Before an associate at your firm is handed access to a general-purpose assistant, four questions need answers in writing: where client information may go and what the client is told, who verifies the output and against what standard, what the firm may bill for work an assistant touched, and which seats get access first. Those are partner questions. The person doing the typing cannot answer any of them, and none of them get easier by being deferred until after the rollout.
Most firms run the sequence backwards: buy licenses, circulate a policy, schedule training for whoever asks. By the time anyone sits in a room, half the associates have formed habits on a personal account.
The gate most firms skip
The 2026 ILTA Technology Survey covers 508 firms and more than 139,000 attorneys. It found 94% using or exploring generative AI, up from 80% a year earlier, with document drafting the top use case at 79%. A formal policy is in place at 78% of firms. Then the same survey found 35% of firms saying they lack the time or resources to configure the features and train people on them.
Four in five firms wrote the rule. One in three admits it has not resourced the part that makes the rule stick.
Firm size predicts the rest of the gap almost perfectly. An analysis of the same survey data by firm size found 72% of firms with more than 700 lawyers require AI training before granting access, against 11% of firms under 50 lawyers. On requiring human review of output the split is 79% to 31%, and on requiring output be cited or independently checked, 55% to 25%.
A firm of 50 to 500 lawyers sits in the middle of that spread, carrying large-firm client expectations on small-firm training infrastructure. The gate matters more than the curriculum, because AI enablement is something a firm builds rather than buys with a license.
| Partner question | The rule it answers to | What partners produce before access opens |
|---|---|---|
| Where may client information go? | Model Rules 1.6 and 1.4 | A two-column tool list and the client-consent language |
| Who verifies, against what? | Rules 1.1, 5.1 and 5.3 | A written review step with a named owner |
| What can we bill? | Rule 1.5 | A billing position and updated engagement-letter terms |
| Who gets access first? | Rules 5.1 and 5.3 | A sequenced rollout, not a firm-wide switch |
Where client information may go, and what the client is told
On July 29, 2024 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, its first formal ethics guidance on generative AI. It created no new rules. It mapped AI use onto six existing Model Rules: 1.1 competence, 1.4 communication, 1.5 fees, 1.6 confidentiality, and 5.1 and 5.3 supervision.
Two conclusions from the confidentiality analysis change how a firm sets up tool access. Informed client consent is needed before client information enters a self-learning tool, and general boilerplate in an engagement letter purporting to authorize AI use does not count as that consent. Contractually walled tools also sit differently from consumer accounts, so the terms a firm has read matter more than the brand on the login screen.
The deliverable from this hour is a one-page list, posted where people work, in two columns: tools inside the line and what may be entered into each, and tools outside it, usable only on hypotheticals. A business-tier account under reviewed terms belongs in column one. A personal ChatGPT or Claude account somebody signed up for last spring does not, whatever the privacy toggle says.
Write the list, skip the session, and you land in the gap our guide to the distance between an AI policy and safe practice describes: a rule nobody has rehearsed against a live tool is a document, not a control.
Who owns verification
Damien Charlotin’s AI Hallucination Cases database tracks decisions in which a court found a party relied on AI-fabricated content. On October 19, 2026 it listed 2,202 matters worldwide, 1,501 of them in the United States, and it excludes mere allegations. These are findings.
Opinion 512 created no external monitor for any of this. It left the duty where the Model Rules already put it, so the firm’s own supervision is the compliance mechanism. Rule 5.1 makes partners and managers responsible for measures giving reasonable assurance that lawyers conform to the rules, and 5.3 extends that to nonlawyer assistance.
So teach the failure mode rather than warning about it. In the room, ask an assistant a question with a confidently wrong answer available: a repealed provision, a threshold that moved, a case name that nearly exists. Then have the group check it against the authority. Lawyers who have watched a fabricated citation arrive in clean prose stop reading polish as a quality signal. Lawyers who were only warned do not.
Then write the standard down so it can be supervised. Every authority opened and read in full, every quotation checked against the source document, every cite verified in the research platform rather than in the assistant that produced it. The reviewing lawyer whose name goes on the filing owns that check, and it belongs in the matter checklist where a supervising partner can see whether it happened.
What the firm can bill
Opinion 512’s fee analysis is the part vendors leave out of the pitch, and partners need it before anyone starts counting savings. For hourly work, a lawyer bills the time spent: prompting, reviewing and verifying are billable, and the time the tool saved is not, because it was not spent. On learning, the opinion is blunter. A lawyer may not charge a client to learn how to use a tool the lawyer will use regularly for clients, with a narrow exception where a client asks for a specific tool the lawyer does not know. The ABA’s own analysis of fee reasonableness works through flat and contingent arrangements too.
For a firm billing by the hour, the gain shows up as capacity rather than revenue until something about the fee structure changes. That is a strategic conversation, not a training one, and it is the same arithmetic agencies run when the billable hour meets a faster draft.
Treat the published value estimates as expectations rather than results. Thomson Reuters’ Future of Professionals Report 2025 found legal professionals expecting to free up nearly 240 hours a year, worth about $19,000 each annually. The same research found organizations with an AI strategy roughly twice as likely to see AI-driven revenue growth as those adopting without one.
Who gets access, and in what order
Access for everyone at once is the default, and the one choice that guarantees the firm’s posture gets set by whoever on the floor is boldest.
A workable order puts partners and practice-group leaders first, since they have to review work the tool touched before they can supervise it. Then the seats with high document volume and a reviewer already in the loop: litigation support, transactional drafting, business services. Junior associates last, with whichever guardrail the firm picked.
Each group needs different material. Role-based AI training rather than one firm-wide session is the argument our role-based enablement playbook works through in full, and the reason a pre-season session for an accounting firm looks nothing like one built for its administrators.
What a partner session covers that an associate session does not
Nearly every AI enablement course on the market addresses the person who will type the prompt. A partner needs four things that curriculum skips.
Reading output from a tool you do not use. The signatures of unverified work are learnable: citations formatted perfectly but absent from the research platform, quotations that paraphrase, a procedural posture fitting the wrong jurisdiction, numbers no source document supports.
The client conversation. Rule 1.4 and the consent analysis in 512 mean someone has to answer the client who asks whether AI touched their matter. Draft that answer once, as a firm, rather than improvising it on a call.
The fee conversation. What the firm bills, what it absorbs, and what it tells a client about either. Partners who leave this open either over-bill or quietly stop recording verification time.
The supervision artifact. Rules 5.1 and 5.3 are satisfied by measures, not intentions, so the session ends with a review step written into the matter checklist and a named owner for it.
The market has noticed the split. AltaClaro runs a supervisory course built for law firm partners covering large language models, the ethical duties AI touches, firm policy and oversight. It carries 2 CLE credits and ends in a checklist the participant writes.
The junior associate question nobody will settle for you
Process-level work was how associates learned to be lawyers: first-pass research, document review, the fourth revision of a standard agreement. Tedious work, and the work that built the judgment partners rely on later. It is also the tier assistants handle fastest.
At the New York State Bar Association’s Law Practice Management Roundtable on April 23, 2026, managing partners set out three different answers. Bond, Schoeneck & King bars new lawyers from the tools while core lawyering competence forms. Barclay Damon requires first- through third-year associates to complete work without AI before submission, then permits AI for refinement. Tully Rinckey grants access but rebuilds development around it, putting new hires into a six-to-twelve-month mentorship with daily check-ins.
The expensive answer is no answer. Leave it open and the firm runs all three postures at once, sorted by temperament, with no way to tell a client which one applied to their matter.
What it costs against the catalog
Market prices, verified October 2026:
| Option | Format | Price |
|---|---|---|
| Clio Legal AI Fundamentals certification | Self-paced, about 2.5 hours | Free |
| Harvey Academy Level 1 | Self-paced video, about 3 hours | Free |
| GC AI prompting cohort | Live, 75 minutes | $225 |
| NBI 2026 Legal AI Masterclass | On-demand, 5h53m, 6 to 7.2 CLE credits | $429, or $99/month unlimited |
| Stanford Law, AI strategy for legal leaders | Self-paced plus two live sessions | $900 |
| Berkeley Law executive education | Nine modules, about 8 hours | $950, discounted to $655 to $855 |
| Harvard Law executive education | In person | About $7,800 |
| Delivered session on the firm’s own matters | Live, hands-on | $2,000 to $15,000 market range |
The free academies are good at what they do. They teach tool mechanics in an afternoon and settle none of the four partner questions, because no vendor can tell your firm where its confidentiality line sits or what it bills. The catalogs buy credit and a framework. A session on the firm’s own matters buys the four answers and the habit, and it is the only format here where people practice on material they will see again on Monday.
So buy the catalog course for credit and the session for capability. Accreditation depends on the provider being approved in your jurisdiction, and those rules are moving. New Jersey now requires one technology-related CLE credit per two-year cycle, effective January 1, 2027. It sits separate from the five ethics credits and names AI developments among the qualifying subjects. Our walkthrough of prompting on client matters without breaching privilege goes deeper on the confidentiality mechanics. Automating a workflow is a separate decision at a separate scale, roughly $25,000 to $150,000 at market, and it should follow fluency rather than stand in for it.
The first thirty days
- Draft the two-column tool list and have the firm’s general counsel or ethics partner review it. One page, posted, no appendices.
- Settle the billing position and update the engagement-letter language to match. The client-facing answer and the internal time-recording rule have to agree.
- Write the verification step into the matter checklist and name its owner. Rule 5.1 asks for measures; this is the measure.
- Run the partner and practice-group-leader session first. Three hours, the firm’s own documents, ending in each attendee’s oversight checklist.
- Open access one group at a time, with a graded exercise on real work product rather than a completion certificate. Completion data only tells you who clicked.
For an outside read on where your firm sits against those four questions before you schedule anything, a free AI-readiness call is the quickest way to get one.
Frequently asked questions
What should AI training for law firms cover?
Four partner decisions first, then skills. The decisions: which tools client information may enter and under what consent, who verifies output and against what standard, what the firm may bill, and which seats get access in which order. The skills follow: drafting with an assistant, verification against primary sources, and reviewing work the tool touched.
Does ABA Formal Opinion 512 require law firms to train their lawyers?
It requires measures, and training is the practical form they take. Opinion 512 applies Rules 5.1 and 5.3, which make managerial and supervisory lawyers responsible for reasonable assurance that lawyers and nonlawyer assistance conform to the rules. No external monitor was created, so the firm’s own policy, training and supervision are the compliance mechanism.
Can we bill clients for time spent using AI?
For hourly work, bill the time spent prompting, reviewing and verifying. Time the tool saved was never spent and cannot be billed. Opinion 512 also holds that a lawyer may not charge a client to learn a tool the lawyer will use regularly, with a narrow exception where the client asks for a tool the lawyer does not know.
Can client information go into ChatGPT or Claude?
Not without resolving consent and terms first. Opinion 512 requires informed client consent before client information enters a self-learning tool, and holds that boilerplate engagement-letter language authorizing AI use is not enough. Most firms land on business-tier accounts under reviewed terms for anything client-identifying, and keep personal accounts to hypotheticals.
Do partners need a different session from associates?
They need a different one. Associates learn to produce with an assistant; partners learn to review work an assistant touched, answer the client question, settle the fee question and own the Rule 5.1 duty. A combined session teaches production and skips oversight, which inverts the risk.
Does a hands-on AI workshop earn CLE credit?
Usually not on its own. Credit depends on the provider being accredited in your jurisdiction for that delivery format, and state rules differ. Satisfy credit through an accredited course such as NBI’s masterclass or AltaClaro’s partner program, and buy the working session separately for capability.
How much does AI training for a law firm cost?
Vendor academies from Clio and Harvey are free. Accredited CLE courses run about $200 to $950 per person, with in-person executive programs reaching $7,800. A delivered session on the firm’s own matters runs roughly $2,000 to $15,000 at market rates, and automating a workflow is a separate $25,000 to $150,000 decision.
How many people should be in the room?
Eight to fifteen, small enough that everyone works on their own matters and gets unstuck. Past that it becomes a demonstration, and recorded courses do those more cheaply. Partner sessions run smaller still, because the output is each participant’s own oversight checklist.
Key takeaways
- Four partner questions precede any rollout: where client information may go, who verifies, what the firm bills, and who gets access first.
- The 2026 ILTA survey found 78% of firms with a formal AI policy and 35% admitting they lack the time or resources to train people on the tools. The rule outpaced the habit.
- Only 11% of firms under 50 lawyers require training before granting access, against 72% of firms over 700. The access gate is the cheapest control available.
- Opinion 512 created no external monitor, which makes the firm’s own training and supervision the compliance mechanism under Rules 5.1 and 5.3.
- Hourly firms cannot bill the time AI saves or the time spent learning the tool, so the gain lives in capacity, fixed-fee work and matters previously declined.
- Teach verification by producing a fabricated citation in the room, not by warning about one.
Arthur Wandzel
