The AI skills for lawyers that actually matter: a map, a self-assessment and how to learn each

Only 9% of lawyers in the State Bar of Texas’s 2026 survey were “extremely confident” evaluating AI output; 40% were not or only slightly confident. LexisNexis asked nearly 900 UK lawyers what AI was doing to juniors and found 72% naming deep legal reasoning as the biggest skills gap and 69% weak verification. Everyone agrees on AI skills for lawyers in the abstract; nobody says which.

The regulators do not help: ABA Formal Opinion 512 asks for “a reasonable understanding of the capabilities and limitations” of the tool, and the EU AI Act’s Article 4 asks firms to support “AI literacy”. So here is a map of the AI skills lawyers need: eight of them, three levels each, and a way to score yourself.

From “AI literacy” to eight AI skills for lawyers

Article 4 has applied since 2 February 2025; the July 2026 Digital Omnibus softened “ensure” to “take measures to support the development of”. The European Commission’s Q&A says “no specific – or ‘sufficient’ - level is mandated” and that “simply relying on the AI systems’ instructions for use or asking the staff to read them might be ineffective”. No fine attaches to Article 4, but national sanctions are “more likely if there is proof of an incident due to lack of appropriate training”; the Article 4 guide covers the paperwork.

A generic definition fails for an empirical reason. In the first randomised trial, Choi, Monahan and Schwarcz found GPT-4 “only slightly and inconsistently improved the quality of participants’ legal analysis but induced large and consistent increases in speed”. Schwarcz and colleagues then found a reasoning model and a RAG tool did improve quality, with productivity gains of 50-130% on five of six tasks; the RAG tool produced three hallucinations, students with no AI four, the reasoning model eleven. The gains depend on the tool, the task and the user, so the skills must be specific about all three.

Skill 1: prompting with context

Zack Shapiro’s contrast is the lesson. Bad: “review this contract”. The good one names the side (“from the vendor’s perspective”), the standard (“risk beyond market norms”), a missing-provisions checklist, the output (“a severity-rated summary with specific counter-language”) and the commercial posture. His summary: “The entire gap between ‘AI is a toy’ and ‘AI changed my practice’ lives in the quality of your instructions.” Clio’s gut check: if the same prompt could apply to a range of different matters without changing a word, it is too broad.

Jurisdiction: [jurisdiction]. I act for [the customer], a [size, sector] business whose posture is [wants to close in two weeks]. Task: [review / draft / extract]. Use only the materials I supply; tag any case, statute or rule you add [VERIFY]. Output: [a table with these columns]. Success criterion: [completeness over brevity]. If the request could be read in materially different ways, ask me up to three questions first.

The prompt engineering guide has the patterns; “think step by step” and “double-check your answer” are unnecessary on reasoning models, and Anthropic says the second now causes over-verification.

Skill 2: verification

Damien Charlotin’s database listed 2,039 court decisions involving hallucinated material by 12 September 2026, 811 of them lawyers’; Illinois’ appellate court says “The only acceptable standard is zero false citations.” Six layers: existence in a real database; names and reporter match; status via a citator; the pinpoint says what you say; quotations match character for character; right jurisdiction and posture. The dangerous failure is misgrounding, a real case cited for a proposition it does not support. Never ask the model whether its own citations are real (Schwartz did, in Mata v. Avianca). The citation verification guide is the protocol.

Skill 3: context and knowledge design

Shapiro again: “The difference between a firm playbook and an individual lawyer’s encoded judgment is the difference between giving someone a recipe and teaching them how to cook.” Level 2 is a Project or custom GPT with safety rules first, then house style, then an anonymised clause bank. Level 3 is versioning that knowledge for a team, which is where Zuva’s Noah Waisberg lands his warning: your prompts and workflows become “software… except without the QA, the versioning, the user feedback loops, or the ability to survive someone leaving the firm”. The Projects and custom GPTs guide is the build manual.

Skill 4: tool selection and confidentiality triage

The distinction that matters is consumer versus commercial tier. ChatGPT Free, Plus and Pro and Claude Free, Pro and Max train on conversations by default; Business, Team, Enterprise and API tiers do not. In United States v. Heppner (S.D.N.Y., February 2026) a defendant’s own consumer-Claude exchanges were held to carry no privilege. Level 2 is a classification you apply without thinking: abstract questions on any tool; anonymised client material on a no-training tier; privileged strategy and anything filed only on enterprise zero-data-retention terms.

Skill 5: workflow and agent design

GC AI’s rule: “An agent you can trust is a workflow you already standardized, running on a schedule, with your name still on the review.” Harvey’s agent guide names six governance dimensions: scope of access, authorised actions, reviewability, matter-level isolation, deployment governance and accountability. The agent explainer has examples.

Skill 6: supervising AI output (Rule 5.3 for machines)

ABA Opinion 512 requires supervisory lawyers to “make reasonable efforts to ensure that the firm’s lawyers and nonlawyers comply with their professional obligations when using GAI tools”; California’s 2026 guidance says lawyers “must not permit AI systems to autonomously file documents, communicate with the court, or make representations on the lawyer’s behalf”. Harvey: “Agents do not sign documents. Lawyers do.” Level 3 is a training problem: a first-year who spent 40 hours on a markup “might now spend 5 hours reviewing an agent’s markup”, so Harvey recommends “intentional friction”, drafting by hand before comparing; without it the replacement question answers itself.

Skill 7: evaluation and benchmarking

Vals’ method is the template: identical instructions and documents to humans and tools, with a deadline. Level 1 is Stanford’s self-tests: “Why did Justice Ginsburg dissent in Obergefell?” (she did not), a fictitious judge, an overruled precedent presented as current. Level 2 is three closed matters where you know the outcome, recorded by model version.

Design five tests I can run on [tool] to probe legal hallucination: (1) a false-premise question about a dissent that was never written; (2) a fictitious judge or party; (3) an overruled precedent presented as current; (4) a jurisdiction trap ([Texas] question, watch for [California] law); (5) an "are these real?" trap where I supply one real and one invented citation. For each: the exact prompt, the correct behaviour and the failure behaviour. I will run them and record results by model version.

Skill 8: pricing and communicating AI work

North Carolina’s 2024 FEO 1 is the ethics floor: the $300-an-hour estate planner whose three-hour draft now takes one “may not bill a client for three hours”. Clio’s framing is the business problem: “If a matter used to take five hours and AI brings it down to one, hourly billing means you’ve just handed your client an 80% discount they never asked for.” Level 3 is pricing one predictable matter type as a flat fee against what it actually costs to deliver today.

Prepare one page of talking points for a conversation with [client] about how we use AI on their [matter types] and how it affects pricing, in this order: what changed about the work; what changed for the client; what value that created; how pricing should reflect it. Include where AI saves time and where it does not, our verification commitment, our billing rule (actual time only) and two alternatives: a fixed fee for [matter type] and a capped hourly rate. Lead with value, not concessions.

The legal-engineer path, and a self-assessment

Legora deploys through “Legal Engineers”, Norm Law employs 35 under that name, and Harvey Academy lists a Certified Legal Engineer path. Richard Susskind: the future lawyer will “build, maintain, supervise, and improve the systems”. None of it requires code: Clifford Chance’s Jamie Tso argues that “by asking the AI to build a tool rather than answer a question, the result becomes deterministic, significantly reducing the error rate”, and an immigration lawyer who asked ChatGPT “How do I open a terminal?” in April 2025 had a practice platform within a year; the vibe-coding guide covers the path and its security limits.

Score each 0 (no), 1 (safe task) or 2 (live work, with a record).

1. Could a colleague run my last prompt on a different matter without changing a word? (If yes, score 0.)
2. Did I open every authority in my last AI-assisted document in a real database, and log it?
3. Do I have a Project or skill with safety rules, house style, anonymised precedents and a “last tested” date?
4. Do I know which tier of each tool I am on and what data it may hold?
5. Have I standardised a workflow before automating it, and can I name the six governance dimensions for any agent I run?
6. Do I read agent output as I would a first-year’s draft, and do juniors still draft by hand first?
7. Have I tested my main tool on three closed matters and the false-premise questions?
8. Can I explain to a client what AI changed about their matter and how the price reflects it?

Under 6: Susskind’s recipe, 30 minutes a day for six months on one tool, the regime that “transformed” a general counsel’s team. 6 to 11: own the lowest-scoring institutional skill for your group, the way Barnes & Thornburg’s AI Practice Champions work “directly inside practice teams”. 12 or more: the firm should build its training around you.

Where to go next: the careers hub has the wider picture, the training guide compares formats, and the prompt library has the patterns behind Skill 1. AI Lab for Lawyers is four two-hour live sessions built around exactly these eight skills, on browser tools, no coding; score yourself on the map before the first session and again after the last.