At scale the constraint is not reading speed, it is defensibility. AI can prioritise, de-duplicate and code a first pass; a lawyer still swears the affidavit.
Key takeaways
Running document review with AI starts with deciding which review you are running. In litigation you are producing a list that a party swears to, and the standard is defensibility to the other side and the court. In a transaction you are finding the provisions that change the deal, and the standard is whether you missed something that later surfaces. The technology overlaps; the obligations do not.
For the litigation case, the output is prescribed. Under rule 223 of the Federal Courts Rules, an affidavit of documents must contain separate lists and descriptions of all relevant documents in the party’s possession, power or control for which no privilege is claimed; those for which privilege is claimed; those no longer in the party’s possession; and those believed to be held by a non-party. It must state the grounds for each claim of privilege, and it must include a statement that the party is not aware of any other relevant document. That last requirement is why an unverified machine classification cannot be the end of the process.
Provincial rules differ, and Ontario practitioners should work from the Ontario framework — Treadstone Law sets out documentary discovery and the affidavit of documents in Ontario, and separately what documentary discovery requires you to produce. Electronic records are squarely within it: see whether emails and texts must be produced.
De-duplication and threading. The first meaningful reduction in a large email collection is mechanical: exact and near-duplicates removed, email chains rolled up so the longest inclusive message is reviewed once. This is not new technology and it is not controversial; it is also where most of the volume goes.
Prioritisation. Ranking documents by likely relevance so that reviewers see the productive material first changes the economics of a review. It does not decide anything — every document still has a disposition — but it front-loads the findings so the legal team can shape strategy in week one rather than week six.
First-pass issue coding. Tagging documents against an agreed issues list, with a confidence indicator, so human review is spent confirming and correcting rather than sorting from zero.
Entity, date and event extraction. Building a chronology and a cast list from the collection is tedious, error-prone by hand, and exactly the sort of task that scales.
Privilege candidates. Flagging documents that involve counsel, or that discuss legal advice, so they route to a privilege reviewer. Note the word candidates.
A tool can surface the documents that look privileged. It cannot decide whether privilege applies, whether it has been waived, or how to describe the grounds — and rule 223 requires the grounds for each claim to be stated. Those are legal conclusions with consequences, and they belong to a lawyer who has read the document.
The practical arrangement that works: the tool routes candidates, a lawyer decides each one, and the decision is recorded with the reasoning so the log can be defended months later. Confidentiality obligations under rule 3.3-1 of the Model Code run through the whole exercise, including the choice of hosting environment for the collection.
The question you will be asked is not whether you used technology. It is what your process was and how you know it worked. A defensible protocol has four parts: a written description of the method, agreed with the other side where the rules contemplate a discovery plan; a measured sample of the documents the process excluded, reviewed by a person; a record of what that sample found; and a remediation step if it found too much.
Ontario practice contemplates agreeing the approach in advance — see Treadstone Law on what a discovery plan is and whether you need one. Agreeing the method before the review, rather than defending it afterwards, is the whole point. Documentary discovery obligations themselves are set out in rule 222 of the Federal Courts Rules for federal matters.
A protocol that survives scrutiny
A two-year email collection from six custodians arrives in a commercial dispute. Manual review of everything is disproportionate to the amount in issue, which is precisely the argument for a technology-assisted process rather than an excuse for one.
The sequence that works: de-duplicate and thread first, which removes the bulk without any judgment being exercised. Agree the issues list and the method with opposing counsel. Have senior counsel code a seed set personally — this is the step firms skip and regret, because the model learns whatever the seed set teaches it. Run prioritisation, review in descending order, and stop when the rate of new relevant material falls and a sample of the unreviewed remainder confirms it.
Route every document involving counsel to a privilege reviewer regardless of rank. Then draft the affidavit from the log, and have the deponent understand what the process was before they swear it. The saving is real and it comes from ordering and reduction, not from anyone skipping the reading.
Where AI-generated content ends up in materials filed with a court, check that forum’s own notices and practice directions first — Canadian courts have addressed this separately and their requirements are not uniform. Whatever the forum, the verification duty is yours: the Office of the Privacy Commissioner of Canada’s generative AI principles expect organisations to evaluate the validity and reliability of a tool for its intended purpose, and providers to disclose contexts in which a system may produce incorrect information. Rule 3.1-2 of the Model Code puts the competence obligation on the lawyer either way.
Is technology-assisted review accepted in Canada?
Proportionality is well established in Canadian civil procedure and technology-assisted processes are used in large matters, but what is accepted in your matter depends on your forum, your rules and what the parties agree. Raise it in the discovery plan rather than deciding unilaterally.
Can we use a general chatbot instead of a review platform?
Not for a production. You need per-document audit trails, defensible sampling and controlled hosting. A chat window gives you none of those.
Who signs the affidavit?
The party, on the basis of a process their lawyers can explain. That has not changed, and it is the reason none of this is automation in the sense clients sometimes imagine.
A 30-minute call is enough to tell you whether AI pays for itself here.
