent by the lawyer.

AI assists

Plain-English status updates and routine letters.

From the lawyer's notes, the model produces a draft in the firm's voice. Lawyer edits. Lawyer sends.

06

Human owns

The signature.

The lawyer's name on the document is the lawyer's professional warranty. AI does not sign.

AI assists

Final-pass formatting, table of authorities, indexing.

The mechanical clean-up that paralegals once owned. AI handles. Lawyer reviews. Lawyer signs.

Figure 02 · workflow swim lanes A six-step swim-lane diagram lives here in v2. Top lane: lawyer. Bottom lane: AI. Hand-off arrows between every step. Built for the printed firm-policy version.

How to know AI is hurting the firm

Six tells the verification step has slipped or the floor has cracked.

Tell 01

A citation in a brief was not opened by a human.

The verifier step decayed. Stop, audit the last thirty days of filings, and rebuild the rule before the next filing.

Tell 02

Client material went into a non-enterprise model.

A confidentiality breach. Disclose to the client per state rules; review the leak, retrain the team, tighten the redaction macro.

Tell 03

A non-client got legal advice from a public-facing chatbot the firm runs.

Unauthorized-lawyering exposure. Pull the chatbot, audit the transcripts, talk to risk counsel.

Tell 04

A junior associate is shipping memos without a senior read.

Supervision duty has slipped. Restore the partner-on-every-output rule for the next quarter.

Tell 05

The AI usage log has not been reviewed in six weeks.

The supervisory artifact is not being supervised. Schedule the review on the calendar; treat the log as a regulator-grade document.

Tell 06

A client asked whether AI touched their matter and the firm could not answer cleanly.

The disclosure clause is too vague or the workflow log is missing. Fix both before the next intake.

Tools and tactics

A second brain shaped for a small firm under bar rules.

The brain is the supervisory artifact. It is also what makes the AI workflow auditable when someone asks.

The Second Brain · small-firm legal stack

Stan's adapted stack for solo and small firms

One folder per matter. Firm policy pinned at the top. Voice file for client communications. Named brief blocks for research, contract review, doc automation, plain-English client letter. Citation verifier in the loop. AI usage log filed against every matter. Quarterly regulator-grade review of the log.

  • One folder per matter, redacted prompts only.
  • Firm voice file pinned to every drafting brief.
  • Five named brief blocks: research, contract review, doc automation, client letter, discovery first pass.
  • Citation verifier integrated; no brief leaves without a checked TOA.
  • AI usage log per matter: prompt, output, who verified, who signed.
  • Quarterly review by the supervising partner.

Documented in full inside the engagement · teaser here

Tactic 02

The firm voice file

One page. How the firm writes letters and memos. Tone, refusal rules, banned phrases. Loaded as the system prompt for every drafting workflow. Stops the firm voice from collapsing into a generic legal voice across associates.

  • One page, plain text.
  • Refreshed annually by partners.
  • Versioned and archived.

Tactic 03

The citation verifier rule

One sentence in the firm policy: no brief leaves the firm without every cited authority opened by a human and confirmed to say what is claimed. Enforced as a checklist on every filing.

  • Rule in firm policy.
  • Checklist on every brief cover sheet.
  • Verified by the signing partner, not the drafter.

Tactic 04

The intake disclosure clause

One paragraph in the engagement letter. Plain English. Names which workflows touch AI, how the firm verifies output, the client's opt-out right. Drafted with risk counsel, versioned with the firm policy.

  • Plain English, two hundred words or fewer.
  • Specific workflow names.
  • Opt-out is a real toggle.

Coming soon

Three rooms held open inside this manual.

Built for firms that want the policy and the stack pre-assembled. Listed here so the road is visible.

In build

The Small-Firm AI Policy Pack

Firm policy template, intake disclosure clause, AI usage log, citation verifier checklist, voice file template. Drafted with risk counsel.

Scoped

The Specialty Brief Library

Named brief blocks for transactional, employment, IP, and disputes. Released when the small-firm pack has been stable for a quarter.

Scoped

The Supervisory Audit

A small structured engagement: read of three months of AI-touched matters, gap report against bar rules, fix list for the firm policy.

What this work is not

AI does not give legal advice. The lawyer does.

This page makes a firm faster, not larger.

The throughput is on workflow. The professional warranty is on the lawyer's signature. The comparison page sets the structural difference between protecting from outcomes and reading the decision before the outcome forms.

Read advisor vs. lawyer →
Pull AI back out of the workflow when
  • The verifier step has slipped on a brief.
  • Client material reached a non-enterprise model.
  • A regulator or bar inquiry has opened.
  • The supervisory log has fallen behind for two reviews.

When the firm question is the wrong question

Run the policy and the stack for one quarter.
If the firm is still stuck, the question is structural.

Application-gated. Personal reply within 48 hours.

Apply for advisory

Tier 01 from $2,500 · Tier 02 from $4,500 / month · All three tiers