The situation
A commercial dispute is heading to mediation. The other side has produced 40,000 documents, and the examination for discovery of their chief financial officer ran two days. A second-year associate needs a chronology and a list of admissions by Friday. They upload the transcripts and a folder of productions to a general-purpose AI tool and ask for a timeline with page references. It works well.
Nobody on the file thinks of this as a disclosure. The material is being used for the case, after all. But the documents did not come to the firm freely. They were compelled, and the law attaches a promise to them.
What the rules actually say
The common law implied undertaking
In Juman v. Doucette, 2008 SCC 8, the Supreme Court of Canada held that parties to civil litigation give an undertaking to the court not to use documents or answers obtained on discovery for any purpose other than the proceeding in which they were compelled (paragraph 27). The rule covers both documentary and oral discovery, and it applies whether or not the material is confidential in origin. The Court also said a party is not generally free to disclose discovery evidence to strangers to the litigation without a court order.
The rationale is privacy and candour: litigants are compelled to disclose, so the law limits the intrusion to what the case needs. Breach can lead to a stay or dismissal, a struck defence, or contempt proceedings (paragraph 29). The case arose in British Columbia, where the rule applies as common law.
Ontario's deemed undertaking
Ontario codified the rule. Rule 30.1.01 of the Rules of Civil Procedure applies to evidence from documentary discovery, examinations for discovery, inspections, medical examinations and written examinations, and to information obtained from that evidence. All parties and their lawyers are deemed to undertake not to use it for purposes other than the proceeding. The listed exceptions include consent of the disclosing party, evidence filed with the court or referred to at a hearing, and impeachment in another proceeding, and the court can relieve against the undertaking.
Quebec
In Lac d'Amiante du Québec, 2001 SCC 51, the Supreme Court recognized an implied rule of confidentiality for pre-trial examinations under Quebec civil procedure. That decision predates the current Code of Civil Procedure, so confirm with Quebec counsel how the rule applies to your file today.
Where AI fits
No Canadian decision we have found addresses AI tools and the undertaking directly. Using software to analyze productions for the same case looks like a use for the proceeding, much as firms already rely on litigation support vendors under contract. The open question is the provider. If its terms allow it to retain content, use it to improve its services, or train models, the material may be serving purposes beyond the case, held by a party that owes no undertaking to anyone. Productions also often contain personal information about people who are not parties.
This sits alongside the professional duties discussed in our guide to what the Law Society of Ontario says about generative AI, and alongside privilege questions for your own client's documents, covered in how one AI prompt can waive privilege.
Why policies and bans fall short
Most firm AI policies say "no client confidential information". Discovery productions are not your client's information, so staff may not think the rule applies. A transcript of the other side's witness feels like fair game.
Bans do not solve that framing problem, and they push work into personal accounts that the firm cannot see. IBM's 2025 Cost of a Data Breach Report found that 63% of organizations have no AI governance policy at all. When a document leaves the firm this way, it cannot be recalled, and the firm may face an uncomfortable conversation with opposing counsel or the court.
What a practical control looks like
- Name discovery material in your AI policy. State that productions, transcripts and information derived from them are covered by the undertaking and may only go into tools the firm has approved for that purpose.
- Approve a tool on the right terms. Look for contractual commitments on no training, limited retention, confidentiality and deletion, so the provider is closer to a litigation support vendor than to a public service.
- Address AI in the discovery plan. Where it matters, agree with opposing counsel, or seek a term in a protective order, on which tools may process productions.
- Train the litigation team on the difference between your client's documents and compelled documents, with a short example.
- Set an incident path. If productions reach an unapproved tool, the responsible lawyer assesses whether the court, opposing counsel, or the client needs to be told, with advice from your practice advisors or counsel.
Sanitized Ai is a browser extension that supports steps 1, 4 and 5 where the risk begins. When someone is about to submit client names and identifiers, personal information or financial data to an AI assistant, it redacts or blocks that content before submission and explains in plain language what was flagged and why. It does not check the accuracy of AI output, citations or legal analysis. Its job is keeping sensitive material from becoming subject to a provider's terms.
Administrators see a dashboard of flagged-event metadata (which tool, what type of data, which policy, when), never prompt content. That record can help show reasonable safeguards if a question arises later, though it does not decide how a court would view any particular use. See how it fits law firms.