veille Product, The agent
The agent

An AI agent that reads regulations the way a lawyer does, and runs 24 hours a day.

Veille evaluates each AI system against its applicable obligations, article by article. It detects regulatory changes, re-evaluates affected systems automatically, and when evidence is missing, it drafts the document itself, bilingually, ready for your team to review and sign.

ModelsClaude family, routed by task
InfrastructureAWS ca-central-1 (inference migrating to Bedrock)
Detection time<4h after a regulatory change

Architecture

Three loops. One continuous cycle.

Veille runs three independent loops on a continuous cadence. Each loop has a distinct purpose. Together they cover the full compliance lifecycle.

Loop 01
Watch Loop
Monitors official regulatory sources for changes and assesses materiality for each registered system.
  • Six official sources: Gazette officielle du Québec, Gazette du Canada, OSFI, EU AI Office, EUR-Lex (series L), Ontario newsroom
  • Scores the materiality of each change per registered system
  • Tags affected frameworks and articles
  • Triggers re-evaluation for linked systems
Loop 02
Discovery Loop
Identifies AI systems in operation but not registered: shadow AI creating untracked compliance exposure.
  • Scans declared environments for unregistered endpoints
  • Flags systems without a designated owner (Art. 3.1)
  • Generates intake form for discovered systems
  • Creates a priority gap for each unregistered system
Loop 03
Compliance Loop
Evaluates each registered system against every applicable obligation for its assigned frameworks.
  • Maps each system to its applicable articles
  • Evaluates each article: compliant, gap, review, or N/A
  • Scores the system's overall posture (0-100)
  • Drafts remediation plan for each identified gap

How it reasons

Not a rules engine. A reasoning system.

Most compliance tools operate on boolean rules: is this checkbox ticked? Veille reasons over the evidence you provide and interprets each obligation in context.

01
Article-by-article evaluation
Each obligation is evaluated individually. Veille reads the article text, maps it to the evidence on file for the system, and produces a finding with a justification. Every compliance status traces back to a specific article and a specific piece of evidence, or the absence of one.
02
Contextual reasoning over evidence
Veille reasons over actual documentation you provide: model cards, PIAs, transparency notices, test results. It identifies what is present, what is missing, and what is insufficient. It does not make decisions based on form fields.
03
It doesn't just flag the gap: it drafts the missing document
When a required piece of evidence doesn't exist: a PIA, a governance policy, an automated-decision notice. Veille drafts it: full content, in French and English (never one without the other), citing the exact article it addresses. The draft is hashed and chained into the evidence vault immediately, fully traceable from the moment it's created, and a notification is queued for the system owner to review the draft. It is a head start on the writing, not a substitute for your DPO's judgment or signature.
And this isn't just a guideline we follow: the platform enforces it at the data layer. Every AI-drafted artifact is stored with a draft_pending_review status, and the compliance engine is built to ignore it: the obligation stays marked as a gap until a qualified human opens it, reviews it against your real operations, and formally validates it. Only that human action flips the status to validated and lets it count as evidence. There is no code path that lets an AI draft promote itself. The 146 encoded obligations are themselves anchored to official source text (drafts), with 0 of 146 validated by counsel to date; an external Canadian AI/privacy law firm partnership is in formalization to anchor that review.
04
Human in the loop, by design
Veille flags gaps and drafts plans. It does not close them unilaterally. Every remediation requires a human sign-off, logged in the audit trail. Veille handles research and drafting. The human carries the legal responsibility.
Model & residency note

Veille uses Claude (Anthropic) for compliance reasoning. LLM inference is being migrated to AWS Bedrock in Canada (Montréal) for full Canadian residency; until that is activated, inference runs on Anthropic's API. Storage and application processing are in Canada. No training data is shared with Anthropic.

Demo

See Veille evaluate a real system.

The demo dashboard shows the full compliance loop on a representative 7-system inventory.

View demo dashboard → Book a call