AI for Notaries and Law Firms: What It Should and Shouldn't Do

Today, AI in a notary's office or a law firm is good for reading documents, pulling out the data and drafting documents that a person reviews. It isn't good for deciding, certifying or giving an opinion without a source. The difference between a pilot that stays and one that gets abandoned almost always comes down to respecting that line from day one.

A note for readers in the US and Canada: in Mexico, a notary (notario público) is a licensed attorney appointed by the state who drafts and authorizes deeds, powers of attorney and corporate documents, closer to a civil-law notary than to a notary public. Most of what follows applies to any firm that turns documents into other documents.

What it can do today

Read IDs and deeds

Vision models read the photo of an ID or the scanned pages of a deed and pull out the data: name, national ID number (CURP), address, property details, registry references. What used to be retyping from a PDF into a form is now reviewing what's already been captured.

That's how Notaría Fácil, in our portfolio, works: it uses AI to extract data from Mexican IDs (INE), CURP records and property deeds to generate notarial documents, cutting manual data entry for notaries. Every value it extracts has to be checkable against the document it came from; otherwise, reviewing takes as long as typing.

Extract data and fill the template

Once read, the data goes into the firm's own templates: the draft deed, the power of attorney, the contract, the letter to the registry. The system doesn't "write" the document from scratch; it fills the template you already use, with your clauses and your style. And it leaves blank what it didn't find, instead of guessing.

Draft with sources

For less mechanical documents, AI can propose a draft from the firm's own templates and precedents, always with a lawyer reviewing. In our portfolio, LegalStock has a library of contracts for small businesses that are generated, customized and signed online, with AI-assisted drafting; and Tlamatini, a tax and legal advisor we keep as a demo, answers questions on Mexican law with the statute or case law behind each statement and an anti-hallucination check before the answer is shown.

The rule we took from those projects: a legal statement without its article or precedent doesn't get shown. It's the same rule behind Papá Presente Michoacán, a public legal guide where every claim cites the exact article of the state's family code that backs it.

Search what you already have

A firm with years of files has a base of precedents almost nobody consults because nobody can find anything. An AI search over those documents answers "how did we handle this in a similar case?" with a citation to the file, not an opinion.

What it must not do

• Decide. Whether a transaction can proceed, whether a client qualifies, which clause is better. AI prepares; a person decides.
• Certify or sign. The notary's public faith and the lawyer's professional responsibility belong to people. No system replaces them, and none should give the impression that it does.
• Make things up. A model without sources fills gaps with something that sounds right. If the value isn't in the document, the field stays blank and flagged; if the statement has no article behind it, it doesn't go out.
• Advise clients on their matter with nobody behind it. It can take details and book appointments; it can't advise on the case. Before putting an agent like that in front of clients, this checklist helps.
• Send client data to just any service. IDs and deeds are personal data; in Mexico, the federal data protection law for private parties applies, and in Canada, PIPEDA. Before uploading the first document, know which AI provider receives it, in which country it's stored, whether it's used to train models, and that the firm's privacy notice covers that use.

How to measure whether it helps

Not with "model accuracy," a number that tells the firm nothing, but with what changes in the reviewer's work:

• Before you start, on a typical matter: how long data entry takes, how many errors review catches, how many documents come back because of a wrong value.
• During the pilot, per document: how many fields came out right and how many the reviewer had to correct, how long review took and how many documents came back.
• The honest comparison: the same files, entered by the team and by the system, side by side.

The goal isn't that nobody reviews. It's that reviewing takes less time than typing and that fewer errors reach the final document than today. If after a month review takes just as long, the pilot didn't work, and it's better to know.

How to start: a small pilot

1. One matter type and one document. A property sale and its draft deed, say. Not "the whole firm."
2. Closed files as the test set. Twenty are enough to start. Since they're already entered, you can compare what the system extracts against what the team typed without touching a live matter.
3. One person at the firm who owns the pilot. Knows the matter type, reviews the results and decides what gets fixed.
4. The data rules in writing before the first document: provider, where it's stored, for how long, who has access, and the accounts in the firm's name, not the builder's.
5. Four weeks and a decision. With the numbers from the section above, you decide whether to extend it to the next matter type or stop. Both are good answers.

A pilot like this doesn't change how the firm works; it adds a step that saves typing. If it works, the next matter type adds itself, because the team asks for it.

If you'd like to talk it through

If you run a notary's office or a law firm and want to know where to start, we'll spend half an hour on your most repeated matter and tell you honestly whether it's worth a pilot. See how we build AI agents or reach us through contact. And if document entry is only one of several repeated tasks, what to automate first may help.