Back to Guides
Legal Services 15 min read

Anatomy of an AI workshop for an in-house legal team

Anatomy of an AI workshop for an in-house legal team

Search for AI training for a legal team and you get registration pages. The first page is event listings, vendor webinars, cohort syllabi and law-firm service pages, and not one of them tells you what happens between the first minute and the last. That is the thing worth knowing before you spend the budget, because the difference between a session that changes next week’s contract review and one that does not is mostly a question of where the minutes went.

Three hours is the right slot for a corporate legal function of two to eight people. Below is the run of show, with the time budget stated — the number no sign-up page publishes.

Minutes Block What it buys
0–10 The frame What this session is for, what is off limits, what each person leaves holding
10–25 One review driven live A real inbound vendor agreement from this company, worked start to finish, including the place the assistant got it wrong
25–75 Everyone reviews their own document Each person works a live agreement from their own queue against a position the company actually holds
75–90 Compare and correct Four people read out what they got, wrong answers first, so the room sees the failure modes together
90–100 Break —
100–135 The second task: drafting from a position First-pass redline language, a response to a counterparty’s pushback, an employment answer
135–155 The never-paste list The room writes, in its own sentences, what does not go into an assistant and on which account
155–170 The save Three prompts worth keeping, written down with the document type each applies to
170–180 One commitment each A named matter each person will handle this way before the next week is out

Eighty-five of the 180 minutes have every participant working on their own live matters. That fraction is what you are paying for, and the first thing to check on any agenda you are sent. The same arithmetic applies at shorter run times, which we worked through for a single ninety-minute slot.

The pre-work is the load-bearing stage

Nothing above is possible without two artifacts collected a week ahead, and this is the stage that separates real AI enablement from a course you could have bought for $200.

A redacted document set. Six to ten agreements the team handles: inbound vendor paper, a mutual NDA, a data processing addendum, an order form with a negotiated service credit, an offer letter, a consulting agreement. Counterparty names, dollar figures, and personal data are stripped. What survives is structure, which is what the exercises need.

A named task list. Not “contract review” but the specific repeat jobs: triaging the inbound paper that lands in a shared inbox, finding the limitation-of-liability clause and comparing it to the company position, summarizing a 40-page master agreement for a sales lead who will not read it, drafting a first response to a counterparty’s security questionnaire.

Collect both from the team, not from the person who bought the session. A general counsel’s list and a contracts manager’s list overlap by less than half, and the second is where the volume is. The broader argument for building curricula around the actual work of each function rather than one shared deck is laid out in our role-based AI enablement playbook.

Who is in the room besides the lawyers

The highest-volume contract work at a 50 to 500-person company is not done by a lawyer. It is done by the sales operations analyst who sends the order form, the finance manager who chases a signed data processing addendum, the recruiter who adapts an offer letter, and the founder who signs an NDA on a phone between meetings.

Train only the legal function and you have trained the people least likely to create the exposure. For a legal team of two or three, the better room is the lawyers plus the three or four business people who send them the most work. Everyone sees the same failure modes at the same time, and the policy conversation that follows has the right people in it.

This is the clearest point of departure from how the market trains law firms, where the audience is genuinely the lawyers and the sequencing question is which lawyers go first. We covered that version of the problem separately in what partners need before associates adopt.

The two exercises that carry the session

Exercise one: a clause against a position. Each person takes a live agreement from their own queue and finds a single clause — indemnity, limitation of liability, termination for convenience, a data processing term — then compares what the counterparty drafted against what the company is willing to accept. The output is three sentences: what the clause does, how it differs from the company position, and what to ask for.

This exercise works because it has a checkable answer. Either the clause says what the participant reported or it does not, and the document is right there. Most published curricula replace it with a generic prompt on an invented scenario, which teaches nothing because there is nothing to be wrong about.

Exercise two: drafting from a stated position. The same person now drafts the first version of the response. Not final language, a first pass they will edit. This is where the time comes back.

The FTI Consulting and Relativity General Counsel Report, in the part released in March 2026, found summarization to be the most common use at 83% of respondents using or experimenting with it, with contract clause identification second at 63%. The reading tasks are where teams start. The writing tasks are where the hours are, and they are consistently the ones nobody has been shown how to do.

The never-paste list the room writes itself

Twenty minutes, late in the session, and the output is one page in the room’s own words. It comes late rather than first because a confidentiality rule means nothing to someone who has not yet seen what the tool does with a document.

Four questions, answered out loud:

  1. Which account. The company tenant, a named enterprise product, or a personal login. This is the single decision that determines everything downstream, and in a legal context it is a legal decision rather than an IT preference.
  2. Which documents never go in, on any account. Anything under a counterparty confidentiality obligation that names the counterparty. Anything containing employee health or compensation data. Anything received from outside counsel.
  3. What gets stripped first. Names, dollar figures, dates, in that order.
  4. What happens to the transcript. Chat histories are company records. Treat them as discoverable, because a court may.

The privilege analysis underneath question one is genuinely unsettled and it deserves more than twenty minutes of a workshop — we treated it at length in the in-house privilege and confidentiality piece, and the policy-versus-practice gap more generally in our safe-use playbook. What the session owes the room is not the analysis. It is a page the team wrote and will therefore follow.

Verification when nobody is filing anything

Most legal AI curricula teach citation checking, because most are written for litigators and the teaching artifact is a sanctions case about fabricated case law in a filing. A corporate legal function at a mid-market company files almost nothing, and its verification problem is less dramatic.

The three checks that matter in-house:

  • Did the clause it quoted actually appear in the document? Open the agreement and search the text. Assistants paraphrase, then attribute the paraphrase.
  • Did it miss a clause that exists? Harder, and the reason the exercise asks for a comparison against a known company position rather than an open-ended summary. Absence is what gets missed.
  • Is the obligation it described the one that binds us, or the one that binds them? Direction errors are the most common failure in contract summarization and the easiest to catch once someone has seen one.

Fifteen minutes on those three checks, worked on the room’s own documents, outlasts any amount of instruction on how language models fail in general.

The two artifacts that have to exist at the end

A session that produces no checkable output was a presentation. Two things have to be on paper when people leave:

A prompt sheet of three to five entries. Each entry names the document type, the task, and the instruction that worked, in the team’s own phrasing. Three good entries beat forty.

The never-paste page. One side of paper, written in the room, with the account decision at the top.

Both are deliberately small. The test is not completeness but whether someone who missed the session can use them a week later without being walked through them.

What the catalog buys instead

The catalog tier is sometimes the right purchase, so be precise about what it does and does not do.

Option What it is Where it stops
ACC Northeast’s on-demand in-house series Free, roughly 50 minutes, vendor-led, not eligible for continuing-education credit A recording. No one in it has seen your paper
Cohort courses on Maven Eight weeks, one 90-minute live session a week plus two to three hours of labs; one leading option is aimed at paralegals and legal coordinators Strong for an individual building a skill; it is not a shared team standard
Coursera’s AI for Legal Professionals and similar catalog courses Low hundreds of dollars, or already inside a subscription you hold Generic documents, generic positions, no artifact that fits your contract workflow

The honest rule: buy the catalog for one person who wants depth, and buy a led session when you need the whole function working the same way. They solve different problems and the failure is treating one as a cheaper version of the other. The general version of that decision is in our enablement manifesto.

What a session like this costs

A led session for a single function runs roughly $2,000 to $15,000 at market rates, with the spread driven by whether the facilitator prepares against your own paper or runs a standard deck. One publicly listed ticket for a four-hour small-group legal AI session runs $1,000 a seat, which crosses the lower end of that range at six people and exceeds it at ten.

Then add the line most budgets omit: staff time. Six people for three hours plus an hour of pre-work is 24 hours of legal and business time, and for a legal function that is the larger number. It is also the strongest argument for the pre-work, because a session run on generic documents spends the same 24 hours and returns less.

What to check thirty days later

Usage dashboards will tell you logins went up. That is not the question. Three checks answerable in a month:

  1. Does the prompt sheet have entries nobody wrote in the room? If the team added two, the habit took.
  2. Has one repeat task moved? Pick the inbound triage job and ask how long the queue takes now against the week before.
  3. Has anyone asked an account question? The first time someone asks whether a specific document can go into a specific tool, the never-paste page is doing its job.

Against that, the adoption numbers are encouraging and not useful. The same March 2026 report put generative AI use within legal teams at 87% of general counsel, nearly double the 44% reported a year earlier, while 53% had a formal technology roadmap — an all-time high in that survey, and still only half. The gap between those two numbers is what a session is for.

Frequently asked questions

Three hours for a function of two to eight people, with at least 85 of those minutes spent on participants’ own live documents. Ninety minutes works for a single task. A full day is usually two sessions that should be separated by a week.

Two things: redact six to ten real agreements the team handles, and write down the specific recurring tasks each person wants to move. Without both, the facilitator runs generic examples and the session becomes a course you could have streamed.

If the legal function is two or three people, yes. Add the three or four business people who send legal the most paper. They generate most of the contract volume and most of the confidentiality exposure.

Can we use ChatGPT or Claude on real contracts in the workshop?

On an enterprise tenant with training on your inputs excluded, and with the documents redacted for the exercises. On a personal account, no. Settle that question before the invitations go out, not in the room.

Does an AI workshop count toward continuing-education credit?

Usually not. Most hands-on sessions are not accredited, and the free in-house series the Association of Corporate Counsel offers states plainly that its recording is not eligible. If credit matters, buy an accredited course separately and do not expect it to change Monday.

What does ABA Formal Opinion 512 require us to train on?

The opinion, issued July 29, 2024, applies existing duties — competence, confidentiality, communication, candor, supervision, and fees — to generative AI, and says a lawyer cannot hand professional judgment to a tool. It is persuasive rather than binding, and several state bars have since gone further on informed consent. A session should cover the duties as they bite on your workflow, not recite the opinion.

How many people should be in the room?

Six to twelve. Below six you lose the compare-and-correct block that makes failure modes visible. Above twelve the hands-on minutes quietly become a lecture, whatever the agenda says.

Roughly $2,000 to $15,000 for a led session at market rates, against $1,000 a seat on one publicly listed four-hour group session and low hundreds for self-paced courses. Add staff time, which for a legal function is usually the bigger figure.

How do we know whether it worked?

Check whether the prompt sheet grew, whether one named task got faster, and whether anyone has asked a new question about which account a document may go into. Login counts answer none of those.

Not for that person alone. Buy them a cohort course, and spend the session budget on the business functions sending them the most work. A shared standard needs at least two people to be shared.

Key takeaways

  • Every result on page one is a sign-up page, a syllabus or a service page. None states a time budget, the one number that predicts whether the session changes anything.
  • Three hours with a legal function of two to eight people, with 85 minutes on participants’ own live agreements, is the shape that holds.
  • The pre-work carries the session: six to ten redacted agreements and a named task list, collected from the team rather than from the buyer.
  • The contract volume at a mid-market company is generated mostly by non-lawyers, so a legal-only room trains the people least likely to create the exposure.
  • In-house verification is not citation checking. It is whether the quoted clause exists, whether a clause was missed, and whether the obligation binds you or them.
  • Two artifacts have to exist at the end: a three-to-five entry prompt sheet and a one-page never-paste list, both written in the room.
  • Catalog courses build one person; a led session builds a shared standard. Buying one expecting the other is the common mistake.
  • Adoption is already high: 87% of general counsel report generative AI use in their teams, against 53% with a formal technology roadmap. The gap is the work.

If you want a read on which of these blocks your legal function needs, book a free AI readiness call. We will walk your contract workflow and tell you where the three hours should go.

Last Updated: Nov 15, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles