AI Non-Training: Legal Requirements and What Leexi Provides

Leexi does not use your meetings to train AI models. This commitment is documented in the privacy policy and compliance FAQ, applies by default to all accounts, and an attestation is provided upon request for your records. The models reading your meetings to produce transcripts and summaries are provided by third parties named in the DPA, all hosted within the European Union, and bound to Leexi by obligations at least as protective as those in the DPA.

Understanding What "Non-Training" Means

Training a model means modifying it with your data. Reading your data to produce a summary is different.

When Leexi transcribes a meeting and drafts the summary, a language model reads the audio or text of that meeting and produces a result. This is the use you request. The model retains nothing from that meeting after producing the summary.

Training involves using thousands of meetings to adjust a model’s parameters to perform better in the future. Your words then become part of the model and may reappear in responses to others. The non-training commitment excludes this.

How Do Buyers Phrase This Requirement?

A wealth management buyer asked Leexi in May 2026 for a "contractual commitment or written clause 'non-training of models' specifying that client data (including logs, transcriptions, embeddings) will not be used to train, improve, or optimize AI models without prior written consent." A public organization requested in March 2026 "guarantees that our data is never used to train your AI models."

Both formulations express the expected level: the commitment must cover transcriptions, technical logs, and vector representations ("embeddings"), and address service improvement as well as strict training.

Is Non-Training a Legal Obligation?

No text explicitly names "non-training." Two GDPR articles produce the same effect.

The first is the purpose limitation principle.

GDPR, Article 5(1)(b) "collected for specified, explicit, and legitimate purposes and not further processed in a manner incompatible with those purposes"

You record a meeting to obtain its summary. Training a model with that meeting is a different purpose you have not specified.

The second is the obligation for processors to act only on your instructions.

GDPR, Article 28(3)(a) "processes personal data only on documented instructions from the controller, including regarding transfers to third countries or international organizations, unless required by Union or Member State law"

Leexi is your processor (Article 4.2 of its DPA). Training on your meetings would be processing without your instruction.

Three levels of force:

• Imposed by law: prohibition to process data for incompatible purposes and processor acting only on instructions.
• Required by buyers: a written clause naming training, improvement, and optimization of models, covering logs and embeddings. GDPR does not require this form, but DPOs request it as it is easily verifiable.
• Imposed by sector: no sector-specific framework mandates it as such. Finance and public sectors obtain it through security questionnaires.

Does Non-Training Exempt Informing Participants?

No. Informing participants is a separate obligation.

EU AI Regulation, Article 50(1) "Providers shall ensure AI systems interacting directly with natural persons are designed so that persons are informed they are interacting with AI, unless obvious to a reasonably well-informed person given context."

This article targets system providers. Your obligation to inform participants comes from GDPR, and the Leexi assistant is visible in meetings to help.

Knowing the Models That Read Your Meetings

Leexi names its AI providers in the DPA, with their hosting locations.

Leexi DPA, Article 6.1 "Microsoft Azure for transcription and generative AI → hosted in the EU" "Gladia for speech-to-text provider → hosted in the EU" "Eleven Labs Inc.: speech-to-text provider → hosted in Europe" "Mistral for generative AI → hosted in the EU"

The compliance FAQ details the setup: the language model is provided by OpenAI and run via Azure France in a closed circuit. Closed circuit means processing stays within a controlled perimeter in France, without exposure to third-party use.

Can the Model Provider Train Their Models on Your Meetings?

The DPA requires Leexi’s processors to have the same obligations as Leexi.

Leexi DPA, Article 6.3 "Leexi will ensure all processors engaged for personal data processing are bound by written agreements with data protection obligations at least as protective as those in this agreement."

This transposes GDPR Article 28(4), requiring "the same data protection obligations" for any downstream processor.

Obtaining Proof of Non-Training

The commitment is documented in two public places, and an attestation completes the file.

In the privacy policy, section "Security and AI Model Training":

Leexi Privacy Policy "We do not use this data to train or improve generalized or non-personalized AI/ML models. All applied AI processing is strictly limited to the features you request."

In the compliance FAQ, question "Does Leexi train its AI on my meetings?": "No. Your conversations are not used to train models. This applies by default to all accounts."

The non-training attestation is among six documents listed in the compliance FAQ for your DPO, along with hosting location, ISO 27001 certificate, DPA, list of processors, and retention policy.

Understanding the Limits of the Commitment

Three points before submitting the file.

• DPA version 2.0 contains no dedicated non-training clause. The commitment is in the privacy policy and FAQ. If your legal team wants a clause in the contract, request an annex at hello@leexi.ai.
• The privacy policy phrase is in the section on data from third-party integrations. The FAQ complements it, targeting the conversations themselves. An attestation clarifies the scope.
• Non-training does not mean AI does not read your meetings. It reads them to produce what you request: transcription, summary, Ask Leexi responses on your meetings. Your meeting content remains in your account, in the library, for the retention period you set.

Providing Documents to Your DPO

What AI does with recordings and what you must tell participants is covered on the record a professional meeting page. All certifications and audits are gathered in the security and compliance section.

The DPA applies from the free trial. You can test Leexi on a real meeting during your DPO’s review. Book a demo

Frequently Asked Questions

What Does "Closed Circuit" Mean for the Model Run via Azure France?

The language model runs within a controlled perimeter in France, and your data is neither used to train this model nor exposed to third-party use. This is the answer given by Leexi’s compliance FAQ and is the first point technical teams verify.

Are My Transcriptions Used to Improve Leexi?

No. The privacy policy excludes training and improving generalized or non-personalized models and limits AI processing to requested features. A DPO wanting coverage for logs and embeddings requests the attestation, which formalizes this scope.

Does a Custom Summary Template Train a Model?

No. A summary template is a set of instructions (a "prompt" in the app) the model reads for each summary. It does not modify the model. You can create one per meeting type without any data entering training.

Does Ask Leexi Learn from My Meetings?

Ask Leexi reads your meetings to answer questions across multiple summaries at once. This is a read at query time, limited to your account. The model is not modified by your meetings.

How to Obtain a Written Non-Training Attestation?

Request it at hello@leexi.ai specifying the scope your DPO expects (transcriptions, logs, embeddings) and your subscription date. It is one of six compliance documents Leexi provides, along with the DPA, processor list, ISO 27001 certificate, hosting location, and retention policy.

Is Non-Training Enough for a GDPR File?

No. It covers only one requirement: purpose limitation. Your file also needs the DPA (Article 28 GDPR), processor list, retention period, and proof participants are informed. The page DPA: Legal Requirements and What Leexi Provides covers the first three.

Summary

Your meetings are not used to train models. The commitment is public, applies to all accounts, and an attestation formalizes it for your DPO.

GDPR achieves the same effect through purpose limitation and the processor’s obligation to act on instruction. Leexi’s AI providers are named in the DPA, hosted in the EU, and bound by the same obligations.

What the DPA lacks—a dedicated clause—can be requested as an annex.

Sources

Texts reviewed from primary sources on August 22 and September 2, 2026.

• Leexi Privacy Policy, section "Security and AI Model Training", leexi.ai/en/privacy-policy
• Leexi Compliance FAQ and Sovereignty, published August 10, 2026
• Leexi Data Processing Agreement, version 2.0 effective July 6, 2026, Articles 4.2, 6.1, and 6.3
• Leexi Trust Center, trust.leexi.ai, accessed September 2, 2026
• GDPR Articles 5 and 28, text published by CNIL
• EU AI Regulation, Article 50

This page describes the applicable framework. It does not constitute legal advice.