Jiberish

AI Consulting

Generative AI for small law firms: what to adopt and which policies come first

Where generative AI fits in a small or solo firm, what ABA Opinion 512 and state guidance expect first, and a policy and 90-day plan to put in place.

Kyle Del FranciaUpdated 13 min read
In this piece

For a small or solo firm, the first AI decision is not which product to buy. It is which tool may receive client information, under which account terms, and who checks the output before it reaches a client, opposing counsel, or a court. Once that is settled, the useful work is fairly clear: intake summaries, first drafts from the firm's own templates, summaries of documents the firm holds, and client updates drafted from matter notes, each reviewed by a lawyer. Research help is worth having too, provided every citation is read in a primary source before anyone relies on it.

Many firms already use these tools without that groundwork. 8am's 2026 Legal Industry Report, published March 5, 2026 from a fall 2025 survey of more than 1,300 legal professionals, reports that 69% use general-purpose AI tools for work and that fewer than half of firms provide training on responsible use. It is a vendor's survey, but it supports the order we recommend: approved tools and a policy, then training, then automation.

Sources were checked on October 31, 2026. This article is general information, not legal or ethics advice. Your own state's rules govern, and they can differ from the ABA Model Rules discussed here.

What the rules expect before client information goes in

The main national reference is ABA Formal Opinion 512, issued July 29, 2024. ABA opinions are not binding, but state bars and courts draw on them, and this one covers six duties.

Competence (Rule 1.1). Lawyers need a reasonable understanding of a tool's capabilities and limitations, not expertise, kept current as tools change. Relying on output without an appropriate degree of independent verification can breach the duty. The opinion scales checking to the tool and task: a lawyer who tested a tool's contract summaries against a sample reviewed by hand may not need to reread every contract. The lawyer remains fully responsible either way.

Confidentiality (Rules 1.6, 1.9(c), and 1.18(b)). Before entering information relating to a representation, including a prospective client's, the lawyer must evaluate whether people outside the firm, or people inside it who should not have it, could access it. For tools that learn from what users enter, the opinion requires the client's informed consent first, and general boilerplate in an engagement letter is not enough. At minimum, lawyers should read the tool's terms of use and privacy policy, or rely on someone who has.

Communication (Rule 1.4). Not every use must be disclosed. Disclosure is required when the client asks, when engagement terms require it, when consent is needed for confidentiality, when the tool bears on the fee, and when its output will influence a significant decision in the matter.

Candor (Rules 3.1, 3.3, and 8.4(c)). Output, including analysis and citations, must be reviewed before it goes to a court. Some courts require disclosure of AI use, so local rules and standing orders need checking.

Supervision (Rules 5.1 and 5.3). Managing lawyers must set clear policies on permissible use, and supervisors must see that lawyers and staff are trained. Vendor diligence applies to AI providers: enforceable confidentiality, breach notice, what the provider retains, including after you stop using it, and its security.

Fees (Rule 1.5). Hourly billing covers actual time, including prompting and review, not time saved. An unchanged flat fee may become unreasonable if AI makes the work much faster. A grammar tool built into a word processor is overhead; a per-use charge for one client's matter can ordinarily be billed at actual cost. Lawyers may not bill clients for learning a tool they will use regularly.

Utah and other states

Utah lawyers practice under the Utah Rules of Professional Conduct, adopted by the Utah Supreme Court, which the Utah Constitution charges with governing the practice of law. As of October 31, 2026, the Utah State Bar's ethics opinions page lists no opinion on generative AI; its most recent opinions, from August 2025, concern other subjects. Bar material such as a January 15, 2026 post on AI tools is informal. In September 2025 the Supreme Court's civil procedure advisory committee formed an AI issues subcommittee to consider requiring attorneys to certify that cited cases were verified. We found no adopted rule on lawyers' AI use in filings.

The clearest Utah authority is a sanctions decision. In Garner v. Kadince, 2025 UT App 80, filed May 22, 2025, the Court of Appeals addressed a petition containing fabricated authority from ChatGPT. The opinion records that an unlicensed law clerk prepared it, the signing attorney did not check it before filing, and the firm had no AI policy. The court ordered the signing attorney to pay the other side's fees for responding and to donate $1,000 to "and Justice for all," and counsel to refund the client's fees for the petition. It relied in part on Mata v. Avianca, where the Southern District of New York imposed a $5,000 penalty on June 22, 2023 while observing that there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance." In both cases the failure was an unchecked filing. Garner also notes a local difference: Utah's Rule 3.3 reaches false statements made knowingly or recklessly.

Other states show where guidance is heading. Florida Bar Ethics Opinion 24-1, issued January 19, 2024, recommends informed consent before using a third-party AI program with confidential information and requires a website chatbot to disclose that it is an AI program, not a lawyer or employee. The State Bar of California approved updated guidance on May 14, 2026. Its Practical Guidance says lawyers must not let agentic systems make substantive legal determinations, communicate legal advice, or file pleadings without meaningful lawyer supervision and review, and California has proposed AI comments to its conduct rules.

Utah's consumer AI disclosure statute is a separate question. Section 13-77-103 requires up-front disclosure when a person in a regulated occupation uses generative AI in a high-risk interaction, and the statute's examples include legal advice. But section 13-77-101 defines a regulated occupation as one regulated by the Department of Commerce that requires a state license or certification. Utah lawyers are admitted and disciplined under the Supreme Court's rules, so on our reading that requirement does not reach them. Another subsection requires a supplier in a consumer transaction to disclose AI use when a person clearly asks; whether that reaches a law firm's chatbot is a question for Utah counsel. We would label a client-facing chatbot as AI from its first message either way.

Check the plan, not the brand

Data terms differ by plan and product, so the confidentiality review has to be done for the account the firm will actually use. As each provider describes it on October 31, 2026:

Provider and planTraining on your contentAlso check
Claude commercial products, including Team, Enterprise, and the APINot used by default (Anthropic)Feedback a user submits can be used; Team and Enterprise owners can turn feedback off
Claude Free, Pro, and MaxSet by each user; if on, de-identified data may be kept up to five years (Anthropic)The firm does not control the setting
ChatGPT Business and EnterpriseNot used by default (OpenAI Academy)Chats are kept until deleted; Enterprise owners can set retention, minimum 90 days
OpenAI APINot used unless you opt in (OpenAI)Abuse-monitoring logs kept up to 30 days by default
Microsoft Copilot with a work accountPrompts, responses, and Microsoft Graph data not used to train foundation models (Microsoft)Can surface any firm data the user has permission to view, so file permissions matter

A training default is one input. Retention, vendor access, connectors to other services, and who inside the firm can see a shared project or folder matter as well. In our view, the first gap to close is staff using personal consumer accounts for client work, because the firm controls none of those settings.

Where AI helps, with a lawyer reviewing

  • Intake and conflict-check preparation. An enquiry becomes a structured summary and a list of every name in it, for the conflict search the firm already runs. The worked example below shows the review.
  • First drafts and summaries from the firm's own materials. A demand or engagement letter from the firm's template and the facts in the file, or a summary of a lease, contract, or transcript. Test summaries against documents a lawyer has already read before relying on them.
  • Research support. Issue lists, search terms, and outlines are reasonable starting points. Every authority is then pulled from a research service or the court's own site and read.
  • Client communication drafts. Status updates written from matter notes, edited and sent by the lawyer.
  • Knowledge management. A shared project for approved templates, checklists, and the firm's style guide, with client documents kept in matter-specific spaces so one client's information does not surface in another's work. Opinion 512 also suggests marking AI-generated material in the file.
  • Practice operations. Internal meeting notes, drafts of firm procedures, and scheduling messages that contain no legal advice.

What stays with the lawyer

Legal judgment stays human: advising, negotiating, deciding whether a conflict exists, and calculating deadlines. A lawyer reads anything filed or sent outside the firm in full first, and no citation is used until someone has read the source. Features that send messages or move files on their own stay off for client work until the firm decides in writing what they may do. California's 2026 guidance says a lawyer must not deploy an agentic system that transmits client information externally without appropriate safeguards and human review.

A one-page firm policy

  1. Approved tools, the plan for each, and required settings for training, feedback, retention, and connectors.
  2. What information may go into each tool, including prospective-client information.
  3. Prohibited uses: personal accounts for firm work, unreviewed filings or client messages, and automatic sending.
  4. The review standard: who reads each cited source, and who signs.
  5. Client communication: when the firm discloses use, engagement-letter language, and when to seek informed consent for a specific tool.
  6. Billing: actual time only, no charge for learning tools, and how tool costs are treated.
  7. Supervision: training for lawyers, staff, and contractors, plus vendor review.
  8. Court requirements: checking local rules and standing orders before filing.
  9. Incidents: how to report a confidentiality slip or a flawed output that went out.
  10. An owner and a review date.

A worked example: from intake email to conflict list

This example is illustrative and was not run through a model. The firm, people, and businesses are invented. Because a prospective client's information is protected, it would run only in a tool the policy approves for client information.

<task>
Prepare an intake summary and a conflict-search list from the
prospective client email below, for the responsible attorney.
Use only the email. Do not give legal advice, assess the merits,
or calculate any deadline.
</task>

<output>
1. Names for the conflict search, grouped as prospective client,
   adverse parties, and others mentioned, each with its source sentence.
2. Facts as the sender states them, attributed to the sender.
3. Dates, amounts, and documents, quoted as written.
4. The sender's questions, listed and unanswered.
5. Anything the attorney should see before any reply, including
   possible connections to the firm. Mark uncertain items unconfirmed.
</output>

<email>
I found your firm online. My landlord, Larkspur Commons LLC, says I
owe $14,200 in back rent and common-area charges for my bakery space,
and I got a notice dated October 23 saying pay or move out. I think
the common-area charges are wrong. My business partner, Dana Ortiz,
keeps our books but is moving to Boise next month. The property
manager is Jeff Lund. I think his wife works at your firm? Do I have
to pay this? I'd also like to go after them for the broken HVAC that
closed us for a week in August.
Sam Reyes, Reyes Baking Co.
</email>

The reviewing attorney compares each item with the email:

ItemAcceptable outputNeeds correction
Conflict namesSam Reyes and Reyes Baking Co.; Larkspur Commons LLC as adverse; Dana Ortiz and Jeff Lund as others mentionedLund or Ortiz missing, or Ortiz listed as a client or adverse party
Firm connectionSender believes Lund's wife works at the firm; unconfirmedDropped, or resolved as "no conflict"
AmountLandlord claims $14,200; sender disputes the common-area charges"Client owes $14,200"
NoticeNotice dated October 23, as the sender describes itAny computed deadline or number of days
Question"Do I have to pay this?" listed, unansweredAny answer, or a reply that gives one
HVAC closurePossible separate issue about an August outageA named legal claim or a view of its strength

The list is an input to the firm's own conflict search, not a conflict decision. Any acknowledgment the prospective client receives should come from a pre-approved template sent by staff. Before using live enquiries, run several closed or fictional ones and record each correction; the Claude starter workflow shows how to keep answer keys for that check.

A first 90 days

In the first month, we would find out what lawyers and staff already use, including personal accounts, then choose one tool on a business plan, configure it, and adopt the policy. In the second, train everyone and run two workflows, such as intake summaries and template drafting, on closed or fictional matters, recording corrections and review time. In the third, decide from those records which workflows move to live matters, align engagement letters and billing practice, and set the policy's first review date. Automating intake comes after the manual version holds up; the lead handoff example covers the exceptions to plan for.

How Jiberish Labs helps

Jaxon Munns, a Jiberish Labs co-founder, is also the founding attorney of Munns Legal, licensed in Utah and Idaho. Jiberish Labs engagements cover tools, workflows, and training; professional-conduct decisions stay with your firm and its ethics counsel.

AI Advisory & Training can train attorneys and staff on the tools your firm approves, using exercises like the intake example; it is tailored and quoted after discussing team size, needs, and format. Custom AI Systems & Agents covers intake workflows, such as routing a website enquiry into a review queue with conflict-search preparation and no automatic replies; focused projects are quoted after discovery, and the Embedded AI Partner is $37,500 paid upfront for three months. For a single decision, such as choosing a tool, an advisory session is $200 for a 60-minute working session.

See AI training and workflows for small law firms for how this work is scoped.

Put this into practice

Bring the decision you’re working through.

A free discovery session is a chance to review your goal, the tools or website you have, and the help you need. We’ll identify a sensible next step and discuss scope before any project commitment.

Custom AI Systems & Agents →AI Advisory & Training →Book a free call