AI for law firms

AI for law firms belongs in one brain, not a pile of legal-AI apps.

That brain talks to Clio, Filevine, email, and billing — the stack you already run. Purpose-built with the owner.

Stay in ChatGPT or Claude. Boxed legal-AI apps cannot match those desktop apps. Same team can run the website, SEO, ads, portal, and onboarding. People keep the phones.

Request demo Managed Firm Brain

Built and run inside Conduit Law: 1,000+ active matters, a sitting non-attorney COO, and receipts, not slides.

Updated September 29, 2026 — added the on-premise vs cloud AI decision for law firms

The honest answer

What AI should a law firm actually use?

A frontier model you control, in a harness that connects it to your real firm context. As of this writing, that means Claude Opus or GPT-5 (whatever the current frontier is when you read this), running in a setup where you can see the folder it works in, watch the browser sessions it runs, and swap the model when the next one lands.

Skip the wrapped “legal AI” that will not tell you which model it uses. If you cannot choose the model, you are not on the frontier. You are on last year’s.

Then build the harness: one firm brain that reads across your case system, documents, email, SMS, and billing. That layer is what makes a frontier model useful for your firm instead of just useful in a chat window. The definition is what is a firm brain. Buying another legal AI seat is not that layer — see legal AI software vs a purpose-built firm brain.

Paralegals and firm ops: Legal AI for paralegals — ranked from the COO / Firm Manager seat, not a DIY lawyer substitute.

Legal operations, not a gadget list

What “AI in legal operations” is actually asking.

The page still answers “AI for law firms.” The queries that moved here in the last window are operations queries: the firm job, the closed loop, and whether you need a box in the closet.

AI in legal operations is the firm job, not the desk job

The queries that now land here are “AI in legal operations,” “AI for legal operations,” and “AI tools for law firm operations.” That is not a writing-assistant search. Operations means intake logging, records chase, provider letters, dormancy, client-update drafts, billing hygiene, and the handoff between those jobs. A chat window can help one lawyer think. A firm brain is what makes the same model useful across the stack the whole firm already runs.

On-prem and “closed” AI are the wrong closed loop

“On premise AI for law firms” and “closed AI systems for law firms” show up because owners want the file to stay inside a boundary they can explain. Buying a GPU rack is one way to feel closed. It is usually the expensive way. The loop that matters is: you choose the model, you bind it to approved context, you watch the work, and a human releases every client-facing or filed action. A wrapper that will not name its model is not closed. It is opaque.

On-premise AI for law firms

Do you need on-premise AI? Usually not. Here is how to decide.

Owners ask about on-prem because confidentiality is the real question. Where the model runs is only one part of the answer. What matters more: who can see the data, whether it trains anything, which model you are on, and who releases the work. Here are the three setups firms actually choose between.

On-premise model

An open-weight model running on a server or workstation the firm owns.

Protects: Prompts and files never leave your network.
Costs: Hardware up front, plus someone to patch, secure, back up, and upgrade it. Open-weight models trail the frontier, so drafts need more review.
Fits: A client contract, protective order, or regulator that forbids third-party processing. Scope it to those matters.

Private enterprise or API account

A frontier model (Claude, GPT) under business terms, reached through an account the firm controls.

Protects: Contractual no-training on your data, admin controls, and retention you can set and explain to a client.
Costs: Per-seat or per-use fees. You still have to decide what context the model sees and who releases its work.
Fits: Most owner-led firms. It is the current model, and the boundary is written down.

Legal-AI wrapper

A vendor app built on someone else's model, often without naming which one.

Protects: Whatever the vendor's terms say, plus whatever its model provider's terms say. Read both.
Costs: Per-seat pricing built for BigLaw, one narrow job per seat, and a model you cannot swap.
Fits: A single task the firm has already scoped, where the vendor will name its model and subprocessors.

Five questions to answer before you buy hardware

Does any client engagement letter, protective order, or carrier contract forbid third-party processing? If not, on-prem is a preference, not a requirement.
Which model runs, and can we change it? If the answer is a brand name, not a model, keep asking.
Is our data used for training, how long is it retained, and who are the subprocessors?
Who patches, monitors, and backs up the box if we host it? A closet server nobody owns is a bigger exposure than a signed cloud agreement.
What can the AI send, file, or write without a person? The answer should be nothing.

If question one comes back yes, scope on-prem to the matters that require it instead of moving the whole firm onto a slower model. If it comes back no, the closed loop you can defend is the one on this page: a named model, approved context, work you can watch, and a person who releases it.

Why the listicles miss

Four reasons the “top 15 AI tools” posts steer you wrong.

The dominant search result for “AI for law firms” is a vendor guide. Read a few and the same problems show up.

Vendors rank themselves

Clio's guide ranks Clio. Spellbook's guide ranks Spellbook. If the tool wrote the review, it is a brochure, not advice.

The tools rot fast

The "best AI for law firms" in 2024 is not the best in 2026. Frontier models jump every quarter. Wrapped tools trail whatever the model underneath was six months ago.

You do not control the model

Legal AI wrappers pick the model for you and rarely say which one. If you cannot swap to the current frontier model, you are not on the frontier. You are on last year's.

They solve one job

A drafting tool drafts. A research tool researches. Neither reads across your case system, documents, email, and billing at once, which is where most real answers live.

Where AI belongs, and where it does not

The line most guides never draw.

The rule is simple. AI everywhere behind the phone. People on the phone. Approvals on everything that leaves the firm.

AI belongs here

Behind the phone (yes)

Drafting, intake logging, records tracking, provider letters, subrogation, demand prep, client updates, follow-up. Everything a firm produces on paper or in email. AI belongs here.

People, not AI

On the phone (no)

Live client calls belong to people. An injured personal-injury client who feels handled by a bot calls the next firm fast. We recommend IVR, smart routing, and human backup, not an AI voice on your intake line.

AI belongs here

Cross-system reasoning (yes)

Reading a Clio matter, the dec page in Dropbox, and the carrier email in one supervised context to catch a mismatch. This is the job a firm brain is built for: one operating picture across the systems this firm actually runs.

People, not AI

Unsupervised sends (no)

AI never sends a client message, files anything, or writes to the case system on its own. Drafts land in your team's email drafts or behind approve, reject, and retry controls. A person always releases.

On AI risk

Hallucinated citations are a workflow problem, not a model problem.

The scary AI-in-court stories all follow the same pattern: someone used ChatGPT raw, without a source, without a gate, and let it send. The bar warnings and ethics opinions are about that shape of use.

The workflow answer is not to avoid AI. It is to design AI in so the risky pattern cannot happen. FirmOps builds the brain to work read-first (the agent cites its sources), draft-only (nothing sends autonomously), and approval-gated (a human clears every client-facing or filed thing). If you never let AI send or file without a person, the risk drops to normal-tool risk.

You still need judgment. You still need attorney sign-off on legal work. What you do not need is fear that a chat window will get you disbarred.

What FirmOps actually builds

The layer, in the spokes that matter.

The Managed Firm Brain is the deliverable. These are the surfaces where owners feel the difference first.

Managed Firm Brain

A firm brain that does the work

One layer across your case system, documents, email, and billing that reads the file, prepares the next step, and drafts the work for approval. Not an AI receptionist: your people own client calls, the brain runs everything behind them.

Explore the Managed Firm Brain →

Managed AI agents

Supervised agents that run the busywork

AI workflow automation for law firms: dormancy digests, intake-to-matter prep, client-update drafts, document and records follow-up, all read-first and behind human approval gates.

Explore Managed AI Agents →

Legal AI

The category map, not a vendor list

Legal AI covers assistants, case management, documents, and intake. Start at the hub, then pick the job you actually have.

Explore Legal AI →

AI for lawyers

The desk job is not the firm job

Owners search AI for lawyers when they want a thinking partner. That is a different page from the operating layer for the whole firm.

Explore AI for lawyers →

Legal AI adoption

A safe rollout path for legal AI tools

Claude, ChatGPT, and legal AI assistants can help when the firm controls the data boundary, review path, and approval gates. Start with supervised drafts and staff-visible proof before expanding.

Explore Legal AI Adoption →

Integration builds

We connect AI to the systems you already run

Custom AI development and integration: we build the connections (including MCP) that let a supervised agent read Clio, Dropbox, Gmail, and your task and reporting tools as one operating layer.

See the Clio automation map →

SmartAdvocate + AI

Runs SmartAdvocate? The layer still works.

Tool-agnostic by design. If SmartAdvocate is your case system, we sit the firm brain on top of it and connect documents, email, and billing across the rest of your stack.

See the SmartAdvocate map →

Filevine + AI

Runs Filevine? Same layer, same stack.

Filevine AI, including LOIS, is Filevine's product. FirmOps builds the firm brain on the stack you already run, including Filevine, and connects documents, email, and billing around it.

See the Filevine map →

CRM automation

Turn lead follow-up into an accountable handoff

Law firm CRM automation should control source, owner, qualification, follow-up, and the case-management handoff before any AI writes to the record.

See the CRM automation map →

Email automation

Turn inbox noise into supervised next steps

Law firm email automation should triage messages, draft from source context, route tasks, and keep client-facing sends behind human approval gates.

See the email automation map →

Case management automation

Show the next action before changing the file

Legal case management automation should surface matter status, stale tasks, missing context, and client-update drafts before any system-of-record write-back.

See the case management automation map →

Legal document automation

Real firm documents, drafted from the real file

LORs, provider letters, subrogation, demand prep, and client updates drafted from the actual matter and staged in staff email drafts for approval. Not a merge template.

See the document automation map →

PI firm focus

Start where personal-injury work drops balls

AI for personal injury law firms should begin with intake, records, treatment follow-up, and case-status visibility. Read-first, source-aware, and approval-gated.

See the PI firm AI guide →

Why FirmOps

AI for law firms, from an operator who runs one.

FirmOps is led by Jonathan Mahler, founder of Conduit Law (non-attorney) and COO of a working personal-injury firm. The system is built and run on real matters, so the advice and the agents are tested against the firm’s actual operating reality, not a slide.

Read the Conduit proof

Built and run inside a real personal-injury firm, not a lab. 1,000+ active matters
Read-first by default: the agent answers across the stack before it changes anything
Every write stays behind an explicit human approval gate
Named operator accountability: a sitting law-firm COO, not an anonymous vendor

Common questions

Straight answers on AI for law firms.

What is the best AI for a law firm?

The best AI is whatever the current frontier model is (right now that is Claude Opus and GPT-5) with a harness that connects it to your real firm context. Wrapped legal AI tools rarely say which model they use and lag the frontier by months. Control your model, control your context, and the tool question mostly goes away.

Are Harvey, Casetext (CoCounsel), Lexis+ AI, and Spellbook worth it for a small firm?

For a small owner-led or personal-injury firm, they are usually overhyped and overpriced. They were built and priced for BigLaw hourly work. You cannot choose the model, and you pay per seat for one narrow job. A frontier model in a real harness (working inside your folder, doing research via a browser, drafting from your file) does more for less.

How are law firms actually using AI right now?

Two patterns. The visible one: staff use ChatGPT or Claude in a browser tab, ungoverned. The one that scales: a supervised firm brain that reads across the case system, documents, email, and billing, and drafts work for human approval. The first is a lawsuit waiting to happen. The second is how a small firm actually gets leverage.

Is AI in a law firm risky? What about hallucinated citations?

The risk is real, and it is a workflow problem. Hallucinations happen when AI writes without a source and sends without a person. The workflow answer is read-first (the agent cites its sources), draft-only (nothing sends autonomously), and approval-gated (a human clears every client-facing or filed thing). If you never let AI send or file without a person, the risk drops to normal-tool risk.

Should AI answer my law firm's phones?

No. For live personal-injury client calls, AI receptionists are not there yet. Use IVR menus, smart routing, and human call-center backup. An injured caller who feels handled by a bot calls another firm. Put AI everywhere behind the phone (drafting, intake logging, records, follow-up, billing), and keep people on the line.

Does AI replace paralegals or lawyers?

No. It replaces the prep that eats their time. A well-built firm brain does the reading, the drafting, and the tracking so paralegals and lawyers spend more time on judgment and client work. Nothing about this model removes attorney sign-off on legal work.

How does AI integrate with Clio, Filevine, or SmartAdvocate?

Through APIs. All three expose enough for a firm brain to read and (with approval) write. FirmOps is tool-agnostic and builds on whatever you run. See the Clio, Filevine, and SmartAdvocate maps for how the same layer sits on each stack. The first thing we do on a fit call is confirm your systems have the APIs we need.

How do small firms compete with BigLaw using AI?

Small firms should not try to buy BigLaw legal AI. They should own the layer BigLaw firms rarely build: a firm brain across the systems the whole firm uses. Small teams turn faster when the layer is theirs. That is the leverage.

Where do I start?

A short fit and feasibility call. Every firm's stack, comfort level, and biggest bottleneck is different, so the honest answer varies. On the call we show you a real firm brain working, map your stack, confirm the APIs, and pick the one workflow worth building first.

What does AI in legal operations actually mean?

Legal operations is the work behind the phone: intake logging, records, provider follow-up, status drafts, billing hygiene, and handoffs. AI belongs there when it reads the real file and prepares the next step for approval. It is a different job from a lawyer asking a chatbot for a memo.

Do law firms need on-premise AI?

Most small and mid-size firms do not. On-premise AI means running an open-weight model on hardware you own, which keeps prompts off third-party servers but trails the frontier models and needs someone to patch, secure, and upgrade it. For most firms the better closed loop is an enterprise or API account that contractually does not train on your data, a named model, approved context only, logged work, and a human release on anything that leaves the firm. Go on-prem when a client contract, protective order, or regulator actually requires it, and scope it to that matter set.

See it live

Watch AI actually running a law firm.

jonny.bot is a public-safe, chronological log of a FirmOps-managed AI agent working inside a real personal-injury firm. Real evidence, not a roadmap.

Open jonny.bot →

Where do I start?

Book the fit call. We will tell you honestly.

Every firm’s stack, comfort, and biggest bottleneck is different, so the honest answer varies. On the 30-minute call we show you a real firm brain working, map your stack, confirm your systems have the APIs to build yours, and pick the one workflow worth doing first.

Request demo How we build and manage it