Artificial Intelligence Policy
How Veille uses artificial intelligence to read texts, summarise them and propose their applicability — and the safeguards that ensure no proposal becomes advice without a consultant’s approval.
On this page (6 sections)
Veille reads official texts for you. This policy describes what artificial intelligence does, what it does not do, and how you stay in control.
1. What the Service does
- Relevance filter — it sets aside texts outside the QSE scope (quality, health & safety, environment, labour); you can restore a text that was set aside.
- Cited summary — it summarises each text in plain language; every sentence links to the passages of the text it comes from. A sentence without a cited passage is removed.
- Obligations — it extracts who must do what, by when, under which conditions and with which penalties, quoting the exact article.
- Matches — for each client, it proposes whether a text applies, may apply, does not apply, or whether information is missing, with its justification and a confidence level.
- Translation — when an AI provider is enabled, it produces the Arabic version from the approved French version, keeping the citations.
2. Two engines
- The built-in engine (default) runs on our servers, deterministically, from French QSE lexicons and rules. It sends no data to third parties.
- An AI model provider (Anthropic) can be enabled for richer summaries, Arabic and criteria that are hard to express as rules. It only receives the necessary passages and information, and is not allowed to use them to train its models.
3. Safeguards
- Nothing without a citation: every statement links to the exact passage of the stored text.
- Nothing without approval: no match enters the register and no report is sent without the approval of an authorised consultant.
- Visible confidence: every proposal shows its confidence level; uncertain cases are flagged.
- Documents are data: the content of a text is never executed as an instruction.
- When in doubt, a human reads: a poorly scanned text or a non-conforming output is marked “needs human reading” rather than filled in by guesswork.
4. Learning
Your corrections (edited summaries, matches rejected with a justification) can become rules and examples specific to your firm. They are never shared with other firms or used to train models.
5. Limits
Artificial intelligence can make mistakes, misread a threshold or miss a legal subtlety. The Service is a decision-support tool: it does not provide legal advice, and only the official texts are authoritative.
6. Contact
In case of discrepancy between this English version and the French version, the French version prevails.