Guide

Preparing a document before sending it to an AI tool

Summarising a contract, rewording a report, analysing a table: AI assistants are useful. What you paste into them leaves your computer.

Updated

In short

  • A document sent to an online AI tool is processed on its provider's servers.
  • Mask the people it names before sending it: names, contact details, identifiers.
  • Keep what the task needs: the structure, the amounts, the dates if they matter.
  • Do the masking on your computer, not with another online service.

What happens to the document

An online AI assistant processes what you send on its provider's servers. Depending on the provider, the plan and the settings, the content may be kept for some time, reviewed for abuse, or used to improve models. Your organisation may also have rules about which tools may receive which documents.

The question is not only whether the provider can be trusted. The people named in a document did not agree to their data being sent to a new recipient, and that decision is often not yours alone.

What to mask

  • Names of people, and their signatures.
  • Postal and e-mail addresses, phone numbers.
  • Identifiers: IBAN, social security number, company numbers, client or file references, licence plates.
  • Indirect identifiers when they point to one person: a rare job title, a precise date of birth, a place, an unusual event.
  • What another rule protects: trade secrets, confidential figures, privileged content.

What to keep

A document stripped of everything is no longer useful. Keep what the question needs: the structure, the clauses, the amounts, the dates when they matter. Masking a name does not stop an assistant from summarising a contract or spotting an inconsistency in a report.

Masked values become black blocks, or █ characters in a text file. The assistant sees that something was there, not what it was. If the task needs to tell people apart, say so in your question ("the first party", "the second party").

Why the masking should happen offline

Many anonymisation tools are themselves online services: to remove the data before sending it to one provider, you send it to another. Masking the document on your own computer means the original never leaves it; only the masked copy does, to the recipient you choose.

Step by step with Obscelis

  1. Import the document: a PDF, a scan or a text file (TXT, Markdown, CSV, JSON, XML, HTML, vCard).
  2. Review the findings: accept what goes, reject what the task needs, draw zones over what is missing.
  3. Export the copy. A text file keeps its format, each masked character replaced by █; a PDF is exported with its pages rebuilt as images, which most assistants read.
  4. Send the masked copy, and only it, to the AI tool.

The limits

Masking names and identifiers lowers the risk; it does not make every document anonymous. What remains (a context, a sequence of events) can still identify someone to a reader who knows the situation. Read the masked copy as its recipient would before sending it.

Questions

Can I simply ask the AI not to keep my data?

Settings and plans that limit retention or training reduce what the provider keeps, but the document is still sent and processed on its servers. Masking before sending is the only way the masked data never leaves your computer.

Will the AI understand a document with masked parts?

Usually, yes: the structure and the meaning remain. Tell it what the masks stand for if it matters ("the employee", "the supplier"), and keep the values the task actually needs.

Does Obscelis use AI itself?

Its optional detection model runs on your computer, like the rest of the analysis. It is downloaded only if you ask for it in Settings, and nothing about your documents is ever sent.

Try Obscelis for 30 days

Download it for macOS, Windows or Linux: the trial needs no account and no card. A question, or a demonstration for your team? Write to us.

Download ObscelisWrite to us

latelierdudev.contact@gmail.com

Opens your e-mail app. We only use your address to reply.