Why US Legal AI Tools Are a Bad Fit for Canadian Practice

Torys announced in April 2026 a firmwide Harvey rollout, joining Davies, Gowling WLG, Caravel, and roughly two dozen other Canadian firms. Harvey now operates from a Toronto office on Bay Street. The marquee US legal AI products have arrived in Canada at scale. The question procurement is not asking with enough seriousness is whether the underlying architecture matches the duties Canadian lawyers actually owe their clients.

The product is good. The fit is not.

Harvey, Hebbia, Casetext (now Thomson Reuters CoCounsel), and Spellbook all do useful things: contract review, due diligence summaries, legal research at speed, document drafting that lawyers keep. Harvey and Legora carry combined valuations near $17 billion in 2026 for a reason.

That is the wrong frame for a Canadian buying decision. A useful tool deployed against the wrong duty profile is still a problem. Four structural mismatches show up for Canadian firms, and they get worse the further east the firm sits.

Section 17 and the prompt as a transfer

The first mismatch is the Quebec cross-border transfer regime. Section 17 of the Act respecting the protection of personal information in the private sector requires a written assessment before personal information is communicated outside Quebec. The assessment weighs sensitivity, purpose, contractual protections, and the legal framework of the destination state. An Ontario vendor already triggers it; a US-hosted LLM triggers a more serious version.

Every prompt that contains a client name, a contract excerpt, an opposing party's address, or a draft pleading is a section 17 event. A mid-size Montreal firm running Harvey across litigation, corporate, and real estate is running thousands of transfer events a month. The statute does not exempt them because the contract is enterprise-grade. The PIA exists or it does not. In practice, it does not.

On January 30 2026 the CAI published new guidance on confidentiality incidents that reads like a roadmap for the file it intends to take to penalty. The administrative regime reaches 2% of worldwide turnover or $10 million; the Court of Quebec can go to 4% or $25 million.

Secrecy, privilege, and the CLOUD Act

The second mismatch is professional secrecy. Quebec professional rules treat solicitor-client information as a personal right of the client; disclosure to third-party processors requires a real analysis, not a click-through to a terms-of-service page. Common-law privilege is more permissive about retained vendors, but the Federation of Law Societies Model Code still imposes a confidentiality duty that survives the engagement and binds the firm to control over disclosure. The default terms of most US legal AI products were drafted against US attorney-client privilege, a narrower and more flexible doctrine.

The third mismatch is training data. The models behind Harvey and CoCounsel were tuned on US legal English, Bluebook citations, federal and state procedural assumptions, and a body of case law that treats consideration, parole evidence, and contractual interpretation differently than the Civil Code of Quebec does. The hallucinations get quieter and worse: not invented case names, but the wrong default rule slipped into a draft because the model assumed a US framing.

The fourth mismatch is the CLOUD Act. BLG published an analysis in April 2026 confirming what the statute has said since 2018: US law enforcement can compel a US-controlled provider to produce data regardless of where it physically sits. A Canadian region label on an AWS bucket does not change the exposure if the parent contract is with a US entity. A Canadian firm using a US-hosted LLM for a client matter has accepted a foreign legal hook on that matter, which sits uneasily with the duty of loyalty.

The Canadian stack we are betting on

This is the design constraint we started SupaCorp with. Canadian small firms doing entity work cannot put a client's beneficial-ownership record, a director's home address, or a minute book extract into a US-hosted tool without a documented section 17 assessment and probably a specific client consent. So we built the storage and the AI processing inside Canadian regions from day one. The trade-off is real; we do not get to ride on top of the cheapest US-only AI infrastructure. The compensating advantage is that any Canadian small firm can adopt SupaCorp without taking on a Law 25 problem on its first matter.

The three viable paths

Canadian firms making generative AI work inside the duty profile are picking one of three architectures.

The first is on-premise or VPC deployment. Cohere, the Canadian foundation-model company that merged with Aleph Alpha in April 2026, ships exactly this through Model Vault: Command, Rerank, and Embed running inside the firm's own infrastructure or a customer-controlled VPC. Data never leaves the boundary. The PIA is short because the destination is the firm itself.

The second is Canadian-hosted commercial models. Azure Canada Central and Canada East, Google Cloud Montreal, AWS Canada, and Anthropic and OpenAI deployments routed through Canadian regions with the right contractual stack. CLOUD Act exposure does not disappear; it gets managed through a written transfer assessment the firm can actually produce when asked.

The third is Canadian-native products built with these constraints from the start. Spellbook, headquartered in St. John's, became the CBA's exclusive AI contract partner in March 2026 on the strength of Canadian data residency and a no-training commitment for CBA-channel customers. The argument is not that Canadian-built is better engineering; it is that the product was designed against the right duty profile and procurement diligence is shorter as a result.

What to watch

The first CAI penalty against a Quebec firm using a US-hosted legal AI tool is the document the rest of the market is waiting for. It will probably involve a complainant whose information surfaced in a vendor relationship the firm could not document. Watch for it in the next twelve to eighteen months. Once it lands, procurement conversations get much shorter and the Canadian-stack option moves from a defensive posture to the default.