Most law firms already use generative AI, whether or not the partners approved it. The tools also changed. In 2024 the question was a chatbot that drafts text. In 2026 the question is AI agents that can read a mailbox, fill in a form or prepare a filing with little human input. The ethics rules changed very little. What changed is how carefully a firm must apply them.
Courts are also less patient. A public database kept by researcher Damien Charlotin listed more than 2,000 court decisions worldwide that deal with AI-fabricated content by October 2026. In 2023 the best-known case was Mata v. Avianca in New York, where two lawyers and their firm were fined $5,000 for fake citations. The penalties since then are much larger.
Below are seven AI oversight mistakes we see in small and mid-sized law firms, and the fix for each one. The fixes follow ABA Formal Opinion 512 (July 2024), the ABA’s first formal ethics opinion on generative AI, and the state guidance that came after it. This article is general information, not legal advice. Check the rules in your own state.
Mistake 1: Treating AI as “set it and forget it”
The problem
A partner approves an AI subscription, or a vendor turns on an AI feature inside software the firm already uses. Nobody checks how it is used after that. Agents make this worse. An agent with access to a mailbox or a document system does not only suggest text. It can act, and it can act on bad instructions hidden in an email or a file.
The fix
- Keep a list of every AI tool, including the AI features built into your practice management, document and email software.
- Review it every quarter: which tools, who uses them, and for what kind of work.
- Give agents the least access they need. Start with read-only access. Require a person to approve anything that leaves the firm: an email, a filing or a payment.
- Turn on logs so you can see what a tool or agent did, and when.
Name one partner as the AI oversight lead. That person reads the quarterly review and decides what to change. A few hours a month is usually enough.
Mistake 2: No clear line between AI work and attorney work
The problem
When associates and staff use AI without clear rules about review, AI output can go straight into a brief or a client email. Opinion 512 makes clear that the supervision duties in Model Rules 5.1 and 5.3 apply to AI use. The State Bar of California went further in its 2026 guidance, which also covers agentic AI: a lawyer’s professional judgment cannot be delegated to AI, and a lawyer must review every AI output, including citations to authority, before it goes to a court.
The fix
Use a simple three-zone rule:
- Green: internal summaries, research starting points and draft outlines. Light review.
- Yellow: client emails, memos and contract analysis. An attorney reviews and approves before it is sent.
- Red: court filings, legal advice and privileged communications. A senior attorney reviews, and a lawyer reads every cited authority in full.
Put the zones in your written AI policy, train everyone on them, and treat a breach of the red zone as a performance issue.
Mistake 3: Choosing AI tools without checking data security
The problem
Firms often pick AI tools on features and price. But Model Rule 1.6 still applies to everything a lawyer types into a prompt. Opinion 512 says that before a lawyer puts information about a representation into a self-learning AI tool, the lawyer needs the client’s informed consent, and boilerplate in an engagement letter is not enough. Free and personal accounts are the biggest gap, because the firm has no contract and no control over them.
The fix
Before you approve an AI tool, get these answers from the vendor in writing:
- A SOC 2 Type II report from the last 12 months
- A contract term that says the vendor does not train its models on your data
- How long prompts and files are kept, and how you delete them
- Where the data is processed and stored, and which vendor staff can see it
- Encryption in transit and at rest
- Single sign-on, multi-factor authentication and admin controls
- Breach notification terms and timelines
Then block personal AI accounts on firm devices and give staff an approved tool instead. Have IT or your managed service provider sign off on each new tool. The review takes a week or two. Cleaning up after a client data incident takes much longer.
Mistake 4: Too little training on what AI gets wrong
The problem
Staff get an AI tool but no training on its limits. That is how fake citations end up in court. Comment 8 to Model Rule 1.1 makes technology part of a lawyer’s duty of competence, and Opinion 512 says lawyers need a reasonable understanding of what the AI tools they use can and cannot do.
The fix
Train everyone who uses AI at least once a year, and again when you add a new tool. Cover:
- How AI fails: invented citations and quotes, out-of-date law and confident wrong answers
- The checks your firm requires before AI output is used
- Which client information may go into which tool
- How agents can be tricked by instructions hidden in emails, web pages and documents
- Your duties under Rules 1.1, 1.6, 3.3 and 5.3
Some states now require technology or cybersecurity CLE hours. Use accredited courses where you can, track who completed the training, and use real sanction cases as examples.
Mistake 5: Engagement letters and bills that ignore AI
The problem
Many firms use AI on client matters without telling clients, or bill the same hours for work that now takes half the time. Opinion 512 applies Rule 1.4 (communication) and Rule 1.5 (fees) to AI. A lawyer who bills by the hour can bill only for the time actually spent. A firm cannot bill a client for time spent learning a tool it will use on many matters. A general AI subscription is normally overhead, not a client expense. A pay-per-use tool used on one matter can be billed at actual cost. States differ on how much disclosure is needed. Florida Bar Ethics Opinion 24-1, for example, recommends informed consent before confidential information goes into a third-party AI tool.
The fix
Update your engagement letters, and talk to clients about them. Cover:
- Which kinds of AI tools may be used on the matter
- How the firm protects confidentiality and checks accuracy
- How AI use affects fees and expenses
- How the client can ask for limits on AI use in their matter
Where a tool learns from what you put into it, get specific informed consent for that matter. A general clause is not enough.
Mistake 6: No system for AI risk
The problem
Most firms deal with AI problems only after one happens. Insurance is a growing blind spot. In January 2026, Verisk’s ISO program released optional generative AI exclusions for commercial general liability policies. Professional liability policies differ from carrier to carrier, and a firm should not assume that its policy covers an AI-related claim.
The fix
- Review risk regularly: new tools, new agent access and new uses.
- Log incidents: AI errors, near-misses and client complaints, so you can see patterns.
- Watch your vendors: changes to terms, data use and security.
- Add AI to your incident response plan: what to do if client data goes into the wrong tool, or an agent takes a wrong action.
- Check your insurance at every renewal: ask your broker in writing whether your professional liability and cyber policies exclude or limit claims that come from AI use, such as a fabricated citation or data leaked through an AI tool.
Mistake 7: Forgetting the people
The problem
Firms focus on the tools and forget the people who use them. When the approved tool is slow or hard to use, staff go back to personal accounts. When nobody explains the rules, people either avoid AI or use it in secret. Both are risks.
The fix
- Start small: run a pilot with a few willing users before a firm-wide rollout.
- Make the approved tool the easy choice, so nobody needs a personal account.
- Talk about concerns openly, including workload and job worries.
- Share what works and what went wrong, without blame.
- Measure results: time saved, errors caught and client feedback.
Build an AI governance framework
These seven fixes work best as one program. Form a small AI group with a partner from each practice area, IT or security, and firm management. Meet monthly at first, then quarterly. Your framework should include:
- A written AI policy with the three review zones
- A list of approved tools, and rules for what agents may access and do
- A vendor review checklist
- Training and a record of who completed it
- Engagement letter and billing language
- A clear path to report AI problems
Most important, make one person accountable. AI oversight must not fall into the gap between IT and practice management. The goal is not zero AI risk. That is not possible. The goal is to manage the risk in a way you can explain to a client, a court or a bar investigator.
Ready to see where your firm stands? Get the law firm scorecard, or book a call with our team.
Sources
- ABA Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024)
- State Bar of California, Practical Guidance for the Use of Generative AI (2026)
- Florida Bar Ethics Opinion 24-1
- Damien Charlotin, AI Hallucination Cases database
- Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023)
- Shumaker, The New AI Coverage Fight: Exclusions, Endorsements, and Denied Claims