Why a U.S. checklist doesn't work for AI legal software in Canada
Most content ranking for AI legal software today is a tools listicle: which product is smartest, which is cheapest, which one plugs into your existing case-management stack. Useful, but incomplete for a Canadian buyer.
Two extra questions sit in front of every capability comparison here. Where does the vendor actually store and process your data, and what does your own law society require of you regardless of what the vendor promises? Neither shows up in a features table, and neither is optional.
If your firm hasn't settled the broader adoption question yet — whether AI belongs in the practice at all, and where lawyers are already using it informally — start with our overview of what's actually working for Canadian firms in 2026 before working through vendor-by-vendor evaluation.
Start with hosting: “Canada” isn't a location until it's a region
“Our data is Canadian” is a marketing sentence, not a technical answer. Ask for the specific cloud region — AWS ca-central-1 (Montréal), Azure Canada Central, or the equivalent — in writing, not a jurisdiction claim.
A vendor can be a Canadian-incorporated company and still route your files through a U.S.-hosted AI provider for the actual inference step. That gap matters: the Office of the Privacy Commissioner of Canada's guidance on cross-border processing treats a transfer for processing as a “use,” not a “disclosure” — but the transferring organization stays fully accountable for what happens to the data once it leaves, and the primary safeguard is a contract, not a claim.
The market is starting to treat data residency as table stakes, not a nice-to-have. When the Canadian Bar Association signed Spellbook as its exclusive AI contract-drafting partner in March 2026 — a deal covering the CBA's network of roughly 40,000 lawyers, judges, notaries and law students — the residency term wasn't buried in a footnote: member data is stored on Canadian soil and is not used to train the model. If the national bar association negotiates that term for its own vendor, it's a reasonable term for any firm to ask for too.
Quebec's Law 25 has required a documented privacy impact assessment before any transfer of personal information outside the province since September 22, 2023. — Borden Ladner Gervais LLP, on Law 25 §17
PIPEDA and Quebec's Law 25 change what “buying” means
PIPEDA does not itself require Canadian data residency. Per the federal Privacy Commissioner's own guidance, a firm can send personal information to a processor anywhere, including outside Canada — but the firm stays accountable for it, and that accountability has to be backed by a real contract, not a checkbox on a terms-of-service page nobody read. That could change: Bill C-36, the Protecting Privacy and Consumer Data Act, was tabled in the House of Commons on June 15, 2026 and would add an explicit requirement to assess and mitigate privacy risks — including a privacy impact assessment — before any cross-border transfer of personal information, closer to Quebec's existing rule. It has not passed; Parliament rose for the summer days after tabling and resumes debate this fall, so the PIPEDA rule above is still the one in force as of this writing.
Quebec's Law 25 sets a materially higher bar. Since September 22, 2023, any transfer of personal information outside Quebec has required a documented privacy impact assessment first — and it applies based on whose personal information is involved, not where your firm's office sits. An Ontario firm with even one Quebec-resident client or opposing party on a file can trigger this requirement.
None of this replaces advice on your firm's specific obligations. Our guide to the three regulatory layers governing AI in Canada walks through how the federal, provincial and law-society layers stack — including why Canada currently has no AI-specific federal statute in force.
LSO duties don't stop at the login screen
Competence. The Law Society of Ontario's generative-AI white paper, published April 25, 2024, is explicit that “certain tasks will require the input of the licensee and the application of their legal competence and cannot be delegated” — not to the software, and not to the vendor. Picking a vendor that can't explain what its AI actually did to produce an output doesn't satisfy this duty; it just moves the gap somewhere less visible.
Confidentiality. The same guidance requires “adequate security measures… where confidential or privileged information may be used as an input” — a vendor-selection question as much as a use-policy question. You can't fix this after the fact if the vendor you picked never had adequate controls to begin with.
Supervision. A firm principal is responsible for how staff use a tool, the same as any other software or service provider the firm engages.
The Canadian Bar Association's own ethics toolkit reaches the same conclusion from a different angle: its guidance on choosing AI tools tells lawyers to scrutinize their relationships with third-party AI providers and set forth “contractual provisions in procurement” that ensure compliance with legal and professional obligations, with clear data-privacy and security terms, before signing anything. For the deeper walkthrough of the four vendor-safety questions this maps to — training, retention, location, autonomy — see our AI vendor safety checklist.
The buyer's checklist: seven questions before you sign
Bring these into the vendor demo or the RFP itself — not a follow-up email you send if you remember.
| Question | Why it matters | Red-flag answer |
|---|---|---|
| Which specific cloud region hosts our data at rest? | “Canada” is not a region — you need the named region (e.g. AWS ca-central-1) written into the contract. | Can't or won't name a specific region. |
| Does the underlying AI model provider train on our inputs or outputs? | Training exposure and storage/retention are two different questions — get both answered separately. | “We don't retain data” without addressing training. |
| How long is data retained before deletion, in writing? | Feeds directly into your firm's own file-retention and privilege obligations. | Vague answer, or retention framed as “for product improvement.” |
| Has a Law 25 privacy impact assessment been done for any transfer involving Quebec-resident data? | Legally required since September 22, 2023 for any such transfer. | Vendor unfamiliar with the requirement. |
| What can the AI act on without a person approving it first? | Maps directly to the LSO's supervision duty. | Vendor can't name a specific, bounded list. |
| Who are the sub-processors, and where are they located? | A support contractor with data access outside Canada can itself be a cross-border transfer. | No published sub-processor list. |
| What happens to our data if the vendor is acquired or goes under? | Ownership changes can move your data's jurisdiction without your firm's consent. | No contractual answer. |
Run an evaluation, not just a demo
A capability demo tells you what the product can do. It doesn't tell you what your firm is agreeing to. Treat the seven questions above as required attachments to any RFP or vendor questionnaire.
Ask for the data-processing addendum before the trial, not after the contract. A vendor that can produce one on request, with the specific region and retention terms named, has usually already done this work for another Canadian firm. A vendor that has to “check with legal” for basic terms usually hasn't.
Run the trial on a real but low-stakes file, and involve whoever signs off on the firm's technology purchases from the start. This mirrors the same lens our Canadian practice-management software comparison uses for the wider category — hosting location, trust-accounting enforcement, AI data handling and cost, evaluated together rather than one at a time.
Where DROZlegal fits
DROZlegal is one answer to the checklist above, not the only one — worth including here because every claim in it is written down, not asserted. Client data is stored in AWS ca-central-1 (Montréal) and never leaves Canada; AI processing runs under Anthropic's commercial API terms — no training on inputs or outputs, and both are auto-deleted within approximately 30 days. That is not the same thing as a negotiated zero-retention arrangement, and it's worth holding any vendor to that same precision rather than a softer claim.
On the supervision question, DROZlegal's seven production AI agents run to a fixed set of six actions — moving trust money, filing with a court, settling, commencing litigation, approving an engagement, and sending email outside the firm — that stay permanently gated to a human sign-off, by design, with no setting that turns that off. The full architecture and compliance posture, including SOC 2 Type II status and PIPEDA/Law 25 alignment, is documented on our security and trust page.
Frequently asked questions
What makes evaluating AI legal software different in Canada than in the U.S.? In Canada, evaluating AI legal software adds two layers on top of the capability-and-price comparison most U.S.-focused guides run: data residency and processing location, and the extra privacy and professional-conduct obligations that apply regardless of what the vendor promises. Quebec's Law 25 and the Law Society of Ontario's generative-AI guidance both attach real requirements to vendor selection itself, not just to how a lawyer uses the tool afterward.
Does PIPEDA require Canadian data residency for legal AI software? No, PIPEDA does not itself require Canadian data residency. Per the Office of the Privacy Commissioner of Canada's guidance, a firm can send personal information to a processor anywhere, including outside Canada, but the firm remains fully accountable for what happens to it, and that accountability has to be backed by a real contract rather than a vendor's marketing claim. A pending federal bill, C-36, tabled June 15, 2026, would require a privacy impact assessment before such transfers if it passes; it had not passed as of this writing.
What does Quebec's Law 25 require before buying AI software that will process client data? Quebec's Law 25 has required a documented privacy impact assessment before any transfer of personal information outside the province since September 22, 2023. It applies based on whose personal information is involved, not where a firm's office is located, so an Ontario firm with even one Quebec-connected file can trigger the requirement.
What Law Society of Ontario duties apply when choosing an AI vendor, not just when using the tool? The Law Society of Ontario's professional-conduct duties of competence, confidentiality and supervision all apply to vendor selection, not only to day-to-day use. The Law Society's generative-AI guidance, published April 25, 2024, states that some tasks cannot be delegated to a tool or its vendor and requires adequate security measures wherever confidential or privileged information is used as an input.