Just over a third of the top 2,000 US law firms are now using AI in production. Not evaluating it. Not running a single practice-group pilot. Using it, as part of how the work actually gets done. What's striking isn't the 36%. It's how broadly AI has spread: it has already landed in six very different kinds of work across the firm, from the lawyer's desk to the billing department.
Each one is a different bet on where AI pays off, and they demand different teams, budgets, and risk tolerance. Here is what each of the six programs actually involves.
1. Firm-built GenAI assistants
One of the most common programs is firms building their own. These are secure, in-house GPT applications that let attorneys draft, summarize, and analyze documents without sending privileged material to public tools. The appeal is control: the firm decides where the data lives, what the model can see, and how the assistant behaves, instead of trusting a consumer chatbot with a client's confidential file.
In practice this looks like a private chat interface wired to the firm's own document store, with guardrails on retention and access. It's the fastest way to give every lawyer a capable assistant while keeping work product inside the walls, which is why so many firms started here.
2. Piloting and rolling out legal AI
A large share of the activity isn't building at all. It's the disciplined work of evaluating purpose-built legal AI, running pilots inside a single practice group, and promoting the winners into firm-wide production. This is the buy side of the decision, and it's its own program because doing it well is a project: defining success criteria, measuring against real matters, and managing the change as a tool graduates from a curious experiment to part of the workflow.
Firms that treat rollout as a deliberate program, rather than letting a hundred individual subscriptions bloom, end up with fewer tools that more people actually use.
3. In-house ML and GenAI engineering
Some firms have gone further and stood up full-stack teams shipping production GenAI features. This is real engineering: embeddings, retrieval-augmented generation over the firm's corpus, and fine-tuned models tailored to how the firm's lawyers work. It's the most resource-intensive of the six programs and the clearest signal that a firm sees AI capability as something to own rather than rent.
The output is bespoke: a retrieval system that understands the firm's precedent, or a model tuned to a niche practice area that no off-the-shelf vendor serves. It's a long game, and the firms investing here are betting the in-house capability compounds.
4. Document automation and drafting
This is the program most lawyers picture first: template-driven assembly plus AI-generated memos, letters, and routine documents that still come back to a human for review. The work is high-volume and repetitive, which is exactly what makes it a strong automation target. The pattern that wins is AI doing the first 80%, fast, and an attorney owning the judgment and the sign-off.
The catch is that real document work rarely lives in one app. A draft has to pull fields from a PDF, land in the right matter folder, and get routed for review across the systems the firm already runs. That last-mile movement, not the drafting itself, is where most efforts stall. Need help with document automation? Caddi builds it from a screen recording, end to end, so the document moves through your systems without anyone re-keying it. Schedule a demo to see one of your own drafting workflows built.
5. Knowledge management and search
Every firm is sitting on decades of its own best work, and most of it is effectively lost the moment a matter closes. This program points NLP and semantic search at the firm's own work product so lawyers can find and reuse what already exists: the brief that won, the clause that survived negotiation, the memo that answered this exact question two years ago.
Done well, it quietly raises the floor on every piece of work. A third-year stops reinventing an argument a partner perfected a decade ago, because the system surfaces it in seconds. The value isn't a flashy demo; it's the slow compounding of a firm that stops solving the same problem twice.
6. Back-office and finance automation
The last program has nothing to do with the practice of law and everything to do with the cost of running the firm: automating billing, accounts payable, intake, and general-ledger processes with OCR and robotic process automation. It's the least glamorous program and often the fastest to pay back, because the work is bounded, repetitive, and currently done by hand.
Every invoice reconciled by a person, every intake form re-typed into the system of record, every payment posted manually is non-billable time the firm never recovers. Moving that work off people lands directly on margin. Need help automating back-office and finance work? Caddi turns billing, AP, and intake processes into unattended automations across your existing tools. Schedule a demo with Caddi to start with your highest-volume process.
The real takeaway
The headline isn't any single number. It's the breadth. In just a couple of years AI has gone from a novelty to something firms apply across six very different kinds of work, from the lawyer's first draft to the finance team's ledger. Most firms have committed to one or two of these programs. Almost none are running all six. The firms that pull ahead over the next few years won't be the ones with the single best model. They'll be the ones that figured out which of these programs to run in parallel, and sequenced them so the fast-payback work (documents and the back office) funds the more ambitious work (in-house engineering) that comes after.
Caddi turns screenshares into AI automations
Two of these six programs, document automation and back-office and finance work, are exactly what Caddi was built for. Instead of hiring an RPA developer or asking your team to learn a builder tool, someone screen-shares the task once. Caddi writes it as deterministic code and runs it unattended on a schedule, with the audit trails an AI committee will sign off on.
Microsoft Outlook
REC- Microsoft Outlook
Explore real legal workflows Caddi runs today, see the law-firm overview, or book a demo to watch one of your own document or back-office workflows built from a screen recording.
Frequently asked questions
What percentage of law firms use AI?
About 36% of the top 2,000 US law firms are now using AI in production rather than only piloting it. What stands out is the breadth: AI has already spread across six very different kinds of work — firm-built GenAI assistants, structured pilots of purpose-built legal AI, in-house ML and GenAI engineering, document automation and drafting, knowledge management and search, and back-office and finance automation. Most firms have started one or two of these; almost none are running all six.
What are the main ways law firms use AI?
There are six major programs. (1) Firm-built GenAI assistants: secure in-house GPT apps for drafting and summarizing without sending data to public tools. (2) Piloting and rolling out legal AI: evaluating purpose-built vendors and promoting the winners firm-wide. (3) In-house ML and GenAI engineering: full-stack teams shipping production features with embeddings, RAG, and fine-tuned models. (4) Document automation and drafting: template assembly and AI-generated routine documents that still come back to a human. (5) Knowledge management and search: semantic search across the firm's own work product. (6) Back-office and finance automation: billing, AP, intake, and ledger processes handled with OCR and automation.
Which legal AI programs deliver ROI the fastest?
Document automation and back-office and finance automation tend to pay back first. Both target high-volume, bounded, repetitive work where the inputs are structured-enough and the governance is lighter than substantive legal drafting. That makes the time saved easy to measure in a single quarter, which is why firms that sequence their rollout by business function usually start there before tackling open-ended practice work.
Do law firms build their own AI or buy it?
Both, and that mix is part of the story. Many firms are building their own secure GenAI assistants, others run structured evaluation and rollout of purpose-built vendors, and a meaningful number have stood up in-house engineering teams shipping production GenAI features. Buying gets a firm moving quickly on a known use case; building keeps sensitive work product off public tools and lets the firm tailor the system to how its lawyers actually work.
How can a law firm automate document and back-office work without developers?
With a record-to-code approach, a paralegal or billing clerk screen-shares the task once, it's written as deterministic code, and it runs unattended on a schedule with audit trails. That removes the need for an RPA developer or a team learning a builder tool, and it's how Caddi turns document automation and back-office processes like billing, AP, and intake into reliable automations across the systems a firm already runs.
See the 6 major ways law firms use AI
36% of the top 2,000 firms run AI in production, across these six kinds of work. Drop your work email to see what each slice is and how firms are using it.
One email unlocks the full guide. No spam, unsubscribe anytime.
Do more with less
See Caddi in action
Tell us where to reach you and the calendar opens right here. In 30 minutes we'll show you how Caddi automates the back-office work that grows with your clients—built, run, and maintained for you.