Klytic

Home›Guides›Documents & document management

Documents & document management · Understand

Classifying or searching documents with AI, without sending their content outside

It is possible, provided the model runs where the documents already are, or with an operator whose role you know, and that neither excerpts nor queries are used for anything else. The question is not the model’s performance, it is data leaving the company. And AI does not replace a folder structure, permissions or a retention period.

Updated October 20266 minOfficial sources cited

What leaves when a document goes to an AI service

Sending a file to an online assistant, or connecting an AI tool to a document space, sends out more than the document’s text:

  • the content, in full or in excerpts;
  • the metadata: file name, author, folder path. “Clients/Dupont/Dispute 2026” already says a lot;
  • the query itself: “find the contracts we can terminate without penalty” reveals an intention;
  • the provider’s logs: queries and answers, kept according to its terms;
  • a derived index: to search by meaning, a tool builds a representation of the content, which is a form of copy;
  • subsequent use: depending on the contract, the data may be used to improve the model.

On top of this comes the question of the law applicable to the provider, covered in protecting a company from the CLOUD Act.

The conditions for processing locally or within the instance

To deploy generative AI, the CNIL recommends choosing a secure deployment mode, for example by favoring local, secure and specialized systems. Failing that, it asks you to determine to what extent the provider is likely to reuse the data supplied, and to adapt use accordingly. In practice, six questions:

  1. Where does the model run? On the workstation, on a company server, in the DMS provider’s instance, or with a third-party AI provider.
  2. Who operates it? The company running the model has access to what it reads.
  3. Are queries and excerpts logged? For how long, and who can read these logs?
  4. Are the documents used to improve the model? The answer must be written into the contract.
  5. Does the index respect permissions? A person must not obtain, by asking the AI a question, the content of a folder they cannot open.
  6. What happens to the index when a document is deleted? Otherwise, the retention period applies only to the file, not to its copy.

The CNIL also recommends governing use with a list of what is permitted and prohibited, for example not supplying personal data to the system when this is not intended.

What AI does, and what it does not decide

Properly governed, AI can suggest tags, extract a date, an amount or a supplier name, or answer a search phrased in everyday language. These are suggestions. They are sometimes wrong, and rarely visibly so.

It does not replace three decisions:

  • the folder structure: who owns which folder, where a contract is filed. Automatic filing into a structure nobody decided on is still a mess;
  • permissions: AI does not know who should see payroll;
  • the retention period: it comes from the law or a policy, see retaining and archiving documents.

For documents that commit the company, a person validates. That is the role of the approval workflow, not of the model.

The limit: a model operated by your provider is still access

Running the model at the DMS provider avoids sending documents to a third party. That does not mean there is no access. To read a document, the model must see it in clear text. The operator running the model therefore technically has access to the content, as when it restores a backup or steps in for support.

Encryption makes this trade-off visible. A document encrypted at the source, with a key derived from the user’s account, is unreadable for the operator. It is also unreadable for a model running on the server. Depending on how the folder is encrypted, you gain confidentiality or server-side processing, rarely both. The question of who holds the keys is developed in what is a sovereign cloud?

Example: finding clauses in contracts

Hypothetical case. A fifteen-person firm wants to find all contracts containing a non-compete clause. Three options.

A consumer online assistant. The contracts are uploaded to it. The content, client names and the query leave the company. How they are kept and reused depends on the service’s terms.

A model on a company server. It indexes only the “Contracts” folder, with the folder’s permissions. Logs are limited and read by the administrator. The content does not leave the infrastructure, but the server has to be run and kept up to date.

No AI. A “non-compete” tag is applied at upload, within the approval workflow. Searching by tag finds the contracts in a few seconds. For a recurring, well-defined need, this is often enough.

The right choice depends on how often the question comes up and how sensitive the documents are, not on the appeal of the tool.

Not to be confused with meeting transcription

A video conference can produce a transcript and a summary. That is processing of the meeting, not classification of files. The summary becomes a document like any other: it is filed, shared and retained under the same rules. It does not make the DMS “intelligent.”

Frequently asked questions

Can a consumer AI assistant be used on internal documents?

Only after reading its terms: retention of queries, reuse of data, applicable law. The CNIL recommends governing its use.

Is an AI installed locally necessarily safe?

No. It prevents data from leaving, but you still need to manage the index’s permissions, logs and deletion.

Can AI set retention periods?

No. They come from the law and from your policy. See retaining and archiving documents.

What about AI-based email filtering?

The topic has its own guide: AI and filtering confidentiality.

Can AI be added to Klytic Docs?

Yes, through an extension you add to your instance. Before enabling it, check where the model runs and what it keeps, using the questions in this guide.

How do you classify without AI?

With team folders, tags and a workflow. See finding, tagging and routing a document.

Where Klytic fits

Klytic Docs does not include AI by default: classification relies on team folders, tags, approval workflows and search. If you wish, you can add an AI extension to your instance, of which you are the administrator. The questions in this guide then apply to that extension: where the model runs, who operates it, what is kept and what leaves the instance.

The video conferencing included in Klytic Docs allows recording, transcription, translation and summarizing of meetings, within the customer’s instance. It is a meeting service, not an intelligent DMS. It is provided for a limited number of users; support for a large number of video conferencing users is subject to a specific quote.

For documents encrypted at the source, the key is generated from the customer’s account: Klytic cannot read their content. Data the service must process is encrypted at rest with keys managed by Klytic. On Klytic’s side, technical access is named and logged.

For email, Klytic MTA uses a model operated by Klytic, or within the customer’s instance, not by a consumer service, to decide only on messages with an ambiguous score. Details are in the guide on filtering.

Hosting is possible in any geographic area matching the applicable jurisdiction, subject to service availability, notably in Europe or Mauritius, or on the customer’s own servers. Klytic is not a SecNumCloud-qualified offering.

This page describes a general framework. It does not replace an analysis of the terms of a specific AI service.

Sources

Consulted in October 2026.

  • CNIL, “Comment déployer une IA générative ? La CNIL apporte de premières précisions” (How to deploy generative AI? The CNIL provides initial guidance), July 18, 2024: favor local, secure and specialized systems; failing that, assess the provider’s reuse of data; govern use, for example by not supplying personal data. cnil.fr
  • Klytic, What is a sovereign cloud?: key for documents encrypted at the source generated from the customer’s account and unreadable for Klytic; data processed by the service encrypted at rest with keys managed by Klytic. klytic.com
  • Klytic, Klytic Docs page: classification by folders and tags, approval workflows, search. klytic.com
  • Klytic, Hosting and sovereignty: technical access limited to operations and support, named and logged. klytic.com

Talk about your sensitive documents

Encryption, permissions, classification: let’s see what suits your folders.

Talk to an advisor →

Discover Klytic Docs

Further reading

Use · Documents & document managementFinding, tagging and routing a documentTeam folders, tags, search and approval workflows: organize Klytic Docs so you can find a document, even in a large organization.6 min readRead →Understand · Email securityAI in email filtering, and what leaves the companyAI helps decide ambiguous messages. Privacy is judged by what the model receives, who operates it, what is kept and what is shared.6 min readRead →GuidesWhat is a sovereign cloud?A sovereign cloud is judged by its operator, the applicable law, control over access and data exit. The datacenter’s location is not enough.GuidesRead →

Welcome offer

30 days free trial, migration support included

No-commitment trial. An advisor calls you back to understand your needs and prepare your Klytic workspace.

🚀 Start the free trial See pricing