Search for an AI workshop with company data and you get two unrelated products wearing one name: employee sessions that promise to be hands-on, and data-readiness assessments and retrieval builds sold by platform vendors to IT. A buyer looking for a one-day fluency session lands on an engineering answer, concludes the real-work version is a six-month infrastructure project, and books the generic session instead.
It is not an infrastructure project. Running a session on your team’s own material takes three things: documents the team handled last week, accounts your company controls, and a written list of what may not enter the room. No model gets retrained and nothing gets indexed. The difference between that session and one built on sample prompts is not technical, and it decides whether anything changes afterward.
The phrase covers two different purchases
Sort the search results and they split cleanly. On one side sit workshop pages: a full day, five modules, a flat rate for twenty seats. On the other sit assessment engagements. Informatica, Citrin Cooperman and a row of Microsoft partners all publish an AI Data Readiness Workshop, which is a scoping exercise for a platform sale. Harvard Business Review’s much-cited 2023 piece on training generative AI with company data is about grounding a system, not teaching anyone.
Three budget lines, and only one of them is AI enablement:
| What you are buying | Who runs it | What you own afterward |
|---|---|---|
| Fluency session on real documents | A facilitator, in the room with the team | People who do their work faster, and a ranked list of tasks worth handing over |
| Data readiness assessment | A platform vendor or its partner | A roadmap, usually priced as the first step of a license |
| Retrieval build over internal documents | An engineering team or build partner | A working internal system, and its running cost |
The third is a real project at a real price, roughly $25,000 to $150,000 at market rates, and it may be the right thing to buy next year. It is not a prerequisite for the first, and vendors blur the two because the blur sells assessments. A session that uses your documents uses them the way your team already does: pasted or uploaded into an assistant, during an exercise. That is the whole mechanism, and the gap it closes is the one our case for building capability rather than buying seats describes.
Why a toy exercise teaches a skill nobody uses
Training research has studied this for forty years and the finding is dull and reliable: the closer the practice conditions sit to the work conditions, the more of the training reaches the job. Identical elements theory holds that a program transfers in proportion to what it shares with the real task. Blume, Ford, Baldwin and Huang’s meta-analysis of 89 transfer studies, in the Journal of Management in 2010, adds the moderator that matters here: they separate closed skills, which have one correct procedure, from open skills, where the trainee decides how to apply what they learned, and the drivers of transfer tie more strongly to open skills.
Writing a prompt about a fictional bakery is a closed skill with one right answer and no stakes. Deciding whether the quarterly supplier review is worth handing to an assistant, assembling the four documents it needs, and catching the confident wrong number in the output is an open skill. Only the second is the job, which is the same argument our case for teaching transferable skill rather than tool coverage makes about breadth.
Sample prompts also hide the failures that matter most. An assistant given a sanitized example produces a clean answer, because sanitized examples are clean. Given a real document, it produces the problems your team will meet: the table it misreads, the clause it paraphrases wrongly, the number it invents when the source is silent. Those moments are the curriculum, and nobody who has not seen your documents can manufacture them.
Your team is already doing this without you
The strongest argument for a sanctioned session is that the unsanctioned one is already running. BCG’s AI at Work 2025 survey, covering more than 10,600 workers in 11 countries and published on June 26, 2025, found 72% using AI regularly while only 36% felt adequately trained, 54% saying they would use AI tools without authorization, and only a quarter of frontline workers reporting enough guidance from their leaders.
The data-handling surveys say what that looks like in practice. TELUS Digital, surveying 1,000 employees at large firms in January 2025, found 57% had entered confidential information into an assistant and 68% were reaching those tools through personal accounts; later surveys from PagerDuty and Zapier put the share who have put work information into a public tool at 88% and 89%. All three are vendor-commissioned and ask different questions, so read the range rather than a figure. The question is not whether company material will meet an assistant, but whether the first meeting happens in a room with a facilitator, on a controlled account, under a rule people have read, or on a personal login at 7pm with a client file open.
The three tiers of workshop material
Facilitator guides nearly all give the same advice: redact the document, or build a synthetic one. Sound counsel, with a cost nobody states.
| Tier | What it is | Transfer | Use it when |
|---|---|---|---|
| 1. Synthetic | Invented documents for a fictional company | Low. Produces people who can prompt, not people who changed a workflow | Public sessions, or a tool policy still being written |
| 2. Redacted | Your documents with names, numbers and identifiers stripped | Medium. Structure survives, judgment calls mostly do not | Regulated records, client-identifying material, any room with outside attendees |
| 3. Real and approved | Low-sensitivity internal material from last week, in a controlled account | High. The output is usable work and the failures are your own | Single-company sessions on a sanctioned tool, which is most of them |
Most teams can fill a day from the third tier without touching the second. The weekly report, the meeting notes, the draft job description, the vendor comparison: none of it is sensitive, all of it is real, and it is where the hours go. Sensitivity and usefulness are different axes, and conflating them is how a session ends up on the bakery.
The pre-work request that makes it possible
Vendors say tailored and mean a scoping call about goals. Tailoring that reaches the exercises needs artifacts, requested by name two to three weeks out:
- Two documents they produced last week that took longer than they should have. Not the hardest thing they do, the most repeated thing.
- One document they had to read and summarize for somebody else.
- One task they tried with an assistant that went wrong, with the prompt and the output if they kept it.
- One task they suspect could be handed over but have not tried.
Item three is the most valuable and the least often requested. A set of failed attempts tells a facilitator where the fluency gap sits, and turns the session from instruction into repair. That information is what you are paying a custom curriculum for, as our comparison of custom and off-the-shelf curricula argues.
Three things belong to the sponsor before the request goes out. Name the tool the exercises run in and confirm the team has company accounts on it. Write the refusal list: personal data about employees or customers, anything under a client confidentiality clause, credentials, unreleased financials and live deal material. And name who signs off on what comes back, usually the functional lead.
Plan for the pre-work to arrive incomplete. In a room of twelve, expect three or four people to bring nothing, and expect that to correlate with who needs the session most. The fix is a facilitator who arrives with exercises built from the artifacts that did come in, so those participants work on a colleague’s real document rather than a fictional one.
Same skill, two exercises
The teaching move is identical on both sides. Only the material changes, and with it everything that follows: the right-hand column produces a finished artifact and a decision log by lunch, the left-hand column produces notes.
| Skill | Toy version | Real-work version |
|---|---|---|
| Assembling context | “Write a product launch email” | Paste the real launch brief, the pricing page and last quarter’s email that worked, then ask for a draft |
| Summarizing | Summarize a public article | Summarize the 30-page vendor proposal on the sponsor’s desk, then check it against the source |
| Checking output | Spot the error in a prepared example | Run the team’s own report through, find where the model guessed, name the tell |
| Judging fit | Discuss which tasks suit AI | Sort the team’s ten most repeated tasks into hand over, assist, leave alone |
| Data judgment | Read the policy slide | Hold up four of the team’s own documents and decide which may be pasted where |
Row four doubles as the cheapest discovery any company runs: a ranked list of candidate automations, drawn from the people who do the work.
What the enterprise terms let you do
The safety question deserves an answer rather than a disclaimer. Under business and enterprise agreements, the major assistants do not train their foundation models on what your people type. Microsoft’s enterprise data protection documentation for Microsoft 365 Copilot states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models; OpenAI says the same for its business products, and Google and Anthropic publish equivalent commitments on their paid tiers. Check the wording for your own plan, and confirm which plan the team is on rather than which one the license says you bought.
Two distinctions survive that commitment and belong in the session rather than a footnote. Not training is not the same as not retaining: Microsoft’s documentation is explicit that prompts and responses under enterprise data protection are logged and kept for audit and eDiscovery. And a company account is not the personal account the same person uses at home, where most of the exposure in those shadow-use surveys sits.
For the governance frame, borrow the vocabulary of the NIST AI Risk Management Framework (AI 100-1, released January 26, 2023). It is voluntary guidance, not a certification, and its four functions give the session a shape a sponsor can defend: Govern the rule, Map what is in scope, Measure what came out, Manage what happens next. Thirty minutes is not a compliance program, but it puts the rule in front of people holding a real document.
When sanitized material is the right call
Three cases, because a format that is always right is a sales pitch.
The room is not one company. A public session cannot run on anybody’s real documents, so synthetic material is correct there, at the cost it looks like: the session teaches prompting rather than changing a workflow. That is a fair reason to prefer an in-house session when the goal is behavior.
The work is the regulated kind. Health records, legal matter files, anything under a client confidentiality clause. Redaction is the floor, and some matters stay out of training entirely. Train the same skill on internal operational documents instead.
The policy is not written yet. If nobody can say which tool is sanctioned, running exercises on real documents makes the facilitator the one who decided. A thirty-minute decision by the sponsor moves the session from tier one to tier three.
Real work gives you a measurable before
The last argument is the one a CFO cares about. A session on sample prompts can only be judged by asking whether people enjoyed it. A session built on real work arrives with a baseline attached, because the pre-work artifacts are the baseline: the report that took three hours, the task everyone avoids.
That makes the follow-up concrete. Thirty days later you are not asking whether confidence improved; you are asking whether that report still takes three hours and whether the tasks sorted into hand over have been handed over. BCG puts the threshold for regular use at five or more hours of training, in person and with coaching, so one session is a start rather than a finish, and what sustains it is the subject of our comparison of one-off sessions and ongoing coaching. For the wider design question, start with our workshop design guide.
Frequently asked questions
Is it safe to use real company documents in an AI workshop?
Yes, for low-sensitivity internal material on a company account. The major vendors state in their business terms that customer prompts and responses are not used to train their foundation models; verify the wording for your plan. Control which account the exercises run in, and write down what stays out.
Does training on company data mean fine-tuning a model on it?
No, and the confusion costs companies money. A fluency session puts documents into a chat window during an exercise. Fine-tuning and retrieval systems are engineering projects with their own budget, roughly $25,000 to $150,000 at market rates.
What should never enter the room?
Personal data about employees or customers, anything under a client confidentiality clause, credentials and keys, unreleased financials, and live deal material. Write the list before the pre-work request goes out, and have the facilitator restate it in the first ten minutes.
Do we need a data readiness assessment first?
Not for a fluency session. Those assessments prepare infrastructure for a platform deployment and usually arrive as the first step of a license sale, which is a different purchase with a different buyer. If your team can open a document and paste from it, your data is ready for this.
What happens if people do not send their pre-work?
Expect a third of the room to send nothing. The facilitator should arrive with exercises built from what did come in, so everyone works on genuine company material even if it is not their own. Keep the request short and have the sponsor send it.
Can we do this if the team only has personal AI accounts?
You can run the session, and you should fix the accounts first. Personal accounts sit outside your administrator controls and outside the enterprise data commitments, so the exercise would teach a habit you do not want on a sanctioned tool.
How do we measure whether training on real work changed anything?
Use the pre-work as the baseline. You know what those tasks cost before the session, so at thirty and sixty days you can ask whether they still cost that, and whether the tasks sorted into hand over have moved. Surveys measure the room; the artifacts measure the work.
What does a workshop using our own documents cost?
Market rates for team workshops run roughly $2,000 to $15,000, set by room size, pre-session design, and whether anything follows. Building exercises from your material adds design hours, not a different category of price. A vendor quoting both versions at one price has not built the second.
Key takeaways
- The search phrase hides three purchases: a fluency session, a data readiness assessment, and a retrieval build. Only the first is enablement, and it carries none of the prerequisites the other two imply.
- Identical elements theory and the Blume et al. 2010 meta-analysis of 89 studies both point to practice conditions that match work conditions. The judgment calls on a real document are the open skill that reaches the job.
- Your material is already meeting these tools. BCG found 72% using AI regularly against 36% who feel adequately trained; shadow-use surveys put real-data exposure between 57% and 89%.
- Request artifacts, not goals: two slow documents, one summary job, one failed attempt and one candidate task per person, plus a written refusal list and a named approver.
- Enterprise terms permit this, with two caveats worth teaching: not training is not the same as not retaining, and a personal account is not a company one.
- Synthetic material is correct for mixed-company rooms, regulated records, and any company whose tool policy is still unwritten. Elsewhere it buys safety you already had.
Before you approve the next session, ask the vendor one question: which of our documents will be open on the screen? If the answer is a sample pack, you are buying a seminar. A free AI-readiness call will tell you which weekly tasks the exercises should be built from.
Arthur Wandzel
