AI Training for Lawyers: What Works, What Fails and a Learning Path for Every Role

Thomson Reuters tells a story about AI training for lawyers, or rather its absence, at a firm that did everything right on paper. The partnership voted unanimously for an AI platform and announced it firmwide. Six months later, “Only a fraction of attorneys had logged into the AI tools more than once”. The lawyers’ complaint was specific: “no one had trained them on the new tools beyond a single 90-minute webinar”. A competitor delivered an AI-assisted preliminary assessment before the engagement letter was signed, and won the pitch. TR’s diagnosis: “a focus on procurement rather than capability building”.

The story is a composite; the numbers are not. In the 8am 2026 Legal Industry Report, 69% of legal professionals use general-purpose generative AI for work, and 54% say their firm has provided no training on responsible use and has no plans to. Training tracks use almost perfectly: Law360 found 80% of frequent AI users had been trained and 71% of non-users had not.

So the question about AI training for lawyers is not whether. Your colleagues already use ChatGPT, mostly on personal accounts. The question is which format changes how a lawyer works, and which produces an attendance list. The evidence follows, with a learning path for each kind of lawyer and an honest placement of my own course among the alternatives.

The gap: 69% use it, 54% have never been trained

Every adoption survey finds the same shape: individual use far ahead of firm governance, and training the weakest link. Law360’s 2025 survey put firm-provided training at roughly two-thirds of BigLaw attorneys, 40% at midsize firms and under 15% at small ones. Wolters Kluwer’s 2026 survey of 810 legal professionals found 39% naming inadequate training and resources as a barrier. In Texas, only 9% of lawyers are “extremely confident” evaluating AI-generated documents; 40% are “not at all” or “slightly” confident.

Two details matter for anyone choosing what to buy. Firm size: Nicole Black, who wrote the 8am report, says mid-sized firms are often the slowest “because they just don’t have the resources available to devote to tech adoption and change management”. And a Thomson Reuters figure from July 2026: 46% of lawyers “do not know enough about AI to answer basic questions posed by their clients about its potential benefits”. That is a competence problem before it is a productivity one.

Why the 90-minute webinar fails

The best evidence on format comes from firms that tried the lecture first and said so publicly.

Paul Weiss’s first generative AI training was a PowerPoint on ethics, confidentiality and prompt engineering. Gina Lynch, its chief knowledge and innovation officer, called it “ineffective”. Iris Skornicki, the firm’s director of AI innovation strategy, describes the replacement: a general education session and a second, hands-on prompting workshop. Lynch’s verdict: “That dynamic training model is critical to really having lawyers understand how the technology works.” Jackson Lewis’s Kristen Baylis: “Off-the-shelf training from a vendor is not going to be as effective as if it’s co-produced with internal folks.”

GC AI, which runs free prompting classes for in-house teams, sees departments that had “a product demo and no structured training” reverting to old workflows a year later. Its verdict is the most useful sentence on the subject: “Self-paced is convenient. Live is effective.”

Regulators agree. The European Commission’s Article 4 guidance warns that “Simply relying on the AI systems’ instructions for use or asking the staff to read them might be ineffective.” The German commentary in beck-aktuell is blunter: “Wer nur Inhalte konsumiert und sich nicht selbst mit praktischen Anwendungsfällen befasst, wird kein nachhaltiges KI-Verständnis aufbauen.”

Why does the lecture fail? Because the skill is not knowledge. Every lawyer who has sat through the ethics slides knows the model can fabricate a case; only one who has watched it happen on a document they care about knows what a fabrication looks like in their own practice area. Nobody learned to cross-examine from a slide deck either.

What works: hands-on, two sessions, your own documents, champions

Put the firm reports, the surveys and the regulators’ guidance side by side and six features recur.

The Wolters Kluwer panel on its 2026 survey said the same. Giulietta Lemmi: “Training should be grounded in real use cases and closely aligned to the workflow that people actually follow.” Eve Vlemincx, on why one-off events fail: “We’re treating training as if work is stable. It isn’t.”

If you are designing an internal programme, this prompt turns those features into an agenda.

Design a two-session generative AI training programme for our [employment / corporate / disputes] group of [N] lawyers in [jurisdiction].
Session 1 (60 minutes, live): how a language model produces text and why it fabricates; consumer versus business tiers for confidentiality; our AI policy in five rules; three live demonstrations of failure (a fabricated citation, a wrong-jurisdiction answer, an over-confident summary).
Session 2 (90 minutes, hands-on): each participant brings one real task on an anonymised document; four exercises on our approved tools; a verification exercise on a draft with two planted errors I will supply; each participant writes one prompt-library entry.
Add 15 minutes of pre-work, a timed challenge, a weekly 15-minute clinic for six weeks, and the attendance record for our training file. Output an agenda with timings and a materials list.

Plant the two errors yourself. If you ask the model to invent fake citations for the exercise it will oblige, and then you have to check that they are genuinely fake.

The regulatory pull: Comment 8, Opinion 512, Article 4, the SRA and the courts

For a decade “technology competence” was a comment in a model rule. In 2026 it has teeth.

United States. Comment 8 to Model Rule 1.1, technology competence, has been adopted by 40 states plus DC and Puerto Rico. ABA Formal Opinion 512 (29 July 2024) sets the AI-specific standard: lawyers “need not become GAI experts” but “must have a reasonable understanding of the capabilities and limitations” of the tools they use, and supervisory lawyers must ensure “subordinate lawyers and nonlawyers are trained”. Three states mandate technology CLE: Florida (3 hours per three-year cycle since 1 January 2017), North Carolina (1 hour a year since 2019) and California (at least 1 hour). The technology CLE requirements guide has the state table.

The courts. Judges now prescribe training as a sanction. Withers v. City of Aberdeen (N.D. Miss., June 2026) added an AI-ethics CLE to fines and disqualification for lawyers on both sides; Beus Gilbert v. BYU (D. Utah, 9 September 2026) ordered two AI-ethics CLE courses alongside a $3,000 fine.

United Kingdom. The SRA’s warning notice of 17 August 2026 ties AI use to Code paragraph 3.6, keeping “professional knowledge and skills… up to date”, and states that “Reliance on an output of AI would not be a suitable defence.” The other direction now exists too: the UK Jurisdiction Taskforce’s July 2026 Legal Statement warns that a professional “could also be liable for failing to use AI” where a reasonably careful peer would have used it.

European Union. Article 4 of the AI Act has required deployers, which includes every firm using ChatGPT, Copilot or DeepL, to support staff AI literacy since 2 February 2025. Article 99 provides no fine, but the Commission’s Q&A is explicit: no certificate or AI officer is required, an internal training record is the expected evidence, and national sanctions become “more likely if there is proof of an incident due to lack of appropriate training”. Austria’s ÖRAK goes further: “Die Nutzung von KI-Systemen ist nur zulässig, wenn eine KI-Kompetenz der Mitarbeiterinnen und Mitarbeiter besteht.” The Article 4 guide for law firms has the checklist.

Learning paths by role

One format; different starting points.

The sceptical partner

Richard Susskind, who calls lawyers the slowest adopters “of any profession except the clergy”, gives the cheapest recipe known to work: a general counsel who made the whole team use AI 30 minutes a day for six months, after which the team “was transformed”. Whole team, not a pilot group: pilots create “two classes of lawyer”.

Start where Suffolk Law’s Dyane O’Leary starts her students: “in a low-stakes, casual environment with topics familiar to them — cooking, travel, home repair”. Week one is that kind of question. Week two is a public judgement in your area, summarised and then attacked. Week three is a false-premise test (“Why did Justice Ginsburg dissent in Obergefell?” is Stanford’s; she did not), so you see a confident fabrication before a client does. Week four is a real, anonymised document. The explainer on how LLMs work gives you the vocabulary to supervise associates who are already ahead of you.

The associate

Your problem is the opposite: you use it daily and nobody has checked how. The LexisNexis Mentorship Gap study of nearly 900 UK lawyers found 72% naming deep legal reasoning as the biggest skills gap and 69% weak verification; only 2% believed AI strengthens learning. Thomson Reuters’ 2026 survey found 48% fearing the effect on independent judgement, and legal respondents expect the path to “trusted judgment” to lengthen by 1.7 years.

Harvey’s own guidance for firms is the answer: a first-year who spent 40 hours on a services-agreement markup “might now spend 5 hours reviewing an agent’s markup”, so firms must build “intentional friction”. Build it yourself.

I am a [first-year associate] training myself to review AI output critically. Set up an exercise on [reviewing a services agreement markup, clauses 8 to 12].
Step 1: I draft the markup by hand in 60 minutes, no AI. Do not help with this step.
Step 2: after I paste my markup, run our playbook review on the same clauses using <playbook>...</playbook>.
Step 3: produce a comparison table: issues I found only | issues you found only | both | neither (leave blank for my supervising partner).
Step 4: give me five written questions on why each of us missed what we missed.
Step 5: I write the final markup myself. Do not write it for me.
Tag any legal proposition you rely on [VERIFY]. Do not cite cases.

Do this once a fortnight on real, anonymised work and you will out-learn the programme your firm has not run. Sidley London now makes an AI seat mandatory in every training contract; you can build the equivalent yourself. The AI skills guide lists what firms hire for.

In-house counsel

You are the most enthusiastic users (87% of general counsel report GenAI use in their teams) and the least well served by generic training. Axiom’s 2026 survey found only 7% of legal teams past piloting and 83% unable to show whether last year’s AI spend paid off. The training gap and the ROI gap are the same gap.

Apply GC AI’s test to every course: can you take a real contract from this week’s queue and apply the method before the session ends? Then follow Axiom’s pilot recipe: eight to twelve weeks, one use case, measurement from day one. NDA triage against a playbook pays fastest.

The solo and small-firm lawyer

Under 15% of small-firm practitioners have been trained by their firm, because the firm is you. Clio’s 2026 data says the constraint is time: for 27% of solos and 33% of small firms, finding time is the biggest hurdle. And 47% of solos use consumer tools such as ChatGPT or Copilot, which train on conversations by default.

So the first hour of your training is a settings exercise: confirm which tier you are on and switch off training, then anonymise one real document with a key table kept offline. Then build a six-week plan in the tool you already pay for; the 30-day plan for solo and small firms is the longer version.

Build me a six-week plan of 30 minutes a day to become competent with [ChatGPT Business / Claude Team / Copilot] for [family / litigation / transactional] work in [jurisdiction]. Each day: one specific task using my own non-confidential or anonymised materials, the prompt pattern it teaches, a success criterion I can check, and one thing that could go wrong.
Week 1: low-stakes personal tasks. Week 2: summarising and extraction. Week 3: first drafts with a verification step. Week 4: review against a checklist I supply. Week 5: a reusable project. Week 6: one full matter workflow.
Never suggest a task that would require entering client-identifying information. End with a ten-question self-assessment.

The German-speaking lawyer

In the Bayerischer AnwaltVerband’s survey (558 participants, 88% from firms of one to ten lawyers), the most-used tools are DeepL (70.7%) and ChatGPT (69.2%). Both fall under Article 4, which beck-aktuell reads as covering “alle KI-Tools, wie zum Beispiel Microsoft CoPilot, DeepL, ChatGPT oder BeckChat”, whatever the firm’s size.

Your training has one module the Anglo-American courses skip: professional secrecy under § 203 StGB, § 43e BRAO and, in Austria, § 9 RAO. BRAK’s line is that public tools should receive only “abstrakte” prompts “die auch im Kontext keinerlei Rückschlüsse auf ein bestimmtes Mandat zulassen”, and that removing names and addresses is regularly not enough. That rule becomes concrete only once you have anonymised a Schriftsatz yourself and seen what still identifies the client. Martin Lorentz, a Fachanwalt für Arbeitsrecht in a five-lawyer firm, reports “Meine Zeitersparnis lag bei mindestens 50 Prozent” on an appeal brief; the anonymisation exercise is how you get there safely.

Courses compared, briefly

The market has split into four camps: vendor academies, university executive education, MOOCs and live cohort courses. Artificial Lawyer’s line on the first camp is worth remembering: a vendor certification “helps with client outreach, and reduces the burden on vendors to provide quite as much hands-on training”. The full comparison of AI courses for lawyers goes course by course; this is the short version, as of September 2026.

None of these is a bad purchase. But a three-hour video cannot watch you prompt, and a leadership course will not teach you to anonymise a document.

Measuring whether training worked

Attendance is not the measure. Thomson Reuters found in 2026 that 66% of professionals in organisations with a named AI strategy said AI met or exceeded expectations, against 22% without. Training is where strategy becomes behaviour, so measure behaviour.

1. Use, four weeks later. Who has logged in more than once? Ashurst measured its trial and reported 88% feeling “more prepared for the future” alongside “frequent hallucinations across all GenAI tools trialled”. Report both.
2. Errors caught. Count the fabricated or misgrounded citations found in verification each month. A rising count means the check is being done.
3. Prompt-library entries with a “last tested on” date. Model updates silently change what a prompt does; an entry untested for a quarter is dead.
4. The record. Dates, attendees, content, tools, follow-ups. It is the Article 4 evidence, the first thing to produce when an underwriter asks “Do you use AI? Do you police it? Do you have protocols in place?” (Aon’s Stan Sterna, April 2026), and the answer to Judge Manasco’s question. The firm implementation playbook shows where it sits in a rollout.

About AI Lab for Lawyers, honestly placed

I teach AI Lab for Lawyers on Maven, so read this section as coming from an interested party.

The format is the one this page argues for: four live two-hour sessions, recorded, hands-on in the browser with ChatGPT, Claude Cowork, Perplexity, NotebookLM and Harvey, no coding, more than 300 practical slides, homework, a community and a certificate of completion. It is rated 4.8 from 320 reviews and has drawn participants from more than 60 firms, among them Norton Rose Fulbright, Paul Hastings, Pillsbury, Taylor Wessing and Withers. Maven’s guarantee applies, and firms often reimburse it.

What it is not: a regulation course. The FAQ answers “Will you deal with AI risks and the regulation of AI?” with “Not really.” That is deliberate, and it is why the ethics and verification guides on this site exist. Nor is it accredited CLE: the Maven page lists a certificate of completion, so treat any credit as self-reported where your state permits it.

Two reviews from the September 2026 cohort describe what the format does. Lukas, an associate at Wolf Theiss: “I do feel much more confident now when handling it.” Christian, an attorney and certified tax expert: “Highly inspiring. Now, the work on AI starts…” The second is the right review. Eight hours starts it; the six months is yours.

Where to go next: what law schools now teach about AI if you hire graduates, and the learning paths hub for students, paralegals and course comparisons. The prompts here, and a hundred more, are in the prompt library; the place to practise them live, on your own documents, is the Lab.