Ask about AI workshop sandbox data and you will be quoted an infrastructure project. The results are cloud lab environments, containerized notebooks, and legal commentary on supervised testing regimes for high-risk systems. All three are real products. None is what a twelve-person fluency session needs, and reaching for one is how a session that could run in two weeks slips a quarter.
The question a sponsor is actually asking is narrower: which of our own documents may twelve people open, in a room, on Thursday, and who signs off. Answering it takes four controls, all settable with licenses most companies already pay for. None requires a build.
Three products answer to one word
Sort the search results and they split into three unrelated purchases.
| What “sandbox” means | What it is | Right for a fluency session? |
|---|---|---|
| Provisioned compute | A pre-built lab. Pluralsight’s AI Sandboxes bundle a prompt sandbox, live AWS, Azure and Google Cloud sessions, and a SageMaker Studio notebook | No. Built for developers learning model work |
| A supervised testing regime | The EU AI Act’s regulatory sandbox. Article 57 requires Member States to stand up at least one, with priority access for small firms | No. It governs providers putting high-risk systems on the EU market |
| A synthetic dataset | Invented documents for an invented company | Rarely. It is the fallback, not the plan |
The compute reading dominates search because vendors sell it. The legal reading dominates what retrieval systems cite, because statutory text is well indexed and workshop practice is not.
A session on ChatGPT, Claude, Gemini, or Microsoft Copilot has no compute to provision. Participants use a browser and an account. The only thing needing a boundary is the material they bring, which makes this a clearance problem rather than an engineering one. Our playbook for designing a session that changes Monday morning treats the exercises as the product; this is the paperwork that lets those exercises use real work.
The license you already bought is most of the control
The objection that stalls these sessions is nearly always the same: if people paste our documents into an assistant, does the vendor train on them? On business plans for the four tools a workshop would use, the published answer is no, by default.
OpenAI’s business data policy states that by default it does not use data from ChatGPT Enterprise, Business, Edu, or the API platform, inputs or outputs, for training or improving its models, and that qualifying organizations can configure retention. Anthropic’s privacy center says of its commercial products: “By default, we will not use your inputs or outputs from our commercial products.” Microsoft’s current Copilot documentation is blunter: “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs.” Google’s Workspace privacy hub states that “Workspace does not use customer data for training models without customer’s prior permission or instruction.”
Read the terms for your own plan rather than taking that on faith, because consumer tiers differ and tiers get renamed. The point holds: the control most sponsors assume they must build has already been bought. What has not been bought is discipline about which account people use, which is the same gap our case for building capability rather than buying seats describes from the other end.
The hazard nobody warns you about: inherited permissions
Here is the part no workshop vendor puts in a proposal. Copilot and Gemini answer from what the person asking can already open. Microsoft states it plainly: Copilot “only surfaces organizational data to which individual users have at least view permissions.” Google’s version: “If the user doesn’t have access to a document or email, Gemini will not retrieve that content.”
Both companies present this as the safety guarantee, and it is one. It is also the real hazard of a live session, because an exercise inherits every permission mistake in your file store. Put twelve people in a room, give them a broad retrieval prompt about a project or a department, and you have run an unplanned access audit with an audience. The folder shared company-wide in 2023 surfaces. So does the compensation sheet somebody dropped in a team drive. Nothing left the company; something worse happened socially, because a person found out in front of peers what they could see all along.
Two cheap moves prevent it. Scope retrieval exercises to named sources rather than open questions, so people practice on a document they chose to bring. And have whoever owns your file store run a sharing report on the folders in scope a week out. That report is worth having anyway, and a workshop is a decent excuse to pull it.
Four controls that make a room safe
This is the whole sandbox: four controls, one named owner each, and a sign-off that fits in a single meeting.
| Control | The decision | Who owns it |
|---|---|---|
| Account boundary | Every exercise runs in a company-administered workspace on the paid plan. No personal logins, no free tiers | IT, or whoever administers the licenses |
| Permission boundary | Exercises reference sources participants name, not open retrieval. A sharing report runs beforehand | The file-store owner |
| Exclusion list | One page naming what may not enter the room, circulated with the pre-work | The sponsor, with legal sign-off where that exists |
| Disposal rule | What happens to chat history, exported artifacts, and the shared prompt document | The sponsor |
Say that in a clearance meeting and the conversation usually ends in twenty minutes, because every item is a setting or a sentence rather than a project. The statutory version shows the contrast: Article 59 of the EU AI Act permits personal data in a regulatory sandbox only where the goal “cannot effectively be fulfilled by processing anonymised, synthetic or other non-personal data,” and only inside “a functionally separate, isolated and protected data processing environment.” A session where people summarize their own meeting notes does not need to pay that price.
The exclusion list fits on one page
Four categories cover nearly everything, and the list works better when it is short enough to read in the room.
- Personal data about people. Employee records, candidate materials, customer or patient identifiers.
- Anything under a confidentiality clause you did not write. Client material under an agreement, third-party data under a license, someone else’s code.
- Credentials and keys. Obvious, routinely forgotten, worth saying twice.
- Live financial and deal material. Unreleased results, active negotiations.
Then the part most lists omit: say what happens when somebody brings an excluded document anyway. Someone will, usually by accident, and usually the most eager person in the room. The facilitator swaps them onto a neighbor’s artifact, says why in one sentence, and moves on. Handled that way it costs thirty seconds and teaches the rule better than the slide did. Handled badly it ends that person’s afternoon, which is the failure mode our notes on facilitating non-technical rooms cover in more depth.
Give the list a named owner who can answer a question by text during the session, because an unowned list stops the room at every edge case.
Redaction is the costliest control, so put it last
Standard advice is to strip identifiers from documents beforehand. It sounds prudent and it is usually the wrong first move, for two reasons.
Cost first. Real de-identification is a defined standard, not a pass with a marker. HIPAA’s safe-harbor method at 45 CFR 164.514(b)(2) requires removing eighteen categories of identifier, covering not only the individual but relatives, employers, and household members, and it still fails if you know the remainder can re-identify someone. Teams that try this by hand discover the hours after they have committed to the date.
Then the lost value. A session on real work teaches judgment: whether a task is worth handing over, which documents it needs, and how to catch the confident wrong number in the output. Strip a document to a skeleton and the exercise becomes a prompt-writing drill, which is the trap our argument for training on last week’s real work takes apart.
Sort material by sensitivity first and spend redaction effort only where the calendar forces it.
| Tier | Example | Clearance effort |
|---|---|---|
| Low-sensitivity internal | Weekly report, meeting notes, draft job posting, vendor comparison | None beyond the exclusion list |
| Internal but named | Performance notes, customer correspondence, pipeline detail | Strip names and account numbers, or substitute placeholders |
| Regulated or client-bound | Health records, financial account data, material under a client agreement | Full de-identification, or build a representative document |
Most teams can fill a full day from the first tier without touching the second. The weekly report nobody enjoys writing is both real and dull, which is exactly what a first session wants.
What outlives the session
Clearance meetings cover the inputs and forget the outputs. Four things need an owner.
Chat history sits in the workspace. Admins can see it, retention policies apply, and on most plans participants can delete their own. Decide before the session, not after a hallway question. Exported artifacts are now files, and if the exercise used client-bound material they carry the same obligations as the source. The shared prompt document is the most valuable thing the day produces and the most likely to end up in somebody’s personal notes app; put it in a company location in hour one, as the pre-work guide asks a sponsor to prepare. Facilitator copies belong in writing: ask what an outside facilitator keeps and for how long, and expect nothing beyond anonymized exercise patterns.
The two-week clearance sequence
Two weeks is enough for a single-company session on a tool you already license. The order matters more than the duration.
- Day 1. Confirm the tool and that every participant has a company account on the paid plan. Request missing seats now, because provisioning is the item that actually slips.
- Days 2 to 3. Draft the one-page exclusion list and send it to legal or the compliance owner with a deadline, asking for edits rather than approval, which gets a faster reply.
- Day 4. File-store owner runs a sharing report on the folders in scope. Fix anything alarming; note the rest.
- Day 5. Pre-work request goes out with the exclusion list attached, naming what participants should bring.
- Days 8 to 10. Facilitator builds exercises from what arrived. Expect a third of the room to bring nothing, and expect that to correlate with who needs the day most, which is why the exercise library carries ready-made substitutes.
- Day 11. Sponsor sets the disposal rule and tells participants what it is.
- Day 14. Session runs, exclusion list read aloud in the first five minutes.
Hand that sequence to whoever owns the calendar rather than to the facilitator. Every item on it is an internal one.
When you genuinely need a built sandbox
Three cases justify the cost, and spotting them early saves an argument.
Regulated records with no safe substitute. If the only work worth practicing on is health data, financial account data, or records under a government handling standard, you need a separated environment with controlled access, and that is a project with a procurement cycle in front of it.
More than one company in the room. Consortium sessions, client-facing demonstrations, and public courses cannot use anyone’s live material. Synthetic documents are correct here, and the transfer cost is the price of the format.
No tool policy yet. If nobody has decided which assistant is sanctioned, do not solve that inside a workshop. Write a one-page interim rule covering the four exclusion categories and run on one named tool.
Outside those three, a built sandbox is a way of looking careful while the behavior you wanted stays unchanged. The teams that get a return on AI enablement clear low-sensitivity real work in two weeks and spend the saved quarter running the next cohort. For a second opinion on where your own material falls, a free AI-readiness call will sort your document tiers and name the items that genuinely need legal review.
Frequently asked questions
What does “sandbox” mean for an AI workshop?
Four controls, not an environment: a company-administered account for every participant, exercises scoped to named sources instead of open retrieval, a one-page list of material that may not enter the room, and a stated rule for chat history and artifacts afterward.
Will the vendor train its models on documents we use in the session?
Not by default on business plans. OpenAI excludes ChatGPT Enterprise, Business, Edu, and API inputs and outputs from training by default; Anthropic says the same of its commercial products; Microsoft says Copilot prompts, responses, and Graph data are not used to train foundation models; Google says Workspace does not use customer data for model training without permission. Check the terms for your own plan, since consumer tiers differ.
Do we need synthetic data for a hands-on AI session?
Only for mixed-company rooms, public courses, or regulated records with no substitute. For a single-company session on a sanctioned tool, low-sensitivity internal documents from last week are both safe and far better for transfer, because the output is usable work and the failures are your own.
Can participants use personal accounts if the company has not bought licenses yet?
No. The account boundary is the one control with no workaround, because consumer tiers carry different training and retention terms and your admins cannot govern that workspace. Buy the seats for the session, or run it on invented documents.
Who has to approve the data used in a workshop?
The sponsor owns the exclusion list, legal or the compliance owner reviews it where that function exists, the file-store owner runs the sharing report, and IT confirms the accounts. One meeting, four names. Assign each item to a person rather than a department.
How long does it take to clear data for an AI workshop?
Two weeks for a single-company session on a tool you already license. The slow items are seat provisioning and legal review, both of which move faster when you ask for edits to a draft list rather than approval of a blank question.
Is it safe to let an assistant read our shared drive during an exercise?
Safe from a leakage standpoint, awkward from a social one. Enterprise assistants return only what the person asking can already open, so a broad retrieval prompt reveals your own oversharing to a room of colleagues. Scope exercises to named sources and pull a sharing report beforehand.
Should we redact documents before the session?
Redact only the tier that needs it. HIPAA’s safe-harbor method requires removing eighteen categories of identifier covering relatives and employers as well as the individual, so hand-redacting a day’s worth of material costs hours and strips out the judgment practice you were buying.
What happens to the prompts and documents after the workshop?
Whatever you decided beforehand. Chat history stays in the workspace under your retention settings, exported drafts carry the same obligations as their sources, and the shared prompt list belongs in a company location from the first hour.
Does the EU AI Act apply to our internal training session?
Not to a US company running a fluency session for its own staff. The AI Act’s sandbox provisions, Articles 57 and 59, govern providers testing high-risk systems before putting them on the EU market. They are a useful benchmark for what a real isolated environment demands.
Key takeaways
- Three unrelated products answer to “sandbox”: provisioned compute, the EU’s statutory testing regime, and synthetic datasets. A fluency session needs none of them.
- Business plans on the four mainstream assistants already exclude your inputs and outputs from model training by default, so the control most sponsors expect to build is already paid for.
- The real hazard is inherited permissions. Assistants return what the person asking can already open, so a broad retrieval exercise doubles as an access audit in front of an audience.
- Four controls cover it: account boundary, permission boundary, a one-page exclusion list with a named owner, and a disposal rule.
- Redaction is the costliest control and belongs last. Sort material by sensitivity and fill the day from the low-sensitivity tier.
- Two weeks is enough. Seat provisioning and legal review are the items that slip, so start both on day one.
Dirk Jan van Veen, PhD