Skip to content
Guilda Work Help center
Go to the app

Help center / AI Assistant / Personal data protection in the assistant

Personal data protection in the assistant

Open in Guilda Work

What is replaced before anything is sent to the AI model, what is not, and what happens if it fails.

Updated on

  • Management
  • Coordination
  • Tax review
  • AI assistant

The assistant uses a language model from an external provider. So that provider does not receive who each person is, Guilda Work swaps personal data for labels before sending each query, and puts them back when the answer arrives. It is automatic and needs no setup.

What is replaced

Data Appears to the provider as
Names of people [PERSONA_1], [PERSONA_2]…
NIF and NIE [NIF_1]…
Company CIF (with check digit) [CIF_1]…
IBAN [IBAN_1]…
Payment cards [TARJETA_1]…
Phone numbers [TELEFONO_1]…
Email addresses [EMAIL_1]…
Postal addresses [DIRECCION_1]…

Your own name is protected too. Companies, cities and dates are not replaced: they are not data of a natural person and removing them would make answers worse.

How it stays useful

The same person always gets the same label within a conversation, so the model can reason about "[PERSONA_1]" without knowing who it is. When it answers, or asks to run an action with those labels (for example, creating a task), Guilda Work puts back the real data: you see the real names and the tools work with real data. You will not notice anything while the answer appears gradually.

If the detector fails, nothing is sent

If the system that recognises personal data is unavailable, the assistant blocks the sending and shows you an error: it never sends your query unprotected. If it happens repeatedly, tell support.

Limits worth knowing

  • Name recognition is statistical: it may miss an uncommon name or, conversely, hide a word that was not a name.
  • Images and PDFs read with AI cannot be protected this way. For example, when the assistant reads a scanned invoice, the document reaches the provider as is. Avoid uploading documents with data you do not want to share with the AI provider.
  • Social Security numbers and health data written in free text are not detected.
  • You or your firm choose the provider and model in Assistant settings: it is wise to choose models with zero retention and no training and to review their privacy policy.

What the firm must do

This protection reduces risk, but it does not replace the firm's obligations as data controller: informing clients and employees, signing the processing agreement with Guilda and, for high-risk data, considering an impact assessment. These decisions should be reviewed with your legal advisers. See also Privacy and retention.

Each person can also delete their history from My data and privacy.

Related guides

Back to AI Assistant

Was this guide helpful?

We only count your vote: we don't store who you are.

Can't find the answer?

Write to us and we'll help.

Write to support