Skip to content
Interface language: English
AI course / Signal

Safer data practices when using AI at work

Prepare a minimal fictional complaint note and check data, attachments and AI output before sharing.

Article language: English

A woman holds a sheet of paper beside an ornate red cabinet of drawers.
In this article

You want AI to draft a reply to a complaint. It needs the problem and the service rules. It does not need the customer's card number, home address or system password. Preparation begins before you paste anything into a conversation.

Consider a fictional shop note. Every identifier is invented:

Customer K-18 bought lamp L-7. After two days its switch stopped working.
They request a replacement. A photograph of the damage was provided.
Internal procedure: first check whether the complaint is complete;
the complaints handler decides on a replacement.

You can practise a draft on this material. A real person's name, address and order number add nothing to the language task. Do not paste them in “just in case”: that is precisely the risk you are trying to reduce.

Establish where you are working with data

Before choosing a privacy setting, check whether your account is personal or organisational and whether your organisation permits the tool for this kind of information. A paid subscription alone does not answer that question.

OpenAI's data controls documentation distinguishes turning off “Improve the model for everyone” from conversation history. OpenAI also describes the default exclusion of content in managed Business, Enterprise and Edu workspaces from model training. Organisational settings and policies still apply.

This is limited information about a provider's data use. It does not replace organisational permission to send a document or an assessment of the agreement, access and retention. If you do not know the rules, use fictional material and ask the security or data protection contact about the specific category of information.

Remove indirect identifiers too

Replacing a name with a letter may not be enough. “The only branch manager in a small town, returning after a particular procedure” may still identify a person. For a task about message tone, “a person requesting a schedule change” is enough.

Prepare a separate working copy. Remove identifiers, contact details, trade secrets and details unnecessary for the purpose. Use replacements consistently: customer A must remain customer A throughout. Otherwise the model may join two cases and leave you to untangle them.

Do not automatically call the result anonymous. The European Commission explains that information allowing a person to be reidentified remains personal data. A table linking labels to people needs protection. Here you are reducing what is shared; leave assessment of a particular dataset to the person responsible for data protection.

OWASP identifies sensitive information disclosure as an AI application risk, including personal, financial and confidential business information. Adding “do not disclose” to a prompt should not be your only safeguard.

Try a short invented example. For a task about response tone, someone prepared: “Customer K-18, phone [NUMBER], address [ADDRESS], lamp L-7, broken switch. Photograph provided. Purchase connected with furnishing a room after treatment.” Brackets contain fictional placeholders; do not insert real data into the exercise.

Assess which fields in the fictional note are needed solely to draft an
acknowledgement of the lamp complaint. For each field say: needed / unnecessary
for this task / requires a decision. Explain the link to the message's purpose.
Prepare a minimal note without contact details, address or treatment information.
Do not assess whether the dataset meets legal conditions for anonymisation.

A model can help identify excess fields, but check the copy yourself. The treatment information does not help acknowledge a broken switch. In a real case, even sending it to a tool would require prior assessment. That is why this test uses invented identifiers, and the decision about company material comes before submission.

If you want a starting point without similar dilemmas, the first free AI exercise (in Polish) uses neutral workshop data. You practise prompting without risking another person's dossier.

A ready prompt after preparing the note

Only after cleaning the material, add the task:

Using the fictional note, draft a reply to the complaint.
Acknowledge receipt and describe the next procedural step.
Do not grant a replacement or refund or accept responsibility for the fault.
If information for the next step is missing, identify it outside the draft.
Do not ask for data the note does not require.
Tone: helpful and factual. Maximum 100 words.

A correct reference answer says the complaint will be checked. It does not announce a replacement: the handler decides that. Check too that the model has not added a request for an identity-document number. An impressive form does not help repair the lamp.

The guide to writing an email prompt (in Polish) explores how to preserve commitments. For larger collections, start with controlled PDF reading (in Polish) instead of sending an entire archive at once.

Apply the check to the answer as well. A fictional faulty draft says: “We received the complaint about lamp L-7. Please send your identity-document number. We confirm a replacement once the number is received.” Neither procedure nor permission for a replacement appears in the note.

Check the faulty draft above against note K-18.
Identify every request for additional data and every new commitment.
For each, say whether it is supported by the described procedure.
Prepare a corrected acknowledgement, completeness check and handover
of the decision to the handler. Do not add questions unless a need is shown.

The corrected message should remove both the document request and the promise of replacement. Check that the model has not replaced them with another unnecessary request. The task is not to find a more elegant way to collect the same data.

Use the same assessment when choosing an AI tool for work (in Polish). If it needs broad system access merely to improve three sentences, record that organisational cost in the comparison. Good wording does not determine acceptable access.

Check attachments and output too

Open the copy before submitting it. Look for comments, other sheets or pages unrelated to the task. Covering a name with a rectangle on screen does not prove it was permanently removed. Check whether the text can still be copied or searched.

After receiving the answer, decide who may see it. A generated summary may contain the same confidential information as the input. Apply retention rules to both and check where the service stores the conversation and files.

Finally compare the cleaned note with the task: can someone still write the acknowledgement and understand the lamp problem? Removing a necessary condition can obstruct the work as effectively as excess data. Keep the information needed for the task, not simply the longest or shortest possible version.

If you connect company systems, permission to perform actions is a separate question, discussed in AI automation with human approval (in Polish). The tools (in Polish) discussed in the AI Course (in Polish) are selected for the task and its constraints; Aeneia provides the wider context. For your next note, ask: “Which details will change the outcome?” Leave the rest outside the chat.

*The source states that documentation was checked on 30 October 2026. This is a fictional exercise, not a legal assessment of a particular processing operation.*

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Explore the AI course (in Polish)

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear