Which AI Writing Tools Actually Protect Your Data in Canada
A privacy lead walks through how we actually evaluated AI writing tools in Canada — the questions asked, the vendor answers, and what we picked.
The thing that stopped us wasn't the AI model at all. It was a sub-processor table buried on page four of a vendor's privacy policy, listing a payment processor in Ohio and an email delivery service in Virginia — for a tool whose entire pitch was "built for Canada."
I handle privacy intake for a mid-size firm, and a department head had asked if we could roll out an AI writing tool for drafting client memos and internal comms. Simple enough request. Six weeks later I understood why it wasn't.
What I actually asked the vendors
I didn't start with "is this secure." Everyone says yes to that. I started with where the words go after someone types them, because that question actually has a factual answer.
The list I sent to four vendors, roughly in this order:
- Where does inference run — which country, which data centre region?
- Is customer input used to train the underlying model, and can that be disabled?
- What's in your sub-processor list, and does any of it touch US infrastructure?
- If a US court or agency issued a request, what's your actual exposure, not your marketing position?
- Where does the data sit at rest, and for how long?
- What happens on failover — if your primary region goes down, where does traffic go?
Three of the four answered the first question with something like "enterprise-grade global infrastructure," which is the kind of phrase that means nothing and is clearly written to mean nothing.
Before I sent that list I'd actually drafted a longer one, about fourteen questions, including things like SOC 2 report cadence and whether they'd sign our standard data processing addendum without redlines. Our security reviewer cut it down hard. His point, which I now think was right, was that a vendor's SOC 2 status tells you about operational maturity, not about jurisdiction — and jurisdiction was the thing we couldn't fix with a better contract clause. We kept the DPA question but moved it to the end, after the six above, because by the time you're negotiating addendum language you've already decided whether the vendor is in the running.
The CLOUD Act question, and the one answer that moved us
My counsel had one line item she would not negotiate on: a tool whose parent company is US-incorporated is, as a matter of law, reachable under the US CLOUD Act regardless of where the servers physically sit. That's not a vendor claim, it's the statute — 18 U.S.C. § 2713 extends US compulsory process to data a US-controlled provider holds, wherever it's stored. A Canadian data centre leased by a US company doesn't solve that; the compulsion runs through corporate control, not geography.
That single point eliminated two of the four tools before we even got to pricing. Both were well-known writing assistants, both technically capable, both with a US parent entity. My counsel's phrasing to me, roughly, was that she'd rather deal with a smaller company with less polish than a large one with a jurisdictional problem she couldn't paper over with a DPA.
One of the eliminated vendors offered what they called a "sovereign data" add-on — a paid tier that kept storage in Canada for an extra fee per seat. We asked what it actually changed about the parent company's exposure under US process. The honest answer, once we pushed past the sales deck, was: nothing. Storage location moved, corporate control didn't. We were not sure at first whether that add-on was worth the premium for optics alone, in case a client ever asked where their data sat, but we decided optics without a legal basis wasn't worth paying for.
Under s. 17 of Quebec's Law 25, a business may only transfer personal information outside Quebec after an assessment confirms the information will receive protection equivalent to what it would receive in Quebec. That's the actual clause our Law 25 assessment ran against, and it's the second place the review bit. We have Quebec-resident clients, so any tool processing their information couldn't just clear PIPEDA — it had to clear that equivalency test for cross-border transfer specifically.
What turned out not to matter
I expected the UI to be a deciding factor. It wasn't, really. Every tool we looked at had a workable interface, and the department head testing them cared more about whether it remembered context from a previous session than how it looked. I'd spent a weird amount of prep time on a comparison chart of editing features — tone adjustment, length control, that sort of thing — and none of it came up again after the first meeting. If I did this again I'd skip that chart entirely and spend the time on the sub-processor tables instead.
I also assumed cost would be the controlling variable, since this was going to a department budget, not a firm-wide line item. It wasn't close. The two tools we eliminated over jurisdiction were actually cheaper per seat than what we picked.
Why we looked at a Canadian AI platform at all
Someone on our review committee asked early why we were even considering a Canadian AI option when the US tools were more established. Fair question. My answer, which I think holds up, was that for a firm handling client documents, "established" doesn't offset "US-jurisdiction," and a Canadian AI platform removes an entire category of question we'd otherwise have to keep re-litigating every renewal cycle.
That's roughly how Augure came up. We weren't looking for a Canadian AI vendor specifically — we were looking for an answer to the CLOUD Act question, and a Canadian AI platform happened to be the cleanest way to get one. Augure's answer on data flow was more specific than I expected: customer data stored in Canada, inference running on Canadian infrastructure with EU capacity for certain model tiers and during failover, and a plain statement that customer conversations and documents are never routed to US providers. They were upfront that email delivery and card payment processing involve some US infrastructure — a limitation, just not the one that mattered to our legal analysis, since our concern was customer content, and that part of the answer held up when our reviewer checked it against the privacy policy.
Pricing was straightforward too: C$20 a month per user for the tier with persistent memory and no message caps, which was the plan department leads actually wanted once they saw that free-tier compute allowances would cap out a heavy user inside a week. I'll say plainly I was skeptical a C$20 tier could compete feature-for-feature with the US incumbents, and on raw model benchmarks I'm still not totally sure it does — but that wasn't the axis we were optimizing.
Where I'd push back on myself
I'm not sure we weighted things perfectly. We treated jurisdiction as close to a hard gate and treated everything else — latency, model quality on long documents, how well it handled French-language drafting for the Quebec team — as secondary. My read now is that was roughly the right order for a firm handling client data, but I could see a different organization, one without the same regulatory profile, reasonably landing somewhere else, and I wouldn't argue hard against it.
One thing I'd redo: we didn't test French-language output rigorously enough before deciding, and it came back as a complaint from the Quebec office about three weeks in. Not a dealbreaker, just a gap in how we ran the pilot.
What we picked, and what I'd tell the next person
We went with the Canadian AI platform for anything touching client material, and we kept a separate, cheaper US tool around for pure marketing copy that never includes client identifiers — a split I'm still not entirely comfortable with, administratively, but one the risk profile justified.
If someone at another firm asked me where to start, I'd tell them to skip the feature comparison entirely and ask for the sub-processor table first. Ask who the parent company is before you ask anything about the model. Everything else is negotiable; jurisdiction mostly isn't, and no vendor's marketing page will volunteer the part that complicates their pitch.
We're still mid-renewal on a couple of other tools, and I don't know yet whether the Quebec office's French-drafting complaints will resurface once usage picks up. I'll probably write the follow-up to this piece in six months, not before.
If you're running this same review, Augure's own site has its privacy policy and sub-processor list published, which is the first thing I'd pull up.
About Augure
Augure is a sovereign AI platform for regulated Canadian organizations. Chat, knowledge base, and compliance tools — all running on Canadian infrastructure.
More insights
View all →Building an AI strategy for Canadian education organizations
We Needed AI Tools That Would Survive a Health Canada Audit. Here's the Shortlist
What Canadian Banks Tell Investors About AI — And What It Means for You
Put this to work: Augure Chat, Canadian-hosted AI →