AI Training for Law Firms: A Practical 90-Day Plan

AI training for law firms means teaching every lawyer and staff member the duties generative AI touches (competence, confidentiality, verification of authorities, supervision and fees), then running a role-based, tool-agnostic program built on a one-page AI use rule and one champion per practice group. Courts in Canada, the US and England and Wales hold the lawyer who signs responsible for fabricated authorities, whichever tool produced them. One full cycle fits in 90 days (an AIDOLS planning estimate).

Key figures (updated 23 October 2026)

• 2,207 court decisions on AI-hallucinated content were listed in a public database on 23 October 2026, 877 involving lawyers and 234 from Canada (AI Hallucination Cases Database).
• 54% of law firms offered no AI training and 43% had no AI governance policy in a legal software vendor's survey of about 1,300 respondents (8am, 2026).
• 1 December 2024: since then, every factum under Ontario's Rules of Civil Procedure must include a signed statement that the signer is satisfied every authority cited is authentic (O. Reg. 384/24).
• Six months: the license suspension the Law Society Tribunal ordered in July 2026 after finding a lawyer relied on a factum of non-existent or irrelevant AI-generated case law and deliberately misled the court (Law Society of Ontario v. Lee).
• 17% to 33%: hallucination rates of legal research tools from LexisNexis and Thomson Reuters in a Stanford-led test of 2024 versions (JELS, 2025).
• 2027: when Microsoft now expects in-country Copilot processing for Canada, a year later than first announced (Microsoft).

Cite as: AIDOLS Research Team, AI Training for Law Firms: A Practical 90-Day Plan, aidolsgroup.com, 23 October 2026.

This guide covers Canadian, US and UK rules, a curriculum by role, which AI accounts can hold client data, a seven-step verification protocol, billing, a 90-day plan and a one-page AI use rule.

Because competence now includes technology, courts sanction the lawyer rather than the tool, and policies alone have not stopped fabricated citations.

Competence now includes technology

The Federation of Law Societies' Model Code says a lawyer "should develop an understanding of, and ability to use, technology relevant to the nature and area of the lawyer's practice" (rule 3.1-2, commentary [4A]) (FLSC Model Code). In the US, Comment 8 to ABA Model Rule 1.1 says competence includes keeping abreast of "the benefits and risks associated with relevant technology" (ABA). Once your people use generative AI, training is part of competence, not a perk.

Courts sanction the lawyer, not the tool

The leading cases share one pattern: a lawyer files authorities a chatbot invented (an AI hallucination), and the court holds the lawyer responsible.

• British Columbia. In Zhang v. Chen, 2024 BCSC 285, counsel cited two non-existent cases invented by ChatGPT. The court said citing fake cases "is an abuse of process and is tantamount to making a false statement to the court" and made counsel personally bear the extra costs (BC Supreme Court).
• Ontario. In Ko v. Li, 2025 ONSC 2766, the court wrote that "it is the lawyer's duty to ensure human review of materials prepared by non-human technology such as generative artificial intelligence" (CanLII). In July 2026 the Law Society Tribunal suspended an Ontario lawyer for six months, with $10,000 in costs, for relying on a factum of non-existent or irrelevant AI-generated case law, deliberately misleading the court and being dishonest with the Law Society (Law Society Tribunal).
• United States. In Mata v. Avianca (S.D.N.Y., 2023), the court imposed a $5,000 penalty on two lawyers and their firm for ChatGPT-invented opinions they kept defending after the court questioned them (sanctions opinion).
• England and Wales. In Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin), the Divisional Court said freely available tools such as ChatGPT "are not capable of conducting reliable legal research", told managing partners and heads of chambers to take effective measures so everyone providing legal services meets their duties, and warned that courts will "inquire whether those leadership responsibilities have been fulfilled" (judgment, 6 June 2025).

Policies and warnings alone have not been enough

In Johnson v. Dunn (N.D. Ala., July 2025), lawyers at a large, well-regarded firm filed fabricated ChatGPT citations despite "repeated warnings, internal controls, and firm policies about the dangers of AI misuse"; three were publicly reprimanded, disqualified from the case and referred to the state bar (sanctions order). The Ayinde court agreed that "promulgating such guidance on its own is insufficient to address the misuse of artificial intelligence" (judgment). A policy is necessary; training and a verification routine make the check happen the night before a deadline.

None of the rules below bans generative AI; all make the lawyer responsible for accuracy and confidentiality. For training, the court differences matter most: Ontario requires a signed statement that the authorities in a factum are authentic, the Federal Court of Canada requires a declaration when AI generated content, and England and Wales has no general declaration rule. This guide is general information, not legal advice; check your law society's or regulator's current guidance.

Canada: law societies and courts

By March 2025, eleven Canadian courts and nine law societies had issued generative AI guidance, and the Law Society of Alberta found they "uniformly emphasize that lawyers are responsible for the truth and accuracy of their work" (Law Society of Alberta).

United States: ABA Opinion 512 and state bars

ABA opinions are advisory and states adopt their own rules, but ABA Formal Opinion 512 (29 July 2024) maps each duty to generative AI.

England and Wales: the SRA, the Law Society, the Bar Council and the courts

Everyone, from partners to the marketing team, takes the same six-module core, one module per duty, before a role track. The LSO organizes its guidance around six duties and lists training on AI's "limitations, potential biases, and ethical pitfalls" among the guidelines firms should give staff (LSO practice note); Opinion 512 adds secure data handling (Opinion 512).

Six core modules, one per duty

1. Competence and verification: tool strengths, failure modes and the firm's protocol. The LSO says verification "should be completed by a human being" (LSO practice note).
2. Confidentiality and privilege: accounts, data and settings (next section).
3. Supervision: treat output as the work of "a law student who requires supervision" (Law Society of Alberta).
4. Candor to the court: each court's certification, declaration and practice-direction rules.
5. Fees: actual time only, and no charge for learning a tool.
6. Client communication: retainer wording, consent and outside counsel guidelines.

Teach tool skills and ethics separately; the Law Society of England and Wales calls them distinct and "both are equally important" (Generative AI: the essentials). End every module with an exercise on redacted firm documents or fictional matters.

Client information goes only into firm-managed accounts with checked contracts, retention and training settings, for approved tasks. The LSO suggests allowing AI only through accounts created with the firm's own credentials (LSO policy checklist), and the SRA warns that once client letters go into an open tool such as ChatGPT, privilege "may be permanently waived and unable to be recovered" (SRA).

Consumer accounts versus firm accounts

Use the four questions that decide what you can paste into an AI tool to sort each document into paste, redact or refuse, then match it to an account:

Data residency for Canadian firms

Do not promise clients Canadian data residency unless the vendor's documents cover both storage and processing. Microsoft says customers outside the EU "may have their queries processed in the US, EU, or other regions" (Microsoft) and expects in-country Copilot processing for Canada in 2027 (Microsoft). OpenAI offers eligible new ChatGPT Enterprise and Edu customers storage at rest in Canada, but inference residency only in Europe, the US and the UAE (OpenAI).

PIPEDA allows processing abroad, but the transferring organization stays accountable and must secure comparable protection by contract (Office of the Privacy Commissioner); Quebec's private-sector law requires a privacy impact assessment before personal information leaves the province, including when a provider outside Québec keeps or uses it for the firm (LégisQuébec). Confidentiality is wider than privacy law and client contracts can demand Canadian residency, so check both. Our guide to PIPEDA, Quebec's Law 25 and what happened to AIDA covers the privacy side.

Client consent, engagement letters and outside counsel guidelines

Opinion 512 requires informed consent before representation information goes into a self-learning tool that could disclose it, even to colleagues behind an ethical wall, and says "general, boiler-plate provisions" in engagement letters are not sufficient (Opinion 512). The LSBC suggests exploring consent where redaction is impossible (LSBC); the Law Society of Alberta says to explain safe AI use in the retainer letter (Playbook). Opinion 512 also requires disclosure when a client asks or when the engagement agreement or outside counsel guidelines require it, so read those guidelines before AI touches a matter.

Permissions and ethical walls in Microsoft 365

Copilot "only surfaces organizational data to which individual users have at least view permissions" (Microsoft), so an over-shared matter folder, including one meant to sit behind an ethical wall, can surface in a colleague's Copilot answers. Our guide on how to fix Copilot oversharing before rollout sets out the steps. The SRA warns that "both paid for and free-to-use AI tools may not provide the contractual, and technical safeguards needed to maintain client confidentiality" (SRA). Put each candidate tool through an AI vendor security questionnaire first.

Rank each task by the cost of an error and whether the output leaves the firm. Opinion 512 says checking "will necessarily depend on the GAI tool and the specific task", and that a lawyer who has tested a tool against a manually reviewed sample may not need to re-check every document (Opinion 512). The Law Society of England and Wales' new practice note asks for human involvement and safeguards scaled to the risk of each use, in both the business and the practice of law (practice note).

Never allowed: client confidential data in free or consumer tools, authorities not pulled from a primary source, and reliance on a client's own AI research. In Al-Haroun v Qatar National Bank, decided with Ayinde, 18 of 45 citations did not exist and the solicitor had relied on the lay client's research (judgment).

Verify every authority and quotation against a primary source before filing, and make the signer responsible. These seven steps combine Ontario's certification rule, the Superior Court's practice direction and the prefiling protocol the Johnson v. Dunn firm committed to once its fake citations surfaced: "confirming the existence, accuracy, and relevance of each citation" (sanctions order).

1. Pull every authority from a primary or presumed-authentic source. Ontario presumes authentic an authority published on a government website or by a government printer, on CanLII, on a court's website or by a commercial publisher of court decisions (rule 4.06.1(2.2)). Never rely on a link, citation or self-check the AI produced.
2. Read the cited passage and confirm it supports the proposition. Ko v. Li stresses the duty to read cases before submitting them (CanLII).
3. Compare every quotation word for word with the source.
4. Confirm the authority is still good law, from the right jurisdiction and level of court.
5. Look for what is missing: contrary authority and qualifications a summary dropped.
6. Log the check: who checked, when, and every error found. The LSO recommends recording AI outcomes, including each hallucination (LSO quick-start checklist); that log becomes your cite-check error rate.
7. Make the signer own the check. Ontario's certification is a signed statement (O. Reg. 384/24); the Johnson v. Dunn firm said junior lawyers who sign a filing must verify it (sanctions order). In the Federal Court, add the AI declaration where AI generated content (Federal Court).

Legal-specific tools need the same steps: tools built on retrieval-augmented generation erred less than a general chatbot in the Stanford-led study, but still erred (JELS study), and Opinion 512 warns that relying on output "without an appropriate degree of independent verification or review" could breach the duty of competence (Opinion 512). Steps 2 to 5 are the legal version of our four-pass verification routine.

Bill hourly work at the time actually spent, including prompting and review; never bill the time a task would have taken without AI; do not charge for learning a tool the firm uses regularly; and treat general AI subscriptions as overhead.

Train timekeepers to record actual time, including prompting and review, and billing partners to track write-offs on AI-assisted matters and consult clients when AI bears on the basis or reasonableness of the fee, as Opinion 512 requires.

Every role takes the shared core, then a short track on its own work, because lawyers sign, clerks prepare, assistants send and IT configures.

Launching its practice note, the Law Society of England and Wales' president warned that over-reliance on AI may cause "a catastrophic gap of future legal skills and knowledge" (Law Society Gazette), so associates should do some work without AI on purpose.

Our AI Fluency course has no legal track; partners fit its Leadership track and the professional-development lead its L&D and People track, so choose a track by the decision you own.

One cycle takes 90 days; the day ranges are AIDOLS planning estimates for a firm of 10 to 500 lawyers.

1. Write the one-page AI use rule (days 1–14). Use the template below, list approved accounts, check engagement letters and outside counsel guidelines, and ask your insurer whether AI use is covered, as the Law Society of England and Wales advises (Generative AI: the essentials).
2. Baseline the measures (days 1–14): turnaround on one recurring work product per practice group, first-draft hours and write-offs, from records the firm already keeps.
3. Enforce the prefiling verification protocol (from day 15, permanently): the seven steps and a log entry for every filing, deadlines included; the log's first month is the cite-check baseline.
4. Train partners on duties, sanctions and billing (days 15–30). They sign filings, set fees and supervise, and Ayinde says courts will ask what leaders did.
5. Appoint one champion per practice group (days 15–30), each with a written charter.
6. Run role tracks on real, redacted work (days 31–60), so everyone leaves with a reusable asset, the standard in our lesson on training that leaves every participant with a working asset. Include drills with planted fake citations.
7. Measure at day 90 and renew the rule (days 61–90): compare with the baseline, review incidents and near-misses, update the rule, renew or rotate champions and pick the next practice groups.

Who should be the practice group champion?

Pick the practitioner colleagues already copy, often a senior associate or clerk who uses the tools on real matters, not the loudest volunteer. The charter names the group, three asks (a clinic every two weeks on live matters, current shared instructions, a Friday note on what broke), hours agreed in writing with the supervising partner (two hours every two weeks, an AIDOLS planning estimate), an end date, and a refusal list routing licensing, data and policy questions to named owners; our lesson on how to write a champion charter with a refusal list has a template.

Make the champion a coach, not a gatekeeper: the Johnson v. Dunn firm required "written permission from the appropriate Practice Group Leader" before AI use, and fabricated citations still reached the court (sanctions order). Our comparison of AI champions versus whole-team training covers the trade-off; in a law firm the duties bind everyone, so the core cannot wait for the cascade.

How to measure whether the training worked

Measure the work, not attendance:

• Turnaround: elapsed time for one recurring work product per practice group, from matter-management timestamps.
• First-draft hours: time entries for first drafts of that work product.
• Write-offs: time written off on AI-assisted matters, from billing records.
• Cite-check error rate: errors found per filing, from the cite-check log, whose first month is the baseline.

Expect the cite-check error count to rise at first, because the checks now find errors; judge the trend over the next quarter, and set targets only once the baseline exists. Our lesson on how to measure adoption as a countable change in work explains why seat counts measure access, not adoption.

No. Train duties, work shapes and the verification habit, then show them in whatever the firm already licenses, so the program survives a vendor change; the AI Fluency course is vendor-neutral by design for that reason. Three categories cover the market (examples, not endorsements):

• General assistants: Microsoft 365 Copilot, ChatGPT Business and Enterprise, Claude Team and Enterprise, and Gemini in Google Workspace.
• Legal-specific platforms: Harvey; research platforms with built-in AI, such as Thomson Reuters' CoCounsel, launched for Canadian Westlaw users in June 2024 (Thomson Reuters), and Lexis+ AI, available to Canadian customers since July 2024 (LexisNexis); and specialists such as Alexi, with offices in Toronto and New York, for research and drafting (Alexi), and Blue J for tax research (Blue J).
• AI inside software the firm already owns: features in document management, practice management and word processing tools.

Legal-specific does not mean error-free, as the Stanford-led study showed, and a paid license does not settle confidentiality on its own. To compare the general assistants, see Microsoft Copilot vs ChatGPT Enterprise vs Claude compared, and for the reasoning behind this stance, why model-agnostic AI training outlasts tool changes.

It can, under each regulator's own rules. In Ontario, lawyers must complete at least 12 CPD hours a year, including at least 3 Professionalism Hours (LSO). Substantive Hours need not be accredited, and non-legal subjects can qualify "if they are relevant to the lawyer's or paralegal's practice and professional development", so practice-relevant AI training can count there; confirm with the LSO. Professionalism Hours count only if the program is accredited by the LSO or a designated Accredited Provider of Professionalism Content (LSO). Florida requires at least 3 of every 30 CLE hours, per three-year cycle, in approved technology programs (The Florida Bar). The AIDOLS AI Fluency course is not LSO-accredited, so it cannot supply Professionalism Hours.

If any of these is true, fix that first:

• No approved firm account exists yet. Training on consumer tools teaches people to put client data in the wrong place. Procure accounts and publish the rule, then train.
• Leadership wants AI research nobody verifies. No program fixes that.
• The bottleneck is a workflow, not a skill. If intake, conflict checks or billing narratives are slow because of the process, map it first with the free workflow map below.
• The firm has only a handful of lawyers. Read the free lessons, adopt the one-page rule and the verification protocol, and skip the program structure.
• A key client's outside counsel guidelines restrict AI on its matters. Comply first, and train on other matters.

A law firm AI policy can run long; the operating rule people recall at the moment of decision fits on one page, as our lesson on the one-page AI ground rule explains. Name tasks, documents and people, not categories.

Header: owner (a named partner), approval date, next review date.

1. Approved tools: the firm accounts, by name. Never personal accounts for firm work.
2. Green, no need to ask: summaries of the firm's own precedents and transcripts, internal memo and email drafts, checklists, marketing and HR drafts without personal data.
3. Amber, ask the practice group champion (answer within one working day): clause comparisons, chronologies from client documents, any new tool or new type of data.
4. Red, no exceptions: client confidential information in free or consumer tools; any authority not pulled from a primary source; client-facing chatbots without controls; filing a client's own AI research.
5. Verify before it leaves the firm: the seven-step protocol for every authority and quotation; the signer owns the check; the Federal Court declaration where AI generated content.
6. Billing and disclosure: actual time only, including prompting and review; no charge for learning a tool; read the engagement letter and outside counsel guidelines first.
7. Incidents: report a suspected confidentiality breach or fabricated citation to the named partner the same day, and log every hallucination found.

• To start this week at no cost: the free AI Fluency course lessons are publicly readable with no sign-up, and one full path takes 228 minutes. Paid enrollment adds guided builds, starter templates, instructor video, cohort office hours, a graded capstone, a certificate issued for a working build that was submitted and marked (never for time spent), and roster and completion reporting for organizations. Individual enrollment is not open yet.
• If one workflow is the problem, such as intake or cite-checking: in the free workflow map of one legal workflow, AI agents interview the 2–3 people who run it, and within three business days you get a one-page map and a written verdict on what each bottleneck needs (software, a process change or AI), with no obligation.
• To plan a firm-wide program around your tools, practice groups and courts, book a 15-minute call.