AI agent or AI model: what notaries need to know
A model answers, an agent acts. The difference changes what you need to check: what the tool does on your behalf, what it spends, how its errors spread, and what remains your responsibility.
Read first. This guide explains general concepts. It is not legal advice and does not replace the standards of the Chambre des notaires du Québec.
"Agent" has become the buzzword of artificial intelligence software. Behind the label lies a real difference in nature from assistants such as ChatGPT, Claude or Gemini. For a notary, it mainly changes what needs to be checked before trusting the tool.
A model answers, an agent acts
An AI model receives a text and produces an answer. It works only on what you give it and on what it learned during training. Ask it what the 30-year period of a title search is: it explains it. Ask it who owns lot 4 523 888: it does not know, and if it answers anyway, it is making it up.
An agent is a model that has been given tools and a mission. It decides the next step itself, uses its tools, reads the result, then starts again until the mission is complete. The tools can be a browser, access to a register, a download or an email.
Take a title search. An agent can:
- sign in to the land register with your credentials;
- open the property index of the target lot;
- keep the deeds that matter for the requested period;
- download and read them;
- notice that the lot comes from a cadastral renewal, and go back to the former lot;
- start again until the period is covered.
None of these steps was dictated one by one: the agent chained the decisions together. That is what makes it valuable. It does in a few minutes work that takes several hours. And that is what calls for new precautions.
| Model (assistant) | Agent | |
|---|---|---|
| Starting point | Your question and your documents | A mission ("search this lot, 30 years") |
| Access to the outside world | None, or a web search | Tools: register, downloads, etc. |
| Who decides the steps | You, question after question | The agent, within the limits set for it |
| Possible spending | Your subscription | That of the services it uses on your behalf |
| What to check | The answer | The answer and the path that led to it |
Five points to check before letting an agent act for you
1. What it does on your behalf, and what it costs
An agent that consults the land register uses your credentials, and each consultation is billed to your account. Before launching it, you need to know:
- which actions it can take, and which it will never take;
- what each action costs, and how the total is shown to you while it works;
- how to stop it: a cancel button that really stops the spending, without waiting for a long step to finish.
2. Traceability: being able to retrace the path
With a model, you see the question and the answer. With an agent, dozens of steps happened in between. A good agent keeps a log: what it consulted, what it kept, what it set aside and why, what it paid. Without this log, you can neither check its work nor explain it to a client or to the syndic.
3. Errors that spread
A model's error stays in its answer. An agent's error can steer everything that follows. A registration number misread on a handwritten index leads to downloading the wrong deed. If the agent carries on from there, the whole chain of title rests on a false reading.
Ask how the tool protects itself: does it check that the downloaded deed really concerns the lot being examined? Does it stop when a reading is doubtful? Does it report its uncertainties, or hide them behind a well-presented answer?
4. Human validation remains mandatory
An agent prepares a file; it does not conclude. The notary who signs a title opinion bears the professional responsibility for it, whatever tool was used. The agent must therefore make its work verifiable: documents attached to the file, sources cited, access to the original documents. It must also leave the final decision to someone who can answer for it.
5. Privacy widens
An agent sees more data than an assistant: it handles your credentials, downloads documents and sometimes sends documents to several services, for example a reading model for scanned documents. Each of these services is a recipient within the meaning of Quebec's Law 25. Ask for the complete list, and what each one receives.
Questions to ask an agent provider
- Which actions does the agent take on my behalf, with which credentials?
- How are costs shown to me, and how can I stop everything?
- Can I consult the full log of each task?
- What does the agent do when a reading is uncertain: does it carry on, stop, or warn me?
- Are the original documents attached, so I can check them myself?
- Which services receive my documents, and where are they located?
In short
A model is a conversation partner: you judge its answer. An agent is a doer: you judge its answer, its path, its spending and its safeguards. Well supervised, an agent saves considerable time on repetitive tasks such as title searches. Poorly supervised, it spends, makes mistakes and chains errors together without anyone noticing.
For choosing a general-purpose assistant, see ChatGPT, Claude or Gemini: a guide for notaries.