AI for Law Firms: An Ethics-First Guide
AI for law firms without the ethics risk: what small firms can safely automate, what ABA Opinion 512 and Texas Opinion 705 require, and a 90-day rollout plan.
Key takeaways
- AI for law firms is allowed and useful, but the ethics rules decide how. ABA Formal Opinion 512 and Texas Opinion 705 both say the same four things: stay competent with the tool, protect confidentiality, verify every output, and never bill for time the software saved.
- The safe starting point for a small firm is low-risk, high-volume work: intake, scheduling, first drafts of routine letters, record summaries and time-entry cleanup. Court filings and legal research come last, and only with a written verification step.
- Most of the risk is in the setup, not the model. A business-tier tool with signed terms, no training on your inputs, single sign-on and a one-page policy removes most of the exposure that sinks firms using personal ChatGPT accounts.
AI for law firms means using generative AI and automation for drafting, document review, summaries, research, intake and billing, under the same ethics rules that govern any other assistant. For a small firm, the right approach is to start with low-risk administrative work, use business-grade tools that do not train on client data, verify every output, and write it down in a short policy.
I am not a lawyer, and nothing here is legal advice. I am the technology side of the table: I run TMG, a Houston managed IT company that supports professional offices including law practices, and I founded Talos Automation, which builds AI intake and workflow automation, so I have a commercial interest in firms adopting this and you should weigh what follows accordingly. What I can offer is the operator view: what the ethics opinions actually require in practice, where the technology breaks, and what a sensible rollout looks like for a firm of 3 to 50 people that does not have a chief innovation officer.
Most of what ranks for this search is written by legal software vendors, and every one of them concludes that you should buy their product. This guide does not recommend a product. It gives you a framework for deciding what to automate, a checklist for vetting any tool, and the rules your state bar will hold you to. If you want the general version first, my operator's guide to AI automation strategy covers the same thinking for any small business.
What Can AI Actually Do for a Small Law Firm?
AI can reliably draft, summarize, classify and extract. That covers a large share of a small firm's non-billable and low-judgment work: first drafts of routine correspondence, summaries of medical records and depositions, intake questionnaires, document sorting for review, deadline extraction and time-entry cleanup. It cannot be trusted to find the law on its own or to exercise professional judgment.
The useful way to think about it is risk against volume. The table below is how I sort a firm's candidate tasks before anyone buys anything.
The conclusion of that table: start with intake, routine drafting and time entries, move to summaries and contract review once your verification habits exist, and leave legal research and anything filed with a court for last. The high-risk tasks are where every public AI disaster in the legal profession has happened so far.
What Do ABA Formal Opinion 512 and Texas Opinion 705 Require?
Both opinions apply rules lawyers already follow to generative AI, and both land in the same place: understand the tool, protect client confidences, verify the output, communicate with the client, supervise staff, and bill fairly. Neither bans AI, and neither creates a new rule. They are advisory opinions, but they tell you how a grievance committee will read the existing rules.
The American Bar Association issued ABA Formal Opinion 512 on generative AI tools on July 29, 2024. It covers competence, confidentiality, client communication, meritorious claims, candor toward the tribunal, supervision and fees. Two lines matter most for a small firm: informed consent is needed before client information goes into a self-learning tool, and general boilerplate in an engagement letter does not count as informed consent.
In Texas, the Professional Ethics Committee issued Opinion 705 on generative AI in the practice of law in February 2025. It works through the Texas Disciplinary Rules of Professional Conduct:
- Rule 1.01, competence: Comment 8 already asks lawyers to keep up with the benefits and risks of relevant technology. Opinion 705 reads that as a duty to understand a generative AI tool before using it on client work.
- Rule 1.05, confidentiality: the lawyer should be reasonably satisfied that a tool will not reveal confidential information, which means reading the terms of service, understanding the data security and training staff.
- Rule 5.03, nonlawyer assistants: applied by analogy. You supervise the software and the people using it the way you supervise a paralegal.
- Rules 3.01, 3.03 and 3.04: unverified output that reaches a court can become a frivolous claim or a false statement to the tribunal.
- Fees: bill actual time spent, including checking the output, but not hourly for time the tool saved.
If you practice outside Texas, check your own state bar; most have now published similar guidance, and the ABA opinion is the common reference point. This is the same structure I recommend for any regulated business in my AI governance framework for small business, just with the bar as the regulator.
Can Lawyers Use ChatGPT for Client Work?
Lawyers can use ChatGPT, Claude, Copilot or Gemini for client work only on an account and configuration they have vetted: a business or enterprise tier with signed terms, inputs excluded from model training, retention you control, and access tied to the firm's identity system. A paralegal's personal free account fails every one of those tests, and it is the most common setup I find.
The distinction that matters is not which model you use but which contract you are under. Consumer tiers of most AI products may keep conversations and may use them to improve the model unless someone finds and changes the setting. Business tiers typically state that they do not train on your inputs and give an administrator control over retention, sharing and who has access. Opinion 705 asks you to read those terms, not to assume them.
Before any tool touches client information, I want written answers to these questions:
- Training: are our inputs and outputs used to train or improve any model, and is that off by default or by setting?
- Retention: how long are prompts, files and outputs stored, and can we set it to the minimum or delete on demand?
- Access: does it support single sign-on and multi-factor authentication, and can we remove a departing employee in one place?
- Location and subprocessors: where is data stored, and which third parties process it?
- Security evidence: is there a current independent audit report, such as SOC 2 Type II, that we can read?
- Exit: can we export our data and have it deleted when we leave?
If you suspect staff are already pasting matter details into personal accounts, you are not alone, and banning it rarely works. My guide to shadow AI and how small businesses fix it covers the approach: give people a sanctioned tool that is better than the one they are sneaking, then close the side doors. For the choice between the two big general assistants, see my Claude vs. ChatGPT comparison for business.
How Do You Keep AI From Inventing Case Law?
You assume it will, and you build a verification step that catches it. General chat models predict plausible text, so a citation that looks right is not evidence the case exists. Every case, statute, quotation and pin cite in AI-assisted work gets checked against a primary source or a citator before it goes to a client, opposing counsel or a court.
The case everyone cites is Mata v. Avianca. In June 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned two lawyers and their firm 5,000 dollars after a brief cited cases that ChatGPT had invented, as LawSites reported on the sanctions order. The court also ordered them to notify their client and each judge falsely named as an author of the fake opinions. The judge was explicit that using AI was not the problem; failing to check, and then not owning the mistake, was.
Texas courts noticed. Days after the Mata filings became public, Judge Brantley Starr of the Northern District of Texas issued a standing order requiring lawyers appearing before him to certify either that no generative AI drafted their filing or that a human checked any AI-drafted language against traditional sources, as Hunton's privacy and cybersecurity blog summarized. Individual judges' standing orders change, so check the current orders of every court you file in.
The verification protocol I recommend is deliberately boring:
- Separate drafting from research. A general assistant can outline an argument; only a citator-backed research platform or a human finds the authority.
- Pull every authority. Open each cited case or statute in a primary source and confirm it exists, says what the draft claims, and is still good law.
- Check every quotation word for word. Models paraphrase and then put the paraphrase in quotation marks.
- Record who verified it. A one-line note in the file with the name and date. That is your evidence of supervision if anyone ever asks.
- Make the responsible lawyer sign off. Their name is on the filing; the tool's is not.
How Should a Law Firm Bill for AI-Assisted Work?
Bill the time a lawyer or paralegal actually spends, including prompting, reviewing and correcting the output, and nothing for the time the tool saved. Do not bill clients for learning a tool. Per-use AI charges can be passed through only if the client agreed in advance, and only at what the firm actually paid.
This is where AI quietly changes a small firm's economics. If a records summary that took four hours now takes one hour of review, the hourly bill drops by three hours. Opinion 705 is clear that efficiency gains under hourly billing belong to the client. The firms that come out ahead are the ones that move suitable work to flat or phased fees, where the efficiency becomes margin, while keeping the fee reasonable under the rules.
- Hourly matters: record actual review time; never reconstruct what the task would have taken by hand.
- Flat-fee matters: price on value and risk, and keep the fee reasonable; AI lowers your cost, not the reasonableness test.
- Subscriptions: treat AI tools as overhead unless the engagement letter says otherwise.
- Time entries: AI can tidy narratives and spot missed entries. It must never invent or inflate them.
Which AI Tools Should a Small Law Firm Evaluate?
Evaluate by category, not by brand. A small firm usually needs at most four kinds of tool: a secure general assistant for drafting and summarizing, the AI features already inside its practice management and document systems, a citator-backed research product if it litigates, and an intake or phone layer. Many firms already pay for the second category and have not switched it on.
Start with what you own. If the firm runs on Microsoft 365, the security baseline matters more than any AI feature: permissions on SharePoint and OneDrive determine what an assistant like Copilot can surface, and over-shared folders become over-shared answers. My Microsoft 365 security service exists largely because of that problem.
- General assistant (business tier): drafting, summarizing, rewriting, internal knowledge questions. Vet with the six questions above.
- Practice management and document AI: features inside the systems that already hold your matters. Lowest change-management cost, but read what each feature sends where.
- Research platforms with citators: answers tied to verifiable authority. Still verify, but the failure rate is lower than a general chat model.
- Intake and phones: answering, screening and scheduling new inquiries around the clock.
Whatever you buy, run a two-week pilot on closed matters or synthetic data before it sees a live client file, and write down what good output looks like before the pilot starts, so the decision is about evidence rather than a demo.
How Do You Roll Out AI in a Law Firm in 90 Days?
A small firm can go from personal ChatGPT accounts to a governed, useful setup in about 90 days: two weeks to inventory and write a policy, a month piloting one or two low-risk workflows, and the remaining time training staff, updating the engagement letter and measuring results before expanding.
- Days 1 to 14, inventory and policy. Ask everyone which AI tools they use, without blame. Write a one-page policy: approved tools, what data may go in, the verification rule and who to ask. My template for writing a small business AI policy is a workable starting point; have a lawyer adapt it to the disciplinary rules.
- Days 15 to 30, secure the platform. Move to a business tier with single sign-on, multi-factor authentication, training off and retention set. Clean up file permissions before any assistant can search them.
- Days 31 to 60, pilot two workflows. Pick from the low-risk rows of the table: intake plus routine letters is a common pair. Measure minutes saved and error rates on a sample.
- Days 61 to 75, client communication. Add a plain-language AI paragraph to the engagement letter and decide when you will seek specific informed consent, following ABA 512 and Opinion 705.
- Days 76 to 90, train and decide. Train every lawyer and staff member on the policy and the verification protocol, then expand, adjust or stop based on the pilot numbers.
The order matters more than the speed. Firms that buy a research tool first and write the policy later are the ones that end up explaining themselves. If you want the broader sequencing logic, it is the same as in what your business should automate first.
Where Does AI Help Law Firm Intake and Phones?
Intake is usually the highest-return place to start because missed calls are lost matters, and the work is structured: capture the caller's facts, screen for practice area and conflicts, and book a consultation. An AI receptionist can do that after hours and during court days, then hand a clean summary to a human who decides whether to take the matter.
Two cautions specific to law. First, an intake system must never give legal advice or create the impression that the firm has accepted the matter; the script and disclaimers matter as much as the voice quality. Second, a prospective client's information is protected even if you never take the case, so intake data needs the same confidentiality controls as matter data. I wrote up the general buyer's view in my guide to AI receptionists for small business. For the legal-specific build, my agency's guide to AI legal intake goes deeper, with the standing disclosure that Talos is my company.
None of this works on a shaky foundation. AI tools inherit the firm's identity system, file permissions, backups and email security, and an assistant can only expose what your permissions already let it reach. If you are a Houston firm, the IT checklist my team at TMG uses for IT support for Houston law firms covers that groundwork.
The Bottom Line
AI for law firms is not a question of whether but of how. The ethics opinions already answer the how: understand the tool, keep client information out of anything you have not vetted, verify every output, tell clients what they need to know, supervise the people and the software, and bill only for real work. A small firm that starts with intake, routine drafting and time entries, on a business-grade platform with a one-page policy and a verification habit, gets most of the benefit with very little of the risk. The firms that get hurt are the ones that skip the setup and go straight to briefs.
If you want a second opinion on your firm's setup from the technology side, book a free strategy call and bring the one workflow your staff complain about most. Pair it with your own ethics counsel for the rules, and you will have both halves covered.
Frequently Asked Questions
How are law firms using AI today?
Small and mid-size law firms mostly use AI for first drafts of letters and routine documents, summarizing depositions and records, document review, legal research inside citator-backed platforms, client intake and scheduling, and time-entry cleanup. The lawyer still reviews and owns every output, and anything filed with a court is verified against a primary source.
Is it ethical for lawyers to use AI?
Yes, if the lawyer stays competent with the tool, protects client confidentiality, verifies every output and bills fairly. ABA Formal Opinion 512 (July 2024) and Texas Professional Ethics Committee Opinion 705 (February 2025) both apply existing rules to generative AI rather than banning it. Neither opinion lets a lawyer hand professional judgment to the software.
Can lawyers put client information into ChatGPT?
Not into a consumer account that may store inputs or use them to train models, unless the client gives informed consent. Texas Opinion 705 says a lawyer should be reasonably satisfied a tool will not reveal confidential information. A business tier with signed terms, no training on inputs, retention controls and single sign-on is the minimum for client work.
Do lawyers have to tell clients they use AI?
Often, yes. ABA Formal Opinion 512 says informed consent is needed before putting client information into self-learning tools and that boilerplate in an engagement letter is not enough. Texas Opinion 705 tells Texas lawyers to run the same analysis. A plain-language AI paragraph in the engagement letter plus a conversation covers most firms.
Can a law firm bill hourly for time saved by AI?
No. Texas Opinion 705 says a lawyer may bill the actual time spent using a tool, including reviewing and correcting its output, but not hourly for time the tool saved. ABA Formal Opinion 512 adds that a lawyer cannot bill a client for learning a tool. Flat fees must still be reasonable under the fee rules.
What is the biggest risk of AI for law firms?
Filing something the AI invented. In Mata v. Avianca, a federal court in New York sanctioned two lawyers and their firm 5,000 dollars in June 2023 over fake cases produced by ChatGPT. The fix is procedural: verify every citation and quotation in a primary source before anything leaves the firm.