AI for Law Firms by Practice Area With a Rollout Plan
AI for law firms mostly means faster legal research, document review, first drafts, client intake and billing work, with a lawyer checking every output before it reaches a client or a court. Use is uneven: in the ABA's 2024 Legal Technology Survey Report, 30% of responding lawyers used AI, from 46% at firms of 100 or more attorneys to 18% of solos.
Below are use cases by practice area, how sensitive each task's data is, where the tools run, and a rollout plan for firms of 1 to 100 lawyers.
How Are Law Firms Using AI Today?
Most firms use AI for five kinds of work: research, review, drafting, intake and back-office tasks. The common thread is text: AI reads a large pile of documents faster than a person, then writes a summary, a draft or a list for a lawyer to check.
- Research: finding and summarizing case law, statutes and secondary sources, with citations the lawyer verifies.
- Review: sorting discovery, flagging clauses in contracts and pulling key facts out of records.
- Drafting: first drafts of letters, motions, memos and clauses from the firm's own templates.
- Intake: answering first questions from prospective clients and turning calls or forms into a matter summary.
- Back office: time-entry narratives, billing descriptions, conflict-check prep and marketing copy.

Firm size changes the picture a lot. The same ABA survey found 30% use at firms of 10 to 49 lawyers, well behind the 46% at large firms. Three-quarters of respondents named AI hallucinations as a reason to hold back.
A small firm rarely has someone whose job is to vet tools, which is why the rollout plan further down changes by size.
AI Use Cases in Law Firms by Practice Area
AI helps most where a task is high volume and the output is easy to check. The table shows the main tasks per practice area, how sensitive the input usually is, and who checks the output.
| Practice area | Main AI tasks | Data sensitivity | Human review |
|---|---|---|---|
| Litigation | Discovery review, deposition summaries, research and briefs | High; medium for research prompts | Lawyer signs off on privilege and production |
| Transactional | Contract review, clause comparison, due diligence | High: deal terms, often under NDA | Lawyer decides pushback; sample-check extracts |
| Regulatory and compliance | Rule and guidance monitoring | Low: public sources | Check summary against the source |
| Intellectual property | Prior art search, office-action drafts | Medium to high: unfiled inventions | Attorney or agent reviews claims |
| Family law and PI | Intake, medical record summaries | Very high: health, money, family | Check every date and diagnosis |
| Back office | Time entries, billing, conflicts, marketing | Medium; low for marketing | Billing lawyer or conflicts counsel clears |
What AI actually does in each area:
- Litigation: ranks documents by relevance, flags likely privilege, groups near-duplicates, summarizes transcripts by topic, builds a dated timeline from exhibits, finds authority, drafts brief sections and suggests counterarguments. Spot-check page and line cites against the transcript.
- Transactional: compares a draft against the firm's playbook, flags missing or off-market clauses, and pulls change-of-control, assignment and term clauses across a data room.
- Regulatory: summarizes new rules and maps them to client obligations.
- IP: searches and summarizes prior art, then drafts office-action responses.
- Family law and PI: turns intake calls into a case summary and medical records into a treatment timeline.
- Back office: drafts time entries from calendar, email and document activity, pre-sorts conflict results, and drafts bios and newsletters.

Read the sensitivity column first. Regulatory summaries or marketing copy can go through almost any tool. Medical records in a PI file or a data room in an M&A deal need a tool whose data terms the firm has checked, or one that runs on hardware the firm controls.
The review column matters just as much. A task saves time only when checking the output is faster than doing the work. Clause extraction across 400 contracts passes that test; a brief built on AI research that nobody pulls and reads does not.
Types of AI Tools for Lawyers
AI tools for lawyers fall into five groups: general chatbots, legal research platforms, practice-management add-ons, contract and review tools, and self-hosted models. Most firms end up with two or three of them, not one.
- General chatbots: ChatGPT, Claude, Gemini and Microsoft Copilot. Each runs on a large language model that drafts and summarizes well but has no built-in check against a legal database, so citations need extra care.
- Legal research platforms with AI: Lexis+ AI, Thomson Reuters CoCounsel and Westlaw, Bloomberg Law, plus legal-specific assistants such as Harvey and Legora. They answer from legal sources and link the authority, which shortens the citation check.
- Practice-management add-ons: AI features inside tools like Clio, MyCase and Smokeball for intake, time entries and document drafting. The data is already in that vendor's system, but the AI feature may send it to a model provider, so check the vendor's AI terms.
- Contract and review tools: contract assistants that work inside Word, such as Spellbook, and eDiscovery platforms such as Relativity and Everlaw.
- Self-hosted models: open models run on the firm's own servers. Nothing is sent to a vendor, but the firm builds the workflow and owns the upkeep.

The category matters more than the brand. A general chatbot fits drafting and summarizing non-confidential text; a research platform fits anything that ends in a citation.
If you are picking a single tool for one lawyer or paralegal, the checklist in our AI legal assistant guide covers that choice; this page stays at the firm level.
Where the AI Runs and Why It Matters for Client Data
Where an AI tool runs decides who holds your client data while the model works on it. There are three options:
- A consumer chatbot sends prompts to the vendor under its standard terms.
- An enterprise plan sends them to the vendor under a contract your firm negotiated.
- A self-hosted model keeps them on hardware in your office.
The Illinois ARDC guide to implementing AI, released in October 2025, frames this as three steps: classify the information, assess the tool's security, then match the two. It also states that confidential data should never go into public AI tools, even with client consent.
Applied to the use-case table above, that gives a simple sort:
| Data in the task | Examples | Where it can run |
|---|---|---|
| Public or non-client | Regulatory monitoring, marketing, general research | Any tool, including consumer chatbots |
| Client confidential | Billing narratives, research with case facts, conflict prep | Enterprise or legal tools; no training, set retention |
| Highly sensitive or privileged | Medical records, discovery, data rooms, deposition strategy | Vetted legal platform or firm-owned hardware |
Running models in-house is the plain version of on-premise AI: the model weights and every prompt sit on a machine in your server room or office, and nothing goes to an outside model provider. The trade-off is that your firm, or its IT partner, now runs that machine.
Autonomous Computer 4 is a business AI computer built for that job. Autonomous says prompts never leave the building and nobody trains on your work. There is no per-token bill, and the 3-year warranty covers the hardware.
- GPUs: 4x RTX 5090 (128 GB of GPU memory) or 4x RTX PRO 6000 (384 GB).
- Software: Ubuntu, NVIDIA drivers, PyTorch, CUDA, vLLM and SGLang preinstalled, so open models can serve the whole office from one box.
- Suits: a firm with an IT lead or outside IT support and enough sensitive work, such as PI medical records or discovery, to keep it in-house.
One honest limit: on-prem only changes where the data lives. Access control, backups and patching are still the firm's job, Computer 4 needs a 240V circuit planned before it arrives, and an open model is not a legal research database, so citation checks stay the same. Hardware ships in 4 to 6 weeks.
Computer 2 is the desk-sized option for a solo lawyer or a small firm, built on two GPUs instead of four.

It comes with 64 GB or 192 GB of GPU memory. Autonomous's own benchmark on a 27B open model shows about 77 tokens per second per user with 4 users at once and 52 per user with 24, which is faster than most people read.
It fits on a desk, but someone still has to set it up and keep it updated.
How AI Changes the Billable Hour and Firm Revenue
AI cuts hours on work billed by the hour, so the money question is pricing, not headcount. In a 2025 Harvard Law School Center on the Legal Profession study, chief operating officers and partners at ten AmLaw100 firms estimated the billable hour at 80% or more of their fee arrangements.
One firm said a complaint-response system cut associate time from 16 hours to 3 to 4 minutes. Those firms did not expect to need fewer lawyers.
- Rates: they expected the value to show up in rates over time, not as separate AI charges to clients.
- Fixed fees: they were building more fixed-fee work, though none felt their pricing skills were mature yet.
- Pilots: many failed to deliver and were dropped, and the study's sample is small.

The ethics rules point the same way. ABA Formal Opinion 512 (July 29, 2024) says lawyers who bill hourly must bill the time they actually spent and in most cases may not charge clients for time spent learning a tool. Charging the same flat fee when AI makes the work much faster may be unreasonable under Rule 1.5.
Where the saved time goes is a choice the partners make. Most firms in the Harvard study expected it to go into better service, and about half would consider taking on lower-margin work their clients now give to smaller firms.
A practical check: before a pilot, write down which matters you will price as flat fees if the tool works, so the gain does not just shrink the invoice.
How to Roll Out AI for Law Firms of Each Size
A firm-wide rollout has five steps in the same order at every size: a written policy, a narrow pilot, training, a review rule and a measure of results. Size only changes who does each step and how formal it is.
- Policy: list approved tools, the data tiers each may touch, and who approves new ones.
- Pilot: pick one or two tasks from the low- or medium-sensitivity rows for 30 to 60 days.
- Training: show staff the approved tools, what not to paste in, and how to check output.
- Review: name who signs off on each output type before it goes to a client or court.
- Measure: track hours per task before and after, error rate found in review, and staff use.
| Step | Solo | Small firm (2-20) | Mid-size (20-100) |
|---|---|---|---|
| Policy | One page: approved tools, what never goes in | Two to three pages, signed by everyone | Partner and IT own it, with tool approval |
| Pilot | Time entries or regulatory summaries | One group, one task, such as billing narratives | Two groups plus a control group |
| Training | Vendor tutorials plus an AI ethics CLE | One in-house session, shared prompt library | Role-based: lawyers, paralegals, billing staff |
| Review | You check everything yourself | Supervising lawyer signs client-facing output | Rules per output type in matter workflows |
| Measure | Hours saved per week | Hours per task, write-offs on pilot matters | Hours, realization, fixed-fee margins by group |
Solo lawyers, often working from a home office, carry the whole policy alone. The one-page version matters most there: it is what you show a client who asks how their files are handled.

AI for small law firms works best when the pilot is boring. Billing narratives, regulatory summaries and marketing copy sit in the low and medium rows of the sensitivity table, show savings in weeks and carry little risk if the output is wrong.
Intake, deposition summaries and contract review come later, and only with a tool cleared for highly sensitive data.
Mid-size firms need the control group most. Partners who doubt the time savings will trust a side-by-side count of hours on matched matters more than a vendor demo. Newer AI agents that run multi-step tasks belong in a second pilot, after the first one has a measured result.
Risks and Guardrails of AI for Law Firms
The three risks that matter most are made-up citations, weak supervision and leaked client data. Each has a plain guardrail.
- Hallucinated citations: A Bloomberg Law analysis quoted by Illinois Courts counted over 280 court filings with AI-invented citations since 2023, as of January 2026. Pull and read every cited case before filing, and keep a short log of how each one was checked.
- Supervision: Opinion 512 applies Model Rules 5.1 and 5.3 to AI, so partners answer for how associates and staff use it. The written review rule from the rollout plan is how a firm shows that supervision exists.
- Confidentiality: Opinion 512 says informed client consent is needed before client information goes into a self-learning tool whose output could disclose it, and that general boilerplate in an engagement letter is not enough. The broader steps for AI privacy and security (retention settings, access control, vendor terms) apply here too.

The ABA opinion is guidance, not a binding rule, and state bars have issued their own opinions, so check your state bar's current guidance before you finalize the policy.
FAQs
Which AI are law firms using?
Law firms mostly use legal research platforms with AI, such as Lexis+ AI, CoCounsel and Westlaw, plus legal-specific tools like Harvey and Legora at larger firms. Many also use AI inside practice-management software such as Clio, and general chatbots like ChatGPT or Claude for drafting non-confidential text.
How are law firms currently using AI?
Law firms currently use AI for legal research, document review, first drafts, client intake and billing narratives. In the ABA's 2024 survey, 30% of responding lawyers used AI, with use highest at firms of 100 or more lawyers. A lawyer reviews the output before it reaches a client or court.
Is AI allowed in law firms?
Yes, AI is allowed in law firms as long as lawyers meet their existing duties of competence, confidentiality, supervision and reasonable fees. ABA Formal Opinion 512 applies those Model Rules to generative AI. State bars add their own guidance, so check the rules where you practice.
Is Claude or ChatGPT better for lawyers?
Neither Claude nor ChatGPT is built for legal research, and both can invent citations, so the better choice depends on the plan's data terms and your workflow. Check whether a plan trains on your prompts and how long it keeps them, and keep citation work in a legal research tool.
Is there a legal AI for lawyers?
Yes, several AI tools are built for lawyers, including Lexis+ AI, Thomson Reuters CoCounsel, Bloomberg Law's AI features, Harvey and Legora. They answer from legal sources and link the authority they cite, which makes checking easier than with a general chatbot. A lawyer still reviews the output.
Will AI reduce billable hours at law firms?
AI reduces the hours spent on tasks like document review and first drafts, so firms that bill only by the hour can lose revenue unless they adjust pricing. Firms in a 2025 Harvard study expected the value to show up in rates over time, not as separate AI charges, and planned more fixed-fee work rather than fewer lawyers.
What is the best way to start using AI at a small law firm?
Start with a one-page policy and a 30 to 60 day pilot on a low-risk task such as billing narratives or regulatory summaries. Measure hours before and after, name who reviews each output, and move to sensitive work like intake only with a tool cleared for that data.
Conclusion
AI for law firms works best on high-volume text work that a lawyer can check quickly: review, summaries, first drafts and billing narratives. Sort each task by how sensitive its data is, pick where the tool runs to match, and roll out with a policy, one pilot and a measured result before going wider.
The firms that get the most from it decide in advance how the saved hours will be priced.
References
- ABA Journal: ABA tech report finds that AI adoption is growing, but some are hesitant (March 2025)
- American Bar Association: ABA issues first ethics guidance on a lawyer's use of AI tools (Formal Opinion 512, July 2024)
- Harvard Law School Center on the Legal Profession: The impact of artificial intelligence on law firms' business models (February 2025)
- Illinois Courts: Paste in Haste, AI hallucinations in court filings and the ARDC's guide to implementing AI (January 2026)
- Harvard Journal of Law and Technology: AI, confidentiality, and the stratified legal profession
- Thomson Reuters Institute: ABA Formal Opinion 512 and billing for generative AI work

