Legal
AI for law firms, built for the rules that bind them
The ethics rules decide your architecture before your roadmap does. Most legal AI projects stall on that, not on the model.
Five rules that change the build
5 things that decide this
- 01Efficiency gains from AI belong to the client, not the firm. ABA Formal Opinion 512 says an hourly lawyer must bill actual time and may not bill for time the tool saved.
- 02The terms your software runs under are now evidence. In February 2026 a federal court held that a defendant's AI chats were not privileged, because the tool's own policy allowed training on inputs.
- 03California says a lawyer must not let an AI system file documents, send communications or transfer client data on its own. The approval gate is the requirement, not a feature.
- 04Opinion 512 says boilerplate in an engagement letter is not informed consent. Consent has to be captured per client and per risk, with a record of what was actually explained.
- 05Your buyer is ethically obliged to audit you. That is why a security questionnaire arrives before the pilot, and why it stalls deals for quarters.
The duty to vet you rolls downhill
A firm's clients impose security obligations on the firm. The firm passes them to you. Opinion 512 makes vetting your reliability, security and breach practices part of a lawyer's own ethical duty, so nobody can wave it through to be helpful.
California adds a sharper line. Reasonable effort there requires more than reliance on generalized marketing assurances, which means somebody reads your terms and your subprocessor list. Answering those questions after a pilot is the most expensive order to do it in.
- 01Firms buy against realisation and liability. A missed deadline is the largest malpractice claim category.
- 02Legal departments buy against outside-counsel spend, which is the number firms are measured on growing. Never sell both the same win.
- 03Legaltech founders buy the product, plus the answers that get it through an AmLaw security review.
Sector context
What the research says about legal AI
>17%
Incorrect answers from Lexis+ AI and Ask Practical Law AI (Stanford RegLab, 2024)
>34%
Hallucination rate for Westlaw AI-Assisted Research (same study)
96%
Of legal buyers demand safeguards for confidential data (Thomson Reuters)
94%
Require outputs grounded in authoritative content (Thomson Reuters)
The work we take in this sector
Ordered by how much of it we have shipped, not by how well it sells.
Intake, qualification and matching
Guided intake, jurisdiction and practice-area routing, after-hours triage and the alert path that decides whether a lead gets called back. This is the Lexpair shape and the one place our proof is direct.
Legaltech product engineering
You know the law and need the product built, then cleared through a firm's review. SSO, audit logs, tenant isolation and a written subprocessor list are the questions that stall these deals.
Approval gates for agent features
Matter-scoped access instead of one firm-wide service account. A human gate on every outbound action. Configurable autonomy, because the duty of oversight rises with how much the system does alone.
Consent and disclosure capture
Recording what a client was told, when, and about which tool. A single global terms acceptance does not meet the standard Opinion 512 sets, so consent is captured per client and per risk, with the explanation stored beside it.
Payments and communications at consumer scale
Stripe and Twilio in a legal intake context, running in Lexpair today. Worth saying plainly: payment experience is not trust accounting. Client ledgers carry invariants and a regulator that Stripe does not.
Jurisdiction as a first-class filter
Law is jurisdiction-scoped, so a retrieval system that mixes jurisdictions is confidently wrong. Jurisdiction belongs in the query, never as a hint in a prompt.
- MatterScoped access, not firm-wide
- PromptDirection of counsel recorded
- EndpointWritten no-training terms
- LoggingEvery vendor is a recipient
- OutputReviewed state, or unreviewed
Transmission to a third party is the disclosure event, not storage. A tracing vendor, a prompt cache or an eval pipeline each add a recipient, and each one has to be named in writing before the architecture is fixed.
A privacy policy decided whether material was privileged
In United States v. Heppner, decided in the Southern District of New York on 10 February 2026, a defendant used a consumer AI tool to research his own case. The court held the material was not privileged. Its published terms allowed disclosure and training on inputs, so there was no reasonable expectation of confidentiality.
Work product failed on a second ground. Nobody had prepared the documents at counsel's direction. Whether an enterprise tool with no-training terms would come out differently was left open, and left open is not the same as safe. Two design rules follow. Keep the terms retrievable, because a firm may have to prove them. Record that the work was directed by counsel, in the prompt itself.
- Log which endpoint saw which matter's data, under which terms, at what time.
- Forwarding an existing AI document to a lawyer afterwards does not make it privileged.
- We build the architecture that supports the claim. Your lawyers make the claim, and nothing here is legal advice.
Intake, matching and routing systems we have shipped
“Their attention to detail, quality of employees, and work ethic were outstanding.”
Nicolas de Quesada · CEO, Lexpair
The generic approach against ours
| Criterion | A generalist agency build | How we scope legal work |
|---|---|---|
| Choosing a model endpoint | Whichever is cheapest, terms unread. | Written no-training terms, kept retrievable. In Heppner the policy decided the outcome. |
| Agent autonomy | Ship it, add approvals if someone asks. | Human approval on every outbound action. California prohibits autonomous filing and transmission. |
| Access scope | One service account across the firm. | Matter-scoped tokens. A firm-wide token breaches confidentiality and the ethical wall at once. |
| Client consent | A tick box at signup. | Per client and per risk, recording what was explained. Opinion 512 rejects boilerplate. |
| Retention against a legal hold | The nightly deletion job wins. | Holds override deletion at run time and fail closed. Deleting under hold is spoliation. |
| Selling the benefit | Bill the same hours in less time. | We say no. The rules give saved time to the client, so we build for capacity instead. |
What these builds run on
Shipped in Lexpair
Open legal data sources
Systems buyers ask us about
Questions firms and founders ask
01Is it safe to use AI on client matters?
Safety here is decided by the terms of the endpoint you send data to, which is now a legal fact rather than a preference. Heppner turned on a published policy that permitted training on inputs. Three design rules follow. Get written no-training terms you can produce later. Log which endpoint saw which matter. Allow no undisclosed third party in the path, because a tracing or eval vendor is another recipient. We are software engineers, not your counsel, and this is not legal advice.
02Can an AI assistant send emails or file documents for us?
Not on its own, and California states that directly. A lawyer must not deploy an agentic system that autonomously transmits client information, files documents or communicates with a court without review. In practice that makes the approval gate a core requirement of the build. We give every outbound action a human checkpoint and log who approved what. The guidance also says oversight duty rises with autonomy, so more capability means more gating, not less.
03Will AI let our lawyers bill more?
No, and any vendor promising it has misread the rules. Opinion 512 says hourly lawyers bill actual time and may not bill for time saved by the tool, and Florida and California agree. General AI subscriptions are treated as overhead, similar to library maintenance, while matter-specific per-use costs may be charged. The honest business case is capacity, faster response to clients, fewer write-offs at realisation, and work the firm previously turned away.
04How do we vet an AI vendor for our firm?
Ask for the four things a marketing page never contains. Written terms with the model provider on training and retention. A full subprocessor list, because every recipient is a disclosure. Whether access is matter-scoped or firm-wide. And whether the product distinguishes lawyer-reviewed output from unreviewed output in the data model, since otherwise nobody can prove review happened. California says generalized marketing assurances are not enough, so a confident answer with no document behind it fails.
05Our matter data lives in systems that do not talk. Is that an AI problem?
Usually not, and treating it as one produces a confident summary of stale data. Status sits in email, the document system, the calendar and somebody's memory. Most single-view-of-the-matter projects are integration work under another name. Fix the plumbing first, then add retrieval on top of data that is actually current. We would rather tell you that during scoping than discover it together in month three.
06What drives the cost of a legaltech build?
Scoping calls are free. Where we need to work inside your existing codebase, we run a paid two-week diagnostic that ends in a fixed price. Three things move the number. Gated integrations count, since vendors like iManage and NetDocuments do not publish readable docs without an account. How much of the workflow must be a deterministic rules engine rather than a model. And how early your buyers' security review lands. We do not quote before seeing those.
Go deeper
- Legal industry overview →Conflicts, ethical walls and the three buyers who share vocabulary and nothing else.
- Does sending client data to an LLM waive privilege? →The February 2026 ruling, and the four questions it raises for a build.
- Legal AI: build vs buy →Where Harvey and Clio win, and the narrow case for building your own.
- AI legal lead matching →Lead quality in legal services is a structure problem.
