AI training for in-house legal teams: privilege, confidentiality, and drafting

Search ai training for legal teams and almost everything on page one is written for a law firm. Firm-size adoption rates, continuing-education credits, billable-hour economics, hallucinated citations in filings. A 300-person company’s legal function is one or two lawyers doing commercial contracts, employment questions, and vendor paperwork, and they never file anything.

The question that matters to a corporate legal department was answered in February 2026 by a federal judge, and no training page in the results mentions it. In United States v. Heppner, Judge Jed Rakoff of the Southern District of New York held that a defendant’s written exchanges with a consumer generative AI platform were protected by neither attorney-client privilege nor the work product doctrine. That ruling, not a course catalog, should set what your legal team and the people around it learn first.

What the market sells under this name

Three products hide behind one phrase, and the prices run two orders of magnitude apart.

A credential for one lawyer. Berkeley Law’s Introduction to GenAI for the Legal Profession: Power User Edition lists at $950, or $800 for alumni and public-sector employees: eight hours of self-paced content over four weeks, 6.5 hours of continuing-education credit, taught by a sitting chief legal officer. It produces one skilled lawyer, not one company standard.

A catalog course. AI for Legal Professionals on Coursera runs nine modules in roughly nine hours at beginner level. Check the price yourself: the best in-house roundup in these results, GC AI’s shortlist, calls it free to audit and about twenty hours, while the live page is a nine-hour program behind a certificate purchase. Catalogs change quarterly; roundups do not re-check.

A credential that does not exist yet. The top organic result is the IAPP’s AI in legal services initiative, whose AI Specialist, Legal certification is slated for January 2027. Waiting for a credentialing body costs you the year.

Nobody sells the fourth thing: two lawyers and the non-lawyers who send them work agreeing on which material enters which account, and what gets checked before it goes out.

Privilege is the design constraint

The reasoning, not the headline, is what you teach from. Heppner, under indictment for securities fraud, used a public assistant to generate reports on his defense strategy after a grand jury subpoena, then shared the output with his lawyer. The privilege claim failed on at least two elements and possibly all three.

• The platform is not an attorney. As Judge Rakoff put it, Heppner “does not, and indeed could not, maintain that Claude is an attorney.”
• The exchanges were not confidential. The platform’s terms reserved rights to collect and share user data, so there was no reasonable expectation of confidentiality to begin with.
• Nobody directed the work. He started on his own rather than at counsel’s direction, which also sank the work product claim: the material may have been prepared in anticipation of litigation, but not at the behest of counsel.

The part to put on a wall: handing the output to his lawyer afterward did not retroactively cloak it in privilege. There is no repair step.

One nuance drives a buying decision. The court left open whether an enterprise product that excludes customer data from training and carries contractual confidentiality terms would be analyzed the same way, while cautioning that a contract alone guarantees nothing. So the account your team uses is a legal question, not an IT preference, and the gap between a personal login and a company tenant is the gap between two outcomes in a discovery dispute.

The exposure is mostly not the lawyers

The person who destroyed the privilege in Heppner was not a lawyer, and at a mid-sized company it will not be one either. It will be the sales lead pasting a customer redline into a personal account at 10pm, the recruiter uploading a separation agreement, the founder asking an assistant how bad the dispute looks.

A July 2026 survey of 500 employed US adults, run on the Pollfish panel and reported through Stacker, found 38% had entered some category of work information into a personal AI account their employer does not control. Of those surveyed, 23% pasted internal emails, memos, or documents and 11.4% entered contracts or legal documents. Only 35.8% said their employer has a clear written policy. Treat a 500-person panel as directional rather than precise; the direction is not in doubt.

So AI enablement that trains only the lawyers leaves the actual failure mode untouched. Our role-by-role view of what each function needs sets out why one company-wide session cannot cover this: it has to say something true to finance and sales at once. Recruiting warrants its own session for the same reason, covered in our look at sourcing, screening, and the legal line, as does the wider people function in our comparison of options for HR and people teams.

Why informed consent lands differently in-house

ABA Formal Opinion 512, issued July 29, 2024, says a lawyer need not become an AI expert but must hold a reasonable understanding of the capabilities and limitations of the tools they use, and may acquire it through reading, continuing education, or training. That makes structured training a recognized compliance pathway rather than a nice-to-have.

On confidentiality it requires a client’s informed consent before information relating to the representation enters a self-learning tool, and says boilerplate consent in an engagement letter will not do. In a firm, that produces a letter to write. In-house, the client is the organization, so there is no outside party to ask, and the question collapses into an authorization decision the company makes about itself: which tenant, which matters, who decides, who is told. That reframing is absent from the course market. The distance between a written policy and Tuesday behavior is the subject of our guide to policy, training, and the gap between them.

Drafting is where the hours are

Privilege is the constraint. Drafting is the payoff, and it is the work a small department drowns in. Three tasks eat a corporate lawyer’s week and current assistants are good at all three.

• First drafts from a position. Your own executed agreements plus the deal terms, producing a draft in house language rather than a form-book template.
• Review against a playbook. Your fallback positions, written down once, applied to an inbound redline to surface every deviation and its severity.
• Translation into business English. The note explaining why the indemnity cap cannot move, written forty times a year.

What they are not good at is knowing when they are wrong, and the failure is not limited to invented case names. The hallucination record in legal work splits three ways: fabricated authority, misstated holdings, and false quotes from genuine sources. Training that covers only the first leaves the harder two in place, and in commercial work the analogue is a confidently wrong claim about what your own contract says. Verification is the skill, not prompting.

The economics will survive a budget request. Thomson Reuters’ Future of Professionals Report 2025, a survey of 2,275 professionals, found respondents expect AI to save about five hours a week, worth roughly $19,000 per person annually, and that firms with an AI strategy were 3.9 times more likely to see benefits than those with no adoption plan.

What a half day with a legal function covers

Four blocks, run on your own executed agreements and open redlines.

The account decision, written down first. One page naming which tenant is approved for legal material, which matters may never enter an assistant, and what happens in the awkward middle case: a demand letter, an internal investigation, a regulator’s question. Write it in the session, because a standard drafted with no live matters open does not survive a Tuesday.

The privilege failure, demonstrated. Walk the three Heppner elements against a real example from your own week, then run the wrong version deliberately and read what the terms of service say about where it went.

Drafting, hands on. Each lawyer brings a live agreement and leaves with a first draft, a playbook-based review of an inbound redline, and the business-English note that gets written last and worst. Most of the session belongs here.

Verification with teeth. A general counsel who tells a CEO the cap is $2M when the agreement says $5M loses standing that takes a year to rebuild. The mechanics behind a session that survives the following week are in our guide to building a session that changes Monday morning.

What it costs when you price the hours

Assume two in-house lawyers and four people in the functions that send them work. The BLS median annual wage for lawyers was $159,670 in May 2025, about $77 an hour before benefits, against $46.89 per hour worked in average employer compensation cost for private-industry workers as of June 2026.

Line	Three credential seats	One facilitated half day for six
Invoice	$2,850 at $950 a seat	$6,000, mid-market for a $2,000–15,000 range
Participant hours	24 lawyer hours, about $1,850	8 lawyer hours plus 16 others, about $1,370
Reinforcement	Unscheduled	Two 1-hour follow-ups, about $680
Coordination	Low	About 8 hours, roughly $400
Loaded total	About $4,700	About $8,450
What you hold afterward	Three skilled individuals	One standard six people share

Read the last row before the first. Three seats is the cheaper line and the longer time commitment, because the participant-hour penalty falls on your most expensive hours, and it does not buy a shared answer to which account the redline goes in. Our comparison for finance and accounting teams runs the same arithmetic against a shorter syllabus and reaches a different answer: the shape of the program decides it, not the sticker. Treat a company-wide program as a separate decision with its own sponsor, for the reason in our argument that buying licenses is not the same as building capability.

What we would buy for a two-lawyer function

Assume 300 employees, two lawyers, licenses already bought, nobody with AI in their job title.

1. One credential seat for whoever will own the standard. Berkeley’s program at $950, one seat, not three.
2. One facilitated half day on live agreements, with both lawyers and the three or four people who send them the most work. Output is the account standard, a drafting method, and one verification check everyone performs identically.
3. Two one-hour follow-ups at three and six weeks. Bring the thing that went wrong, and see where the method holds or quietly lapses. About $680 of paid attention.
4. A named owner with two hours a month, with the measure in their objectives rather than a vendor’s completion report, and one written question to your contract tooling vendor: does any enabled feature send document text to a model outside our tenant.

Loaded cost lands near $9,400 against $4,700 for credentials alone. Invert it if your legal function is one person: buy the credential and skip the session, because a shared standard needs more than one head.

How to tell in ninety days whether it worked

An ACC and Everlaw survey of 284 chief legal officers, general counsel, and legal operations professionals, published November 17, 2025, found 81% citing greater speed of legal support and matter resolution as a top benefit of generative AI. In the same survey, 12% track technology return on investment and 9% measure impact on business outcomes. Four fifths believe it is faster; one in eight can show it.

Before the session, time two recurring items, because rough timekeeping beats a number reconstructed afterward: a commercial agreement from intake to signature-ready draft, and an inbound redline from receipt to a summary of deviations. At ninety days, check four things.

• Cycle time on those two items, same definition, same people.
• Whether the account standard is in use. Ask each of the six which account a customer redline goes into. Six answers means it did not land.
• Whether anything new entered the stack. Assistants and contract-tool features arrive without a purchase order.
• One quality check. Pull five assistant-drafted summaries and verify every factual claim against the executed document.

Self-reported confidence is deliberately absent. Turning a plausible gain into a defensible one is the subject of our guide to proving capability moved.

Frequently asked questions

What should an in-house legal team learn about AI first?

The account standard: which tenant is approved for legal material, which matters may never enter an assistant, and who decides the edge cases. Prompting technique is easy to pick up later. Privilege destroyed by a paste into a personal account is not recoverable, and Heppner confirmed that showing the output to a lawyer afterward does not repair it.

Does using ChatGPT or Claude waive attorney-client privilege?

On consumer tiers, assume yes for anything privileged. In United States v. Heppner, decided February 2026, the Southern District of New York held that exchanges with a consumer platform were neither privileged nor work product, partly because the terms permitted collection and sharing of user data. The court left open whether enterprise products that exclude customer data from training would be treated differently, and warned that a contract alone is no guarantee.

Is an AI certification for lawyers worth the money?

For one person on the team, yes. Berkeley Law’s program at $950 for eight hours is the most substantial option a single lawyer can buy, and it carries continuing-education credit. Three seats produce three skilled individuals and no shared standard, the thing a two-lawyer department is missing.

How much does AI training for a legal team cost?

Credential seats run roughly $800 to $950 a person, catalog courses sit in the low hundreds or inside a subscription you may hold already, and facilitated sessions run $2,000 to $15,000 at market rates. Add participant time at about $77 an hour for a lawyer, the line most budgets omit.

Does AI training satisfy our legal and regulatory obligations?

One of them, and only if it applies to you. The EU AI Act’s Article 4 literacy duty, in force since February 2, 2025, is discharged by teaching the staff who operate AI systems, and binds only if you have staff in the EU. Your records, notices, and tenant terms are governance problems no session resolves.

Should we train the lawyers or the business functions?

Both, in the same room, if your legal function is small. Survey data puts 11.4% of workers entering contracts or legal documents into personal AI accounts, and the defendant in Heppner was not a lawyer. A session covering only the legal team leaves the people generating the exposure untrained.

What about the contract tool we already pay for?

Ask the vendor in writing whether any enabled feature sends document text to a model outside your tenant, and whether that text is excluded from training. Features ship quarterly and the legal team finds out last. The question underneath that choice is covered in our playbook on building the capability against buying it.

Our legal team is one person. What changes?

Buy the credential, skip the session for the legal function alone, and spend that budget on the business functions sending that person the most work. A standard needs two people to be a standard. The adjacent case for customer-facing teams is in our comparison for customer support and success teams.

Key takeaways

• Page one for this term answers a law firm’s question. A department of one or two lawyers does commercial contracts and employment questions, not filings, and needs a different curriculum.
• United States v. Heppner, February 2026, is the datable instruction: a consumer platform is not an attorney, its terms defeat confidentiality, nobody directed the work, and showing counsel the output afterward repairs nothing.
• The account is a legal decision, not an IT preference. The court left open whether an enterprise tenant that excludes data from training analyzes differently, and said a contract alone is no guarantee.
• The exposure is mostly generated by non-lawyers: 11.4% of surveyed workers entered contracts or legal documents into personal AI accounts, and only about a third reported a clear written policy.
• ABA Formal Opinion 512’s informed-consent requirement does not transplant in-house, because the client is the organization: it becomes an authorization decision the company makes about itself.
• Three self-paced seats cost about $4,700 loaded and buy three habits; a half day for six costs about $8,450 and buys one shared standard. One legal department in eight tracks technology return on investment well enough to tell the difference afterward.

If you are weighing a credential against a session for your legal function, book a free AI readiness call. We will walk your contract workflow and tell you which of the two you are missing.