Contract review is a deliverable with a client waiting on it. AI can do the reading and the first pass. The advice, the risk call and the signature stay with the lawyer.
Key takeaways
Canadian firms put AI into contract review in four places: pulling the clauses out of the document, comparing them against the firm’s own positions, extracting the dates and obligations, and drafting the first version of the issues list. Everything after that — what is material, what to concede, what to advise — is the lawyer’s work and does not move.
It helps to be precise about what the client is buying. A contract review file is rarely a marked-up document on its own. It is usually a short issues list ranked by risk, a redline carrying the firm’s preferred language, and a covering note that says which points are worth fighting for. The document reading is the input; the ranking and the recommendation are the deliverable.
That split is why the technology lands unevenly. Reading 80 pages and finding every assignment clause is a mechanical task that scales badly with human attention and well with software. Deciding whether an uncapped indemnity is acceptable in this deal, for this client, at this price, does not scale at all — and should not.
Clause extraction and classification. Given a signed agreement, current tools will locate and label the recurring provisions — term and renewal, termination, assignment and change of control, limitation of liability, indemnity, governing law, confidentiality, dispute resolution. Output is a table with clause references, which a lawyer reads far faster than the contract.
Comparison against a playbook. If the firm has written down its standard and fallback positions, a tool can mark each clause as at-standard, off-standard or missing. This is the highest-value use because it converts an experience-dependent task into a checkable one, and because it makes junior work reviewable at a glance.
Obligations, dates and triggers. Notice periods, renewal windows, reporting duties, price-adjustment mechanics and conditions precedent can be pulled into a single schedule. For a client with a portfolio of agreements, that schedule is often worth more than the review itself.
Internal consistency. Defined terms used but never defined, cross-references that point at the wrong section, numbering that breaks after an amendment, two clauses that contradict each other. These are the errors that survive human proofreading and that software catches reliably.
The professional obligations do not change because a tool was involved. Under the Federation of Law Societies of Canada’s Model Code of Professional Conduct, rule 3.1-2 requires a lawyer to perform all legal services to the standard of a competent lawyer, and the commentary to that rule says a lawyer should develop an understanding of, and ability to use, technology relevant to their practice — including the benefits and the risks, and recognising the duty to protect confidential information. Competence in this area is not optional extra credit; it is part of the competence rule itself.
The Model Code is a national template. The rules that bind you are your own regulator’s — in Ontario, the Law Society of Ontario’s Rules of Professional Conduct, with equivalent codes in every other province and the Barreau du Québec for Quebec practice. Check your own jurisdiction’s wording before you rely on a rule number.
Model Code rule 6.1-1 puts complete professional responsibility for all business entrusted to a lawyer on that lawyer, and requires direct supervision of staff and assistants to whom tasks are delegated. Software is not an assistant within the meaning of that rule, but the principle underneath it is exactly the one that matters here: the file is yours, the output is yours, and an unchecked extraction table that reaches a client is your error, not the vendor’s.
The line, stated plainly
Model Code rule 3.3-1 requires a lawyer to hold in strict confidence all information concerning the business and affairs of a client acquired in the course of the professional relationship, subject to narrow exceptions. Putting a client’s unsigned agreement into a consumer chatbot is a disclosure decision, and it should be made deliberately rather than by an associate at 11pm.
The practical questions are short. Where is the document processed and stored? Is the content of prompts used to train or refine the provider’s model? Who at the vendor can see it? How long is it retained, and can you delete it? The Office of the Privacy Commissioner of Canada’s principles for responsible, trustworthy and privacy-protective generative AI put the same points in privacy terms: providers should tell organisations when information from prompts is used for further training, and organisations should establish that the tool is necessary and proportionate for the purpose rather than merely useful. Where a contract contains personal information, PIPEDA’s fair information principles apply to it in the tool exactly as they apply to it in your document management system.
Firms that have settled this well tend to have done one boring thing: written a one-page rule that names the approved tools, states what may be put into them, and says what must never be. The Law Society of Ontario’s practice supports and resources are the place to start for Ontario licensees looking for practice-management material.
The single biggest determinant of whether contract review AI is useful at a firm is whether the firm has written down its positions. A playbook does not need to be elaborate: for each recurring clause, the standard position, the acceptable fallback, the point at which a partner must be consulted, and the preferred language for each. Twenty clauses covers most commercial work.
Without it, the tool has nothing to compare against and will produce generic observations that a competent articling student could have written. With it, the same tool produces a marked-up contract against your firm’s house standard — which is a genuinely different product, and one clients notice.
A manufacturing client sends 60 supplier agreements ahead of a renewal cycle and asks which ones expose it if a supplier is acquired. Historically this is a week of associate time and the answer arrives after the client has already started renegotiating.
Run through a tool with an extraction task rather than an advice task, the first pass produces a table: for each agreement, the assignment and change-of-control language quoted verbatim with its clause number, the termination rights, the notice period, and a flag where no change-of-control provision was found. That table is generated in an afternoon and is wrong in places — scanned agreements and hand-amended schedules are where extraction fails.
The lawyer then does the part that matters: checks every flagged agreement against the source document, reads the ones the tool marked as silent (silence is the risky answer, so those get human eyes first), ranks the exposures, and writes the note. The client gets a ranked answer in days rather than weeks, and the file still contains a lawyer’s judgment on every agreement that mattered.
Two guardrails made that work. The extraction prompt asked for quoted text and clause numbers, not conclusions — so every line is checkable against the contract in seconds. And nothing was reported to the client that a lawyer had not opened the underlying agreement to confirm. For background on the clauses that usually drive this analysis, Treadstone Law’s note on key clauses in an Ontario service agreement and its explainer on the difference between an NDA and a confidentiality clause are useful starting points.
Can AI do the review and a lawyer just sign it off?
No, and framing it that way is where firms get into trouble. Sign-off on unread output is not review. The workable model is that the tool produces checkable artefacts — quoted clauses with references — and the lawyer verifies and then advises.
Do we have to tell the client we used it?
That depends on your regulator, your retainer and the nature of the work, and the cleanest place to settle it is the engagement letter rather than mid-file. Where the work product is destined for a court or tribunal, read that forum’s own published notices and practice directions before filing — they are not uniform across Canada.
What breaks most often?
Scanned and photographed documents, hand-amended schedules, and agreements incorporating other documents by reference. If a clause lives in an appendix that was never sent to you, no tool will find it — and it will not tell you it is missing unless you ask it to report gaps explicitly.
A 30-minute call is enough to tell you whether AI pays for itself here.
