AI reads the data room. It does not run the searches, and it does not decide what goes into the disclosure schedule — which is where the deal risk actually sits.
Key takeaways
AI is used in Canadian legal due diligence almost entirely on one part of the job: reading the contracts in the data room and pulling out the provisions that change the deal. It does not run the corporate, security or title searches, and it does not decide what belongs in the disclosure schedule. Those two facts shape everything below.
The deliverable on a buy-side file is usually a due diligence report organised by risk, a set of issues that need to be fixed before closing or priced into the deal, and input into the representations, warranties and disclosure schedules. Treadstone Law’s overview of what to investigate when buying an Ontario business is a good statement of the scope, and the structure of the deal drives it: an asset purchase and a share purchase produce very different diligence lists.
Registry work is verification against an official source, and the output has to be a search result, not a summary of one. Federal corporate status is confirmed through Corporations Canada, which also issues certificates of compliance and existence and holds annual return filings. Ontario entities are searched through the Ontario Business Registry, which took over from the previous system in October 2021 and offers search products such as a profile report; note also that corporations must file an annual return through the registry each year within six months of fiscal year end, and that changes must be reported within 15 days — both of which produce diligence findings when they have not been done.
Land holdings are confirmed through the provincial land registration system. In Ontario, historical land registry records are searched through Ontario Land Registry Access, with current title work done in the electronic registration system. Personal property security registrations, court records and licence status each have their own official source. A generative tool asked what encumbrances exist will answer; the answer is not a search result and cannot go in a report.
Registry work: the rule
Contract triage across the data room. Sorting several hundred agreements into material and immaterial, by type, counterparty and value, so the team knows what it is dealing with on day one rather than day five.
Change of control and assignment extraction. This is the highest-value extraction on almost every deal. Which agreements terminate, or give the counterparty a right to terminate, on a change of control? Which require consent to assign? Which are silent? Output should be a table with quoted clause text and clause numbers, because every entry will be checked. The same question arises on premises: see Treadstone Law on assigning an Ontario commercial lease and on the lease terms a buyer should review carefully.
Other deal-relevant provisions. Exclusivity, most-favoured-nation clauses, minimum-volume commitments, non-competes, unusual termination rights, uncapped liability, automatic renewals about to trigger.
Consistency across a document set. Where the same form was used a hundred times, finding the twelve that were amended is a mechanical task. Those twelve are usually the story.
First-draft summaries. A one-paragraph description of each material contract for the report annex, drafted from the extracted text and then checked.
Materiality is the first judgment: a change-of-control right in a small supply agreement may be noise, and the same right in the agreement with the customer representing most of the revenue may be the deal. Only someone who understands the target’s business can tell those apart.
The second is what to do about it — consent, waiver, price adjustment, indemnity, or walk. The third is the disclosure schedule, which is a drafting exercise with direct legal consequence and belongs entirely to the lawyer. And where a shareholders agreement has to be unwound or amended as a condition, that is its own workstream: see how a shareholder agreement is changed in Ontario.
Rule 3.1-2 of the Model Code of Professional Conduct requires legal services to be performed to the standard of a competent lawyer, with commentary expecting an understanding of the benefits and risks of relevant technology; your own regulator’s code governs, such as the Law Society of Ontario’s Rules of Professional Conduct in Ontario.
A buyer signs a letter of intent for an Ontario manufacturer. The data room contains around four hundred documents, and the diligence window is short.
Day one is triage: the tool sorts the room by document type and counterparty, and produces the first extraction table on change of control and assignment. The team immediately knows which thirty agreements matter. Day two, a lawyer reads those thirty in full against the extraction and corrects it — on this kind of file the corrections cluster in scanned amendments and in agreements that incorporate standard terms by reference.
In parallel, and entirely outside the tool, corporate searches are ordered from Corporations Canada and the Ontario Business Registry, land and security searches are run, and the minute book is reviewed for share issuances and resolutions. The report is written by the lawyer, from verified inputs. The tool bought a faster route to the thirty documents that mattered; it did not shorten the part where somebody had to be right. For a commercial property in the deal, the Ontario commercial real estate diligence checklist sets out the parallel workstream.
Data-room material is confidential by definition, frequently subject to a non-disclosure agreement, and often contains personal information about the target’s employees. Rule 3.3-1 of the Model Code requires strict confidence; PIPEDA’s fair information principles apply to the personal information within it; and the Office of the Privacy Commissioner’s generative AI principles expect organisations to establish that a tool is necessary and proportionate and to know whether prompt content is used for further training. Confirm the NDA permits processing in the tool you intend to use before a single document is uploaded.
Can AI write the due diligence report?
It can draft the descriptive annexes from verified extractions. The risk assessment, the recommendations and the executive summary are advice.
What about vendor due diligence?
Same tooling, different objective: you are finding what a buyer will find, early enough to fix it. The extraction work is if anything more useful, because you control the timetable.
How much checking is realistic?
Check every entry you rely on in the report. Extraction with quoted text and clause numbers makes that fast, which is exactly why the output format matters more than the model.
A 30-minute call is enough to tell you whether AI pays for itself here.
