AI in Québec notarial practice: what a notary can delegate, and what they never can.
A Québec notary may use artificial intelligence on one condition: that it prepares the work without ever concluding it. Research, reading the file's documents and flagging anomalies can be delegated; the duty to advise, legal characterization and the signature cannot. This guide draws the line.
ATOM Solutions · August 20, 2026 · 12 min read
What Québec's professional framework already requires
No Québec rule forbids a notary from using artificial intelligence. The obligations that govern its use are not new either: they are the ones already governing the whole practice, and they apply to a tool exactly as they apply to staff.
The useful question is therefore not whether a notary may use AI, but on what terms. Those terms follow from existing obligations, with no need to wait for dedicated legislation: the duty to advise forbids letting a tool conclude, competence requires knowing its limits, professional secrecy demands knowing where the data goes.
Two points are worth stating up front. Liability is not shared with the software vendor: a defective act engages the notary, whatever the assistance received. And the argument that the tool said so carries no weight anywhere — not with a client, not with the syndic, not before a court.
This guide is informational and does not constitute legal advice. The provisions cited link to LégisQuébec and CanLII: verify them at the source. For the exact scope of your obligations, refer to the Chambre des notaires du Québec.
Public officer and duty to advise
The Notaries Act (RLRQ c. N-3) makes the notary a public officer bound to act impartially and to advise all parties to the act fairly. That duty is personal: it cannot be subcontracted to software, or to anyone.
Competence and diligence
The Code of ethics of notaries (RLRQ c. N-3, r. 2) requires practising with competence. Using a tool whose limits and answer provenance you do not understand breaches that requirement — it does not delegate it.
Professional secrecy
Protected by section 9 of the Charter of human rights and freedoms (RLRQ c. C-12). Feeding a client's file into a service that stores or reuses the data raises a question that comes before answer quality.
Personal-information protection
Law 25, amending the Act respecting the protection of personal information in the private sector (RLRQ c. P-39.1), governs communication outside Québec and requires a prior assessment of protection factors.
The dividing line: prepare, never conclude
One question settles every task: does this work produce a verifiable finding, or a legal decision? The first can be delegated and checked; the second belongs to the notary, entirely.
The boundary is not technical, it is legal. A tool that reconstructs a chain of title does reading work: it reports what the register contains. The same tool declaring a title good and marketable would be doing characterization work — standing in for the public officer. The machine's capability is identical in both cases; the legitimacy is not.
That line has a practical consequence for tool selection: prefer the ones that make their work checkable over the ones that make their answer convincing. A finding that is cited, dated and graded by severity is verified in seconds; a fluent, sourceless synthesis forces you to redo the research before you can sign anything.
| Task | Delegable to AI? | What stays with the notary |
|---|---|---|
| Finding the applicable provision | Yes, if every answer cites its official source | Reading the cited text, confirming it is in force, judging its relevance |
| Reading and summarizing file documents | Yes — extraction, indexing, spotting dates and parties | Checking the summary against the original documents before relying on it |
| Reconstructing a chain of title | Yes, link by link, with breaks flagged | Concluding on the validity of the title and stating any reservations |
| Detecting a missing clause or a stale reference | Yes — that is a comparison, not a judgment | Deciding whether the gap is a defect, a deliberate adaptation or a non-issue |
| Drafting a proposed act | Yes, as a marked draft | Rewriting, characterizing, arbitrating clauses and owning the final text |
| Legally characterizing a situation | No | The analysis and the characterization, in full |
| Advising the parties | No | The duty to advise, personal and impartial toward each of them |
| Receiving and signing the act | No | Reception, signature and retention as a minute |
Five concrete risks, and how to defuse them
The incidents seen in legal practice come down to five specific defects. Every one of them can be checked before adopting a tool — and none of them can be fixed afterwards.
The invented citation
A language model produces plausible text, not necessarily true text: it can fabricate a section number or a decision that does not exist. Defusing it: accept only a tool where every statement links to an official text openable in one click — then click.
Stale law
A repealed or replaced provision lives on in training corpora, sometimes for years. Defusing it: require the tool to display the date and in-force status of what it cites, and distrust any undated answer.
The leaking file
Many consumer services retain conversations and use them to improve their models. Defusing it: get the model provider's retention policy in writing, not merely the policy of the vendor reselling it.
Leaving Québec
Processing hosted outside Canada triggers the outside-Québec communication analysis required by Law 25. Defusing it: ask for the exact processing region — a country is not an answer, a region is.
The automated decision
Section 12.1 of the Act respecting the protection of personal information in the private sector requires informing the person concerned when a decision is based exclusively on automated processing. Defusing it: keep a human decision at every step, which the ethical rules require in any event.
How NotarIA applies that line
NotarIA is built on this division, not merely compatible with it. Every legal answer cites its source — Civil Code of Québec, LégisQuébec, Justice Canada, CanLII — and every finding is graded by severity rather than presented as a conclusion. Every document produced comes out marked as a draft to validate: the tool never finalizes, signs or sends anything on its own.
On the data side, model calls stay in a Canadian region, in zero-retention mode: no client file trains a model. Text recognition and dictation are self-hosted, with no network egress. Every access, generation and validation is written to an immutable audit log — enough to answer, years later, the only question that matters then: who verified what, and when.
The limits are stated as plainly as the capabilities. There is no direct Land Register access: documents are uploaded or entered. The AI never declares a title good and marketable. There is no electronic signature. And the published performance figures come with their protocol, their corpus and their limits, on the methodology page — a figure without a denominator is worth nothing.
Governing AI in a practice: what to write down
Governed use fits in a short document and a few shared habits. Here is what an internal policy should say — and what a client, an insurer or the syndic might one day ask to read.
The approved tools, by name
A closed list, each with its permitted and forbidden use. A tool absent from the list is not approved by default: it is the only principle that holds up over time.
What never leaves the file
Name the categories that must never be entered into an unapproved service: identity documents, financial statements, testamentary dispositions, and more broadly any information identifying a client.
The verification rule
No AI statement reaches an act or a piece of advice until its source has been opened and read. The rule must stay simple to survive a busy Tuesday afternoon.
Who decides, and where it is written
The name of the notary accountable for each validation, and the precise place in the file where that validation is recorded. A validation that leaves no trace did not happen.
What clients are told
How the practice informs its clients of its use of AI. Chosen transparency always costs less than the same information discovered by a third party.
The steps at a glance
Inventory actual usage
List what the practice already does with AI tools, including informal use on personal accounts. You cannot govern what you have not mapped, and the real footprint is almost always wider than assumed.
Pick the tools and write the policy
Settle on a closed list of approved tools, the permitted use for each, the data categories that must never be entered, and the person accountable. Two pages is enough; a document nobody reads protects nobody.
Verify residency and retention
Get the processing region and the retention policy in writing from the model provider, not only from the software vendor. File those answers in the practice's personal-information register.
Test the tool on a real file
Take a closed file whose traps you already know and compare: what the tool flags, what it misses, what it invents. Click every citation through to the official text.
Train the team to challenge an answer
The skill to build is not prompt writing, it is doubting an answer: open the source, check the in-force date, notice what is missing.
Record and revisit
Log who validated what in the file, and reread the policy at least once a year. The tool catalogue and the applicable law both move, rarely at the same time.
Frequently asked questions
Can a Québec notary use artificial intelligence?
Yes. No Québec rule forbids it, and AI already ships inside everyday office software. The applicable obligations are not new: competence, duty to advise, professional secrecy, personal-information protection. In practice a tool may prepare, extract, compare and flag; legal characterization, advising the parties and the signature stay with the notary. For your own situation, refer to the Chambre des notaires du Québec.
Can ChatGPT be used for notarial legal research?
A consumer assistant raises two separate problems. Confidentiality first: exchanges may be retained and used to improve the service, which sits poorly with professional secrecy. Verifiability second: without an openable, dated citation, nothing separates a grounded answer from a plausible invention. A professional tool is judged on those two points — documented retention and clickable sources — before it is judged on its prose.
Can AI draft a notarial act?
It can produce a draft, never an act. The difference is not one of form: a notarial act is received by a public officer who owns its content, has explained it to the parties and keeps it as a minute. Generated text must therefore be marked as a draft, reread, rewritten and owned. NotarIA marks every document it produces as a draft requiring a notary's explicit validation.
What does Law 25 say about automated decisions?
Section 12.1 of the Act respecting the protection of personal information in the private sector (RLRQ c. P-39.1) provides that an enterprise making a decision based exclusively on automated processing of personal information must inform the person concerned no later than at the time of the decision, and allow them to submit observations. In a notarial practice the answer is structural rather than procedural: no decision should be exclusively automated.
How do you check that an AI answer is not invented?
Three moves, in this order. Open the cited source and confirm the text exists. Check its date and status: a provision may have been amended or repealed since. Read the full excerpt rather than the quoted sentence alone — that is where the exceptions hide. A tool that does not allow those three moves in seconds is not usable in professional practice.
Must clients be told that AI is used?
No general disclosure obligation arises from merely using an assistive tool, just as none arises from using drafting software. Two situations change that: a decision based exclusively on automated processing (s. 12.1 P-39.1), and communication of personal information outside Québec, which requires a prior assessment. Beyond the legal minimum, a clear mention in the practice's privacy policy is a prudent choice.
Test the line on your own files.
Request a demo: bring a real act, click every citation through to the official text, and judge what the tool flags — and what it leaves you to decide.
Assistant — not legal advice. The notary reviews, signs and decides on every act.