Bokili
Responsible AI Practice3 min read

Build a Prompt Data Boundary Before Work Enters AI

Use a four-step prompt data boundary to drop, mask, generalise or route information for approval before an employee sends it to an AI tool.

Bokili Editorial· Verified August 15, 2026
ShareX
Work documents passing through four filters before a clean excerpt reaches an AI tool

The safest prompt is often the one that contains less information. Before an employee pastes a document into an AI tool, they need a small decision process that works in seconds. A long policy may explain what is confidential, but it rarely helps at the exact moment someone is holding a customer email, a contract or a spreadsheet.

A prompt data boundary turns data minimisation into a visible work habit. The ICO describes data minimisation as using personal data that is adequate, relevant and limited to what is necessary for the purpose. CNIL advises organisations to define authorised and prohibited uses and says end users should submit only information they are allowed to share. The boundary below translates those principles into four actions.

The DROP–MASK–GENERALISE–APPROVE boundary

Four actions before the prompt

1

DROP

Remove details the task does not need. If the model can answer without a field, do not include it.

2

MASK

Replace direct identifiers with stable placeholders such as [CUSTOMER], [SUPPLIER A] or [CASE 14].

3

GENERALISE

Reduce precision when exact values are unnecessary: use a range, category or short description.

4

APPROVE

If sensitive or confidential detail remains necessary, use the approved system and route—or ask the named owner before proceeding.

These actions are not interchangeable. Masking a name may still leave a person identifiable from job title, location and event details. Generalising a figure may make a task useless when the exact value drives the decision. Approval is not a magic word either: it should point to a real policy, system or person.

Start with the purpose, not the document

A 60-second boundary check

  1. 1

    1. Write the task in one sentence

    For example: “Rewrite this complaint so the next action is clear.”

  2. 2

    2. Mark the minimum facts

    Highlight only the facts the output genuinely depends on.

  3. 3

    3. Apply the four actions

    Drop irrelevant detail, mask identifiers, generalise precision and route anything still sensitive for approval.

  4. 4

    4. Re-read the clean input

    Ask whether a colleague who sees only this excerpt could still identify a person, client, deal or confidential plan.

  5. 5

    5. Check the destination

    Confirm that the chosen AI tool and account are approved for the remaining information.

Worked example: rewriting a customer complaint

Reading is a start. Practice makes it stick.

Start learning

Imagine a fictional complaint containing a customer’s full name, account number, health detail, exact purchase value, store location and a long history. The task is only to draft a calm acknowledgement and list the next three service actions.

DROP the account number and unrelated history. MASK the name as [CUSTOMER]. GENERALISE the purchase as “a high-value order” if the exact price does not affect the wording. If the health detail is essential for an accessibility response, do not assume masking is enough: use the organisation’s approved route or ask the privacy owner. The final prompt can focus on the service facts without carrying the entire case file.

Boundary questions employees can answer

  • What exact output am I asking the AI to produce?
  • Which facts are essential to that output?
  • Can I remove direct and indirect identifiers?
  • Can I replace exact values with ranges or categories?
  • Does any remaining information require an approved enterprise tool or human permission?
  • Could the task be practised with a synthetic example instead?

What the boundary does—and does not—solve

The boundary reduces unnecessary disclosure. It does not replace your organisation’s data classification, contracts, access controls, retention rules or tool settings. It also does not make every masked prompt anonymous. Treat it as a front-line behaviour that supports those controls.

The rule should be easy to find at the moment of work: next to the approved-tool list, inside a prompt template or in a short practice mission. Managers should test it with realistic examples. A rule that employees cannot apply under time pressure is not yet an operating control.

Try the boundary in ten minutes
  1. Choose a low-risk work document you are allowed to use—or create a synthetic example.
  2. Write the AI task in one sentence.
  3. Make four copies labelled DROP, MASK, GENERALISE and APPROVE.
  4. Mark which details move through each action.
  5. Build a clean prompt from only the minimum approved information.
  6. Ask a colleague to identify one detail you could still remove.

Make safe use a practised reflex

A safe AI practice sandbox lets employees rehearse this boundary before live data is involved. Pair it with a clear workflow stop rule and a method to verify AI output. Bokili turns these behaviours into short workplace missions with feedback, so safety becomes something people do rather than a policy they once read.

Sources

  1. Principle (c): Data minimisation — Information Commissioner's Office
  2. CNIL's Q&A on the Use of Generative AI Systems — CNIL
  3. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST
ShareX

Reading is a start. Practice makes it stick.

Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.

Start learning

Keep reading

All articles