← Back to Insights
Compliance

Documenting AI Compliance Under Law 25: What Quebec Actually Requires

Law 25 requires a privacy impact assessment for AI tools, not a vendor's word. What Canadian organizations need on file, and why.

By Augure Newsroom·
A man sitting at a desk working on a computer

Quebec's privacy regulator closed 37 files involving automated decision systems and personal information in the twelve months following Law 25's September 2023 deadline, according to the Commission d'accès à l'information's most recent activity report. Most of those files did not end in a fine. They ended in a document request the organization couldn't satisfy.

Procurement teams evaluating a Canadian AI platform or a US chatbot subscription tend to miss that part. The CAI is not generally showing up to test whether an AI tool is good. It wants a paper trail: a privacy impact assessment, a data flow diagram, a record of who approved the deployment and why. Organizations that adopted a Canadian AI tool because it sounded safer have, in more than one case, discovered they still hadn't written anything down.

What Law 25 Asks For

Quebec's Act respecting the protection of personal information in the private sector, the law everyone calls Law 25, requires a privacy impact assessment, or PIA, for any project to acquire, develop, or overhaul an information system or electronic service delivery system that involves personal information. An AI chat tool that retains conversation history, a document Q&A system indexing client files, a contract review tool parsing NDAs: all three trigger this requirement if personal information passes through them.

The CAI has published guidance on what a PIA should contain, and it is less prescriptive than a lot of compliance teams expect. There's no official template. What the Commission has said, consistently, is that the assessment has to be contemporaneous: done before or during deployment, not reconstructed after an inquiry letter arrives. A PIA written six months after a tool went live reads differently to an investigator than one with a timestamp from procurement.

The assessment answers a short list of questions, and most organizations can draft a reasonable version in an afternoon once they have the facts from their vendor. What data categories move through the tool. Where that data is stored and processed, including any cross-border transfer. Who at the vendor can access it, under what contract terms. What happens to the data if the contract ends. Whether the tool trains on customer input. Whether there's a subprocessor table, and whether it's current.

A reasonable objection here is that an afternoon sounds optimistic for anything a regulator later scrutinizes. It is, for the first one. Privacy officers who have done this more than once describe the first PIA for a given vendor as the expensive one: it requires chasing down a data processing agreement, reading a subprocessor table that may not be formatted for easy extraction, and sometimes emailing a vendor's support desk because the public-facing documentation doesn't answer a specific question about retention after contract termination. The second PIA, for a different tool from the same category, reuses most of that structure. What rarely gets reused is the vendor correspondence — each product has its own gaps, and assuming one vendor's disclosure practices apply to the next is itself a documentation risk.

The Cross-Border Question

Getting a straight answer from a vendor about where processing happens is harder than it should be, and this is where documentation efforts usually stall. Law 25 requires an assessment of transfers outside Quebec specifically: a factual account of what leaves the province and under what safeguards, not a general assurance that things are fine.

Augure's privacy documentation, for comparison, discloses that customer data is stored in Canada and that inference for its chat and document products runs on Canadian infrastructure for some model tiers, with EU partners handling other tiers and failover under zero-data-retention agreements. Never US providers, for AI inference or customer content. Payment processing and email delivery involve limited US-based processing, which the privacy policy also lists. That is the kind of specificity a transfer assessment needs: which flows exist, and where they go.

One Montreal privacy lawyer put it this way at a recent compliance conference panel: the assessment has to say what actually happens to the data, not what the organization hopes happens to it. That framing tracks the Commission's own published guidance on conducting PIAs.

Organizations running US-headquartered AI tools face a harder version of this exercise. They also have to account for US jurisdictional exposure: the CLOUD Act, FISA 702, and whatever their vendor's own disclosures say about government access requests. A Canadian AI platform with no US corporate parent and no US investors doesn't make that question disappear, but it narrows what has to be analyzed for customer content, since no US-jurisdiction entity sits in that part of the processing chain. That is a genuine distinction in the documentation exercise, not a compliance shortcut. It says nothing about the payment and email processors that still sit in the US.

Developer-facing tools complicate the transfer assessment further, and this is a category most PIA templates don't anticipate. Augure's MCP server, used by developers connecting tools like Claude Code, processes text in Canada, with inference served from Ontario capacity and embeddings, reranking, and OCR routed to Calgary when a tool calls for them. If Canadian capacity is unavailable, the call fails outright rather than routing elsewhere — it returns a canadian_capacity_unavailable error instead of leaving Canada. That failure mode is itself worth documenting in a PIA, because it means the tool has no silent cross-border fallback to disclose; the alternative to Canadian processing is no processing, not foreign processing. Organizations should not extend that disclosure to claim the third-party assistant itself becomes Canadian or sovereign — the assistant's own provider still receives the conversation and tool results under its own terms. The MCP disclosure covers what Augure does with the call, not what the client application does afterward.

Building the File

Most PIAs that survive CAI scrutiny share a structure, even without a mandated template. A privacy officer typically assembles five components:

  • A data inventory: what personal information categories the AI tool will touch, pulled from actual use cases, not hypothetical ones.
  • A processing map: storage location, inference location, subprocessors, and retention period, sourced from the vendor's privacy policy and any signed data processing agreement.
  • A necessity and proportionality test, required under CAI guidance, explaining why the AI tool is the appropriate means to the stated end.
  • A risk mitigation section covering access controls, encryption, and incident response.
  • Sign-off from whoever holds the privacy officer designation Law 25 requires every organization to name, a role requirement that predates the PIA obligation but gets checked alongside it.

Teams using Augure for contract review or internal knowledge bases still have to produce this file themselves. No vendor's architecture, Canadian or otherwise, completes a PIA on an organization's behalf. The obligation sits with the organization doing business in Quebec, and the CAI has been clear that vendor marketing material doesn't substitute for an internal assessment.

The necessity and proportionality test is the component most often done badly, mainly because it's the one with no vendor documentation to lean on. A processing map can be assembled from a subprocessor table. Necessity has to be argued from the organization's own workflow — why a contract review tool, specifically, rather than a manual process or a narrower rules-based system. Skipping this step doesn't usually surface until a CAI investigator asks for it directly, at which point reconstructing the reasoning after the fact looks exactly like what it is.

PIPEDA and the Federal Layer

Outside Quebec, PIPEDA governs private-sector data handling. It doesn't mandate a PIA by name the way Law 25 does, but the Office of the Privacy Commissioner of Canada has said organizations conducting systematic data processing should be prepared to demonstrate accountability under their Schedule 1 obligations. That means a documented decision trail, even informally kept, matters there too. An organization operating in both Quebec and the rest of Canada is generally better served by one assessment framework applied consistently than by two separate compliance efforts running in parallel.

The federal picture is also shifting. The Artificial Intelligence and Data Act, introduced as part of Bill C-27, died when Parliament prorogued in January 2025, and no replacement framework has passed as of this writing. PIPEDA and Law 25 remain the operative rules for Canadian organizations deploying AI tools, with sector-specific guidance from bodies like the Law Society of Ontario filling some of the gaps for regulated professions.

One consequence of that gap rarely gets spelled out: without AIDA or an equivalent, there is no federal requirement specific to algorithmic decision-making or model behaviour, only the general personal-information obligations under PIPEDA. An organization whose AI tool denies a loan application or flags an employee file isn't working from a dedicated algorithmic accountability statute the way it would under the EU's AI Act. Law 25's PIA requirement ends up doing double duty in Quebec, covering both data handling and, indirectly, the automated-decision disclosure obligations the law added in 2023. Elsewhere in Canada, that second layer simply isn't legislated yet.

Where the Habit Breaks Down

The failure mode isn't usually bad faith. It's sequencing. A team adopts a tool, uses it for three months, gets value from it, and only then does someone in legal or privacy ask for the assessment that should have come first. By that point the PIA reads as a retrofit, and retrofits are exactly what draw CAI scrutiny when a complaint triggers an inquiry.

Organizations evaluating AI vendors now have an easier version of this problem than they did two years ago, mostly because vendors have gotten better at publishing the specifics a PIA needs: subprocessor tables, data flow disclosures, retention policies, without requiring a phone call to extract them. Augure publishes this information on its site rather than behind a sales conversation, which shortens the drafting work but doesn't remove it. Disclosure supports an organization's compliance case. It doesn't substitute for one.

The organizations that come through a CAI inquiry cleanly can hand over a dated file, not a reconstructed one. Writing it before deployment, not after someone asks, is what separates the two.

A

About Augure

Augure is a sovereign AI platform for regulated Canadian organizations. Chat, knowledge base, and compliance tools — all running on Canadian infrastructure.

Ready to try sovereign AI?

Start free. No credit card required.

Get Started