How Harvey AI's next generation deals with Canadian data

Harvey ii, which legal artificial intelligence provider Harvey rolled out a few weeks ago, addresses a basic limitation of AI tools used in legal work: they tend to forget everything between sessions, forcing lawyers to re-explain context every time they open a new query.

The first fix, memory, operates at the individual user level, storing details like a lawyer's practice area and preferred response format; users can accept, reject, edit or delete what gets remembered. The second, spaces, organises the platform around individual client matters rather than individual queries, pulling in relevant documents, playbooks and task lists tied to a specific file.

Joe Marando, Harvey's Toronto-based legal innovation partner, framed the governance logic behind both features as the more important design choice than the features themselves.

Law firms operate under "ethical wall" rules that restrict which lawyers can see which client files, so both memory and spaces are scoped narrowly to keep information from leaking across those walls. That's a real constraint: a firm cannot afford an AI feature that exposes one lawyer's client file to a colleague who's supposed to be walled off from it.

"For a lot of folks, they might want to use Harvey to build out a knowledge base of everything relevant for the matter that they're working on, regardless of what practice area they're in. For a litigator, having that knowledge base sit in a space and being able to query that can really help them bring their best on the day of court. So, for example, you know, might help you generate some ideas with questions to come up with witnesses, and that's particularly helpful if there's an expert witness in you know an area that you don't know personally," said Marando.

The Canadian question of data storage and retention

Canadian law firms working with Harvey, the legal AI platform, can choose to store their data inside Canada's borders.

Anita Gorney, Head of Privacy and AI, Legal at Harvey, said Canadian customers can choose which country stores their data, including in Canadian Microsoft data centres, but when data needs to be processed, it has to leave the country and be routed through OpenAI infrastructure in the United States, the European Union, or Australia.

Harvey uses Microsoft Azure as its primary cloud infrastructure, while OpenAI supplies key foundation models through Azure OpenAI.

"What that really means is when we're sending the inference to our AI providers to process the inference and the customer data, we're using their API endpoints that are connected to specific locations," she said. "We do use OpenAI models, and that is served either directly through OpenAI, so first party or third party on Microsoft Azure. The other models that we make available in the tool are Anthropic's models and also Gemini. There is a model selector, and customers have the choice of model."

Verification is still the lawyer's responsibility

Marando said Harvey addresses hallucination by attributing "sources" that ground the tool's output in citations a user specifies, returning pinpoint links back to primary case law rather than letting the model draw solely on its training data.

Canadian courts have not been immune to AI-fabricated citations. Zhang v. Chen, a 2024 B.C. Supreme Court decision, became the country's most-cited example after a lawyer submitted two fabricated case citations generated by ChatGPT in a child travel application; the court ordered her to personally cover the opposing side's costs of tracking down the fakes.

A similar episode surfaced in Ontario's Ko v. Li in 2025, where a factum cited case law that either didn't exist or was linked to the wrong decision entirely. 

Marando described Canadian law firm buyers as uniquely rigorous: stress-testing vendor claims and evaluating not just whether a tool meets today's needs but whether the company will keep developing it fast enough to stay useful six months from now.

"I'd say they're very savvy buyers, and sort of keep us innovating from that perspective," said Marando. "The sociological bit is that they're also very interested in making successful adoption, which doesn't just look like number of logins. It's what value is this tool delivering for the firm, and then ultimately the firm's clients or for in-house team."