What AI actually changes in a notarial practice, and what it does not

Every notarial practice in the Netherlands has now sat through at least one presentation about AI. The slides promise faster dossiers, automatic checks and a future where the protocol somehow runs itself. Partners leave those presentations with the same two feelings: this is probably going to matter, and nobody in the room could say precisely how.

This article is an attempt to say precisely how. It is written by people who build systems for professional firms, so read it knowing where we stand. But the argument does not depend on buying anything from us, and parts of it argue against buying anything at all yet.

Start with what a notary actually sells

A notarial practice does not sell documents. It sells certainty: the assurance that this transfer, this incorporation, this will is exactly what the law and the parties require, checked by someone who carries personal liability for being right. Every discussion about AI in the notariat that skips this point drifts into fantasy within two paragraphs.

Because the product is certainty, the constraint on any new way of working is not speed. It is whether the work can be shown to be right. A deed produced twice as fast is worthless if the deed it feeds cannot be defended. This is why general chat assistants, whatever their brilliance, have made so little serious progress inside practices: they produce plausible text, and plausible is precisely the quality a notary cannot accept on faith.

What AI genuinely changes today

Strip away the theatre and the real, present value sits in three places.

The dossier before the appointment. Most delay in a practice is not legal reasoning. It is assembly: identity documents, extracts, prior deeds, the register checks, the source-of-funds questions, chased across mail and phone in no fixed order. This is structured, repetitive and checkable work, which is exactly the profile where current AI is strong. A system that follows what has arrived, notices what is missing and prepares the file for review can genuinely take days out of a matter without touching a single legal judgement.

The first draft that knows the house. Deeds are not written from a blank page; they are assembled from the office's own precedents and conventions. AI drafting is safe exactly to the degree that it drafts from the office's own model documents and marks every deviation for a person to see. Generic drafting from an open model, trained on nobody knows what, is the version that deserves the profession's suspicion.

The memory that does not retire. An experienced practice holds decades of judgement: how this kind of clause failed once, why the office stopped accepting that structure. Today that memory lives in senior heads and leaves with them. Systems that make an office's own history searchable and present at the moment of drafting change something economically real: the cost of experience walking out of the door.

What AI does not change

Three things stay exactly where they are, and any supplier who suggests otherwise is describing a different profession.

The judgement stays personal. No system decides whether the parties understand the deed, whether the will reflects a free mind, whether the transaction smells wrong. The law places those at a person, and so does common sense.

The liability stays personal. A notary cannot outsource being wrong. Every output that matters must therefore arrive as a proposal to a person, with its reasoning attached, and with a record of what was proposed and what the person did with it. A system that cannot show its trail does not belong in a practice, whoever built it.

The client data stays yours. A practice's files are among the most sensitive commercial records in existence. Anything that pools them, learns across offices from them, or moves them into a shared model is disqualified before the demonstration starts. This single question, where does our data go and who else benefits from it, removes most of the market.

How to actually evaluate this, if you run a practice

Do not start with a supplier list. Start with one workflow that hurts, usually the intake and assembly of a standard transfer, and write down what a week of it costs today. Then hold every conversation to four questions. Where does our data live, and is it ever shared? Can every automated step be seen, approved and reversed by a person? Does it work from our own documents and our own way of working, or from someone's average of the profession? And what does leaving look like: do we get everything back, in a form we can read without the supplier in the room?

Then insist on a proving period against the real work, in the background, before anything touches a client. If a supplier resists running in shadow first, that tells you what they expect the shadow to show.

Where we stand

What the market calls AI, we build as something narrower: one Brain per firm, built on that firm's own files, its own conventions and its own rules, running in shadow until the record shows it is ready, with every action logged and a person in approval wherever the practice wants one. Nothing pooled between offices, ever. Whether that is the right answer for your practice is exactly what a short conversation is for; the page on ATTEST, the Notarial Brain sets out how it works for the notariat in particular.