AI Privacy

A middle-aged woman in glasses on a chain and a burgundy blouse stands by a window holding a cyan spiral-bound folder. With a pen she ticks one of the boxes on a slip attached to the folder.

Documents for an AI training session come from two sources: a public stand-in or an approved working copy. Either way, you prepare the material for a specific exercise: remove what is not needed, keep the logic the task depends on, and test it on an example. If it passes, approve it. If not, revise or replace it.

What documents are safe to bring to an AI training session

A public stand-in is an open document whose structure matches your real task - an expense table, a leave policy, a request template. It works when the underlying logic is transparent and does not depend on closed formulas or internal rules.

An approved working copy is your actual document with the unnecessary parts removed. The material owner decides what stays. IT checks that no hidden traces remain after cleanup. System actions are rehearsed in a test environment - no sends, no changes to live data.

Public does not mean automatically safe. A website may carry third-party personal data, someone else's copyright, or outdated terms. Take the specific facts you need, rewrite them in your own words, and cite the source.

How to prepare a card for the exercise

This is a training example, not client data. Source: GSA (General Services Administration - the US federal procurement agency), M&IE breakdowns, updated 31 July 2026.

M&IE (Meals and Incidental Expenses) is the per-diem allowance for food and small expenses in US federal travel. Rate for an intermediate full day within CONUS (the continental United States - 48 states and the District of Columbia): 68 USD. The rate breaks down into breakfast 16, lunch 19, dinner 28, incidentals 5. Government-provided meals subtract from the rate. A free hotel breakfast and carrier meals do not reduce it.

From the card, remove other rate rows, site navigation, and first-and-last-day rules. What remains is one condition: intermediate full day, rate already given.

Exercise:

A. Breakfast was provided by the government. How much of the daily allowance remains?

B. The hotel fed the employee breakfast for free. How much remains?

Correct answers: A - 52 USD, B - 68 USD. Without the right exception, both answers come out as 52. The difference is who provided the breakfast, not whether it cost money. Check: 16 + 19 + 28 + 5 = 68.

A hand holds a cyan roller stamp; a magenta pattern covers the first column with grey lines visible beneath. A pencil, ruler, and cap sit nearby.

How to check that the material is ready for the session

Before the session, the organiser answers two questions.

  1. Does the exercise train exactly the skill you promised to teach? In this example - the ability to distinguish deduction conditions in rate calculations.
  2. Does the material contain every dependency and exception needed for a correct answer? In this example - both breakfast variants and the explicit rule wording.

Two "yes" answers mean the design is sound. But before approving the card, you still need to: verify the values against the source, check the formulas, run a trial exercise in an approved environment, and record the decision - approve, revise, or replace - with a reason.

Approval for one specific exercise does not mean the material may go to any service. The learning effect is a separate check.

What to keep and what to cut in texts, tables, and correspondence

Format Keep Change
Text Argument structure, step order, if-then conditions Real rules and dates stay only under an approved transfer; otherwise build a synthetic stand-in or replace the material
Table Columns, linking formulas, condition ranges Row identifiers
Correspondence Sequence, intervals, and deadlines Replace recurring names and references under an agreed approach; real dates stay only under an approved transfer; otherwise use different material

Three roles: the employee prepares a draft; the material owner checks that the working logic survived and no confidential information remains; IT verifies the approved environment, recipients, storage, and access. Hidden content is checked on a copy, not the original - and after the check it is not automatically clean.

A desktop tower and a monitor with a yellow frame and an empty grey screen. A network cable lies coiled on the desk, its free plug connected to nothing. A hand ticks fields on a form standing on a holder in front of the monitor.

Why cleaning a document does not guarantee safety

Hidden data is information inside a file that is invisible during normal viewing.

Microsoft Document Inspector is the built-in Word tool for finding hidden information (guide). It runs on a copy - removal can be irreversible. The tool name does not guarantee cleanliness: it will not find white text on a white background or objects covered by other objects.

NIST SP 800-188 (September 2023, full text) is a voluntary framework for US federal agencies. It states that removing identifiers does not eliminate all risks, and that over-removal makes data unusable for the task.

The ICO (Information Commissioner's Office - the UK data regulator) anonymisation guidance (ico.org.uk) notes that risk depends on available information, recipients, and environment. Replacing a name is not proof of anonymity. This is UK guidance, not a legal opinion for other jurisdictions.

The right to edit a file and the right to send it are different permissions. Even a clean file must not go to a training environment without a separate approval for the transfer. Assess what the file carries in your company's context and get permission separately.

What to agree on before the session starts

Before the session, agree on five things: the material, the tool, access, permitted actions, and the verifiable outcome. Rehearse system actions on a test copy - no sends, no changes to live data. If permission is not in place, swap the material or move the exercise.

Questions and answers

Can I just replace all numbers with random ones and send the file?

No. If the document structure itself is sensitive, neutral values do not remove the risk. Two columns in a table combined may identify a specific person even after the numbers are replaced. The owner needs to review the whole document and decide what to keep, what to replace, and what to remove. Then IT checks for hidden traces. Only then is the material ready to transfer.

Can I load a publicly available file into an AI tool without checking?

Before uploading a file from someone else's site to an AI tool, check: does it contain third-party personal data, what are the terms of use, and is the material still current? A public URL is not a licence for the whole site. Take the specific facts, rewrite them in your own words, and cite the source. Paraphrasing does not remove copyright restrictions - confirm that your purpose is compatible with the licence.

What if the needed skill cannot be learned on public material?

If the skill requires closed formulas or internal rules, public material will not work. You then have three options: get an approved copy with the owner's consent, find a public example closer in structure to the actual task, or move the exercise to a session that uses a working copy in a test environment. The right choice depends on the task and what the owner is willing to approve.

Do I need permission from the material owner before sending it to the training environment?

For an approved copy, we recommend getting the owner's confirmation for the specific material and environment in line with your company policy. For a public stand-in adapted in your own words, the transfer check still applies - confirm there is no personal data and that the terms of use allow the transfer to the chosen environment. Paraphrasing does not remove those restrictions.

What is an approved synthetic stand-in?

An approved synthetic stand-in is a training document created specifically when a real document cannot be safely prepared for the session. It reproduces the logical dependencies without real data. It must not be presented as a real document or as anonymised data. The organiser creates it together with the material owner - the owner checks that the needed logic is preserved. This is the standard fallback when public material does not fit the required structure and an approved copy cannot be prepared.

How do I tell whether a table's structure is itself sensitive?

Ask yourself: if you combine two columns from the table, can you identify a specific person? If yes, replacing the values is not enough - the owner needs to assess the document as a whole. The owner evaluates which combinations carry risk in your company's context. Based on that, the decision is: what to keep, what to replace with placeholder values, and what to turn into a synthetic stand-in. Replacing numbers on your own, without that assessment, does not remove the risk.

Conclusion

A safe document for an AI session is material prepared for the specific exercise, verified, and approved before the session starts. The employee, the material owner, and IT do this together.

If you want to sort through your own documents and decide what can go to a session, write to t.me/shimaoz or hello@majento.ai. Also useful on this topic: AI team training and how to prepare a team for AI work.

Sources

  1. nvlpubs.nist.gov publication
  2. ico.org.uk publication
  3. gsa.gov publication
  4. support.microsoft.com publication

View as Markdown