What Canadian Federal and Provincial Crown Counsel Need to Know About AI in Evidence Review

Crown counsel occupy an awkward position in the AI conversation. The volumes are as large as anything in commercial litigation, the disclosure obligations are heavier, and the tolerance for error is close to zero. At the same time, public sector litigators cannot simply sign up for a tool because it looks useful. Procurement rules, departmental IT security review, and a growing body of federal guidance on the responsible use of artificial intelligence all sit between the idea and the deployment.

The result is a familiar stall. Everyone agrees that manual review of a several hundred thousand document production is neither proportionate nor realistic. Nobody wants to be the person who authorized an AI tool without a defensible answer to the questions that will follow. What follows is a practical map of those questions, and of the answers that already exist.

The disclosure obligation is the starting point, not the technology

Crown counsel do not adopt review technology because it is modern. They adopt it because disclosure obligations are constitutional in criminal matters and demanding in civil ones, and because those obligations must be met within finite budgets and statutory timelines.

This reframing matters. When a review technology decision is presented to a departmental approver as an innovation initiative, it invites scrutiny about novelty and risk appetite. When the same decision is presented as a proportionality measure required to meet an existing legal obligation, it sits in far more familiar territory. The Sedona Canada Principles already supply the vocabulary. We have written before about how Sedona Canada now makes supervised AI review the defensible baseline rather than the risky departure, and that analysis applies with equal force to public sector litigants. Proportionality and cooperation are not private sector conveniences. They are the standards a court will apply to the Crown as readily as to anyone else.

The practical consequence is that the question facing a Crown litigation team is not whether AI review is permitted. It is whether the specific tool, and the specific workflow around it, can be explained and defended.

What the Directive on Automated Decision-Making does and does not cover

Federal counsel will inevitably encounter the Treasury Board Directive on Automated Decision-Making in this discussion, often raised as an objection. It is worth being precise about its scope, because imprecision here kills good projects.

The Directive governs automated systems used to make or assist administrative decisions about clients, the kind of decision that determines a person's rights, privileges, or entitlements. It imposes Algorithmic Impact Assessment requirements, notice obligations, and explanation requirements calibrated to the impact level of the system.

AI-assisted document review does not obviously fit that description. A tool that reads a document and proposes a relevance or privilege call for a lawyer to accept or reject is not making an administrative decision about a client. It is a decision support tool operating entirely inside counsel's own analytical process, in the same conceptual category as a search index or a keyword filter, albeit a considerably more capable one. The lawyer remains the decision maker, and the lawyer's judgment is what the file records.

That said, you should confirm this analysis with your own institution's AI governance function rather than take it from a vendor. The safer posture is to treat the Directive as a statement of the values your department expects, transparency, accountability, human oversight, and auditability, and then demonstrate that your review workflow satisfies those values regardless of whether the Directive strictly applies. A workflow that would survive an Algorithmic Impact Assessment is a workflow you will never have to apologize for.

Where the data goes is the question that stops most procurements

In our experience, the single question that ends more public sector evaluations than any other is the one about data residency. Crown counsel handle material that is protected, sometimes classified, frequently subject to statutory confidentiality, and always politically sensitive. A tool that ships documents to an American data centre, or that uses departmental content to improve a commercial model, is not going to clear security review, and it should not.

This is where a clear architectural answer replaces a long negotiation. Claira's Canadian deployment runs in Google Cloud's northamerica-northeast1 region in Montreal, with no cross-region replication and no cross-region egress of customer content. Model requests are pinned to endpoints in the same region. Customer content is processed per request and is not used to train Claira-owned models, and operational events are logged to support audit and troubleshooting. Those specifics, along with the encryption and access control posture, are documented in the Claira privacy and security reference rather than buried in a sales conversation, which is precisely what a departmental security assessor needs.

The point is not that any one architecture is uniquely virtuous. The point is that these questions have concrete, documented answers, and that a procurement conversation moves quickly once they do.

Plugin architecture is the procurement shortcut nobody uses enough

Public sector technology stacks are approved slowly and changed reluctantly. Many federal and provincial legal services already run Nuix Discover. That existing approval is an asset, and it is routinely underused.

A tool that operates as an extension inside Nuix Discover, rather than as a separate platform that documents must be exported to, presents a materially smaller procurement and security surface. There is no new document repository to assess. There is no additional copy of the evidence to track, secure, and eventually destroy. The chain of custody stays where it already is. Reviewers work in the interface they have already been trained on, and the coding they produce lands in the Nuix fields the case team already relies on.

Framing an AI review capability as an addition to an approved platform rather than as a new platform is not a rhetorical trick. It is an accurate description of a genuinely narrower risk profile, and it is the framing most likely to survive contact with a departmental architecture review board.

Human oversight is a feature, not a concession

Every credible framework for public sector AI use, and every law society guidance document a Crown counsel is bound by, converges on the same requirement. A human must remain accountable for the output.

This is not a limitation you must reluctantly design around. It is the workflow that actually works. Claira reads the full document set and proposes coding with reasoning attached. Counsel reviews, accepts, corrects, and retains the record of having done so. The AI carries the reading. Counsel carries the judgment, and the file reflects it.

Where to start

Start narrow. Pick one closed or low-sensitivity matter, run a supervised review against coding your team has already completed, and measure agreement. That exercise produces exactly the artifact a departmental approver wants, which is evidence from your own environment rather than a vendor's claim.

If you would find it useful to work through the security and procurement questions specific to your department, book a working session with us. We have had this conversation before, and it goes faster than you expect.