Minimising data before using AI: what should you remove from a document?
Prepare a limited copy for an AI task: remove unnecessary data, preserve meaning, check attachments and review the result.
Article language: English

In this article
To improve the layout of a letter, AI usually does not need a client's phone number, signature or the entire history of attachments. Before sharing material, define the task, keep the necessary context and remove information that does not help produce the result. This is data minimisation. Less does not mean worse preparation: sometimes it simply means that an entire company binder does not leave the office to edit one paragraph.
This is a preparation procedure, not automatic confirmation of anonymity or compliance for the whole process. A document without a name may still identify someone or disclose a trade secret. Before using it in a tool, you need to know company rules, the permitted data scope and the terms of the specific service. The method below helps prepare a more limited working copy and check its usefulness.
How do you start with the task instead of the file?
Write one sentence: what exactly should AI do? “Improve the language of the letter” requires different material from “compare the parties' obligations” or “find inconsistent dates”. Without that sentence it is hard to decide what is unnecessary. Every fragment looks potentially useful, so the whole document and its extras reach the tool.
Next, list the facts needed to perform the task correctly. Editing may require only tone, audience and the meaning of a paragraph. Comparing deadlines requires the sequence of events, but not necessarily their real dates. In analysing a specific contract, actual values may matter; do not change them without understanding the effect on the result.
Also establish which materials may be processed in the chosen environment at all. The AI use policy adopted by your company should help make that decision before a copy is prepared. Removing a few fields does not replace the required assessment of the tool or automatically make a prohibited use permissible. The guide to a company AI use policy (in Polish) can help when preparing these rules.
Consider three routes: share limited material, use a synthetic example, or work in an environment approved for that data category. If none satisfies company conditions, stop the transfer and ask the appropriate owner to decide. Do not try to “blur it a little more” without a defined criterion.
How is removing a name different from anonymity?
Masking replaces or conceals selected elements. Pseudonymisation processes personal data so that it cannot be attributed to a particular person without additional information. That additional information is kept separately and protected by technical and organisational measures against attribution to an identified or identifiable person. CNIL explains that pseudonymised data can still be personal data, and identification may be possible by combining characteristics. The label “Client A” is therefore not proof that a document is anonymous.
In a practical check, look for direct identifiers and distinctive combinations. A rare job, a particular event, a small town and an exact date may be enough for someone who knows the context. A national identification number need not be published for material to reveal who is involved. Also check project names and unique descriptions of business clients.
Distinguish protecting people from protecting a company's interests. A document may concern no identifiable person and still contain a margin, tender strategy or unpublished specification. Review both categories in light of the task. A personal-data filter alone does not answer every confidentiality question.
If the company needs a formal anonymisation assessment of a dataset, it should ask the appropriate people to assess the specific use. This guide does not make that classification. Its outcome is a deliberately limited working copy, a description of changes made and a list of information that still must not be shared without separate approval.
How do you prepare a copy that keeps its meaning?
Work in a separate file and keep the source in an approved location. First remove whole sections unnecessary for the task: a footer, earlier correspondence, recipient lists and attachments unrelated to the question. Replace individual details only after that. Narrowing the scope often works better than painstakingly searching for names across dozens of pages.
Use consistent role labels, such as “Supplier A” and “Recipient B”. If one party appears in several places, use the same label each time. Do not change it randomly in every paragraph: the model might treat one party as several. Preserving relationships can matter more to the analysis than preserving the original name. For test sets, extend the check to consistent relationships and event sequences in synthetic data (in Polish).
Change values only with the task in mind. If you are evaluating the order of deadlines, relative days may work if you preserve the intervals. If you are checking proportions, do not replace every amount with the same number. When an exact value is necessary, record that requirement and choose an appropriate processing route. A result on fictional data will not resolve the entire real case in that situation.
Add a short note to the copy explaining what changed and what AI should not infer. For example: dates shifted, roles preserved, amounts replaced proportionally. Do not include a key that reconstructs real data if the task does not need it. Decide separately who may open the replacement map.
What should you check beyond visible text?
A document may contain comments, revision history, hidden sheets, speaker notes or author information. The relevant details depend on the format and program. Make a checklist for your file type and inspect the actual export you will send. A print view does not necessarily show everything available to the receiving tool.

In a file with images, inspect captions, seals, codes and background details. A screenshot can reveal open tabs, a folder name or a side panel about another matter. If a transcribed paragraph is enough for the task, consider sharing only the text. That is simpler than hiding many unnecessary elements in an entire interface.
Check file names and attachment descriptions too. A name such as “case surname contract number” carries information outside the document body. The AI data-flow register maintained by your company should account for file metadata because it can reach task history or logs. Preparing a copy should cover what will actually be sent. The guide to recording AI data flows (in Polish) can help organise the register.
Do not assume a drawn rectangle permanently removes the text underneath. The outcome depends on method and format. Check your output file in an approved tool: read it, search it and inspect the elements meant to disappear. If you cannot confirm the result, choose a simpler form or ask a competent person to prepare a shareable version. If a model detects and removes content, also follow the acceptance process for automated document redaction (in Polish), from the detected passage to the checked output copy.
What might a hypothetical complaint letter look like?
Suppose a team wants to improve the readability of a complaint response. The original contains client details, a device address, order numbers and earlier messages. The task is to structure the response and explain the next step. Full identification and the history of other purchases are not needed for that purpose.
An employee prepares a copy: they retain the problem description, the company's decision and the proposed action. They replace roles, remove contact details and unrelated threads. The deadline is expressed as a number of days because the task is not assessing its correctness in the specific case. A second person checks whether a client name remains in a heading or file name.
AI proposes a better-organised text but changes the cautious “we can offer” into the binding “we guarantee”. That is a meaning error, independent of data minimisation. The response owner compares the result with the approved decision and corrects the wording. Without that check, a more safely prepared input could still lead to an inappropriate business response.
Before sending, the employee adds the actual details and correct deadline in the approved system. The working copy has a planned disposition, and the responsible person accepts the result. This hypothetical example shows the full flow: limit the input, check the result's meaning and deliberately incorporate it into the real case.
How do you check the copy and the answer?
For more consequential material, arrange a short two-person check. One person assesses data reduction; the other checks whether information necessary for the task was removed. For simple material, roles may be combined under company rules, but both questions should remain. Information protection and analysis quality need parallel attention.
Use several artificial examples in team training on AI verification (in Polish). One can contain a name in a comment, another a distinctive description of a person, and a third numbers changed too aggressively. The participant should identify the problem and explain how it affects the decision to use the material. Finding a word alone is not enough.
After preparing the answer, compare it with the original task. Did the model do the work or add missing facts? If it treated fictional values as real, label or reject the result as appropriate. A result cannot settle issues that deliberately changed data do not allow it to assess.
If prohibited material was shared, follow the organisation's AI data incident response procedure. Preserve the facts needed to assess the scope and notify the appropriate person. Sending a corrected copy does not automatically undo the earlier transfer, so do not treat it as closure. The AI data incident guide (in Polish) can help when preparing your organisation's procedure.
Before adopting a template, measure on several artificial examples how long it takes to prepare the copy and check the result. If the procedure is harder than doing the task itself, the team may start bypassing it. Narrow its application or choose another workflow. Acceptance should confirm both data reduction and whether the rule is practical at the team's everyday pace.
For document translation, align minimisation with checks for AI-assisted translation (in Polish). Replacing a name or number should not change the task's meaning. Record which elements must return to the final version and who checks their correct restoration.
How do you maintain the procedure without excessive work?
Create a few templates for typical tasks instead of expecting a complete analysis of any document each time. Separate templates for editing a paragraph, comparing terms and analysing a table help staff choose the right scope. They must also say when to stop if material falls outside the approved example.
- Record the task before selecting document fragments.
- Send only necessary elements of approved material to the tool.
- Keep role labels consistent and preserve relationships.
- Check comments, attachments, file name and other relevant format elements.
- Ensure replacements did not destroy information needed for analysis.
- Compare the result with the task and have the appropriate person accept it.
- Decide how the working copy will be stored or deleted.
Include copy-handling rules in the AI data retention plan for your company. To develop a procedure suited to your own documents, see Aeneia Audit. The guide to retaining and deleting AI data (in Polish) can help prepare this plan.
Aeneia Audit: Start with a synthetic example and a description of the task for which the team wants to use AI. We can agree on the scope for preparing the copy and checking its usefulness. Plan an AI data preparation procedure.










