Drafting is where a language model is most useful and most dangerous at the same time: useful because the shape of a deed is repetitive and the machine knows it well, dangerous because the draft you are working on is already full of the parties’ details.
The two are not tied together. What the model needs to write a clause is the legal structure; what sits in the draft is the legal structure plus the personal details of four people.
The precedent is shareable, the matter is not
A warranty against eviction, a waiver of a legal mortgage, a declaration of cadastral conformity: these live in a precedent bank, not in a file. Asking a model to reword them, adapt them to a set of facts or check them for internal consistency needs no real data at all.
The well-written question describes the facts and not the people: “sale of a property held in community of property, part of the price paid by assumption of the existing mortgage; draft the clause on payment terms”. There is nothing to anonymise in a question like that, because nothing went in.
When the real draft is needed whole
Some checks only make sense on the whole document: consistency between the recitals and the operative part, the presence of the mandatory declarations, comparing two versions after one side’s amendments. The text has to leave in full there, and the question becomes what state it leaves in.
With placeholders the check survives intact, and this is the part that surprises anyone who has tried black boxes. If every party has a stable identifier, the model still sees that the seller in the recitals is the seller in the operative part, and flags it when they are not. With the text covered over, that check cannot even be set up.
Substituting too much is the opposite mistake
People who have understood the problem then tend to overshoot the other way, and the result is a draft with the legal situation removed as well. If the model receives “[PARTY_1] sells [ASSET_1] to [PARTY_2] for [AMOUNT_1]”, the answer that comes back is as generic as the question, because the question no longer describes anything.
The line is sharp and worth keeping in mind: substitute what identifies a person, keep what describes a transaction. So the matrimonial property regime stays, and so do the nature of the asset and its functional category, whether there is financing and of what kind, the order of the steps, and the chain of title understood as a type of deed: “a 2011 gift”, not “a deed of notary So-and-so of 3 March 2011”. What goes is everything that is a key: names, tax codes, IBANs, land registry references, addresses, dates of birth, repertory numbers.
Dates are where this goes wrong most often, which is a pity because the fix is simple. A precise date identifies, but what the model needs is not the date: it is the sequence. A placeholder that preserves the order ([DATE_1] before [DATE_2]) still allows you to check that the preliminary contract precedes the final one and that a pre-emption right was exercised in time, without saying when.
The return trip is where the time is saved
Substitution has a second effect that only shows up later. When the model’s answer comes back carrying the same placeholders, it can be reassembled onto the real matter instead of retyped: the rewritten clause comes out with the right names in the right places.
That is the difference between a tool that redacts and one that substitutes, and it is not a security point: there is nothing to put back under a black box, so the work the model did has to be transcribed again. The difference between the two routes is set out elsewhere.
What stays with the notary
Italian Law 132/2025opens in a new tab requires the professional’s own intellectual work to remain predominant, and in drafting that is not an empty formula: a model proposes a plausible form, and plausibility is exactly how a wrong clause goes unnoticed. Reading it back is not an added formality, it is the job.
It helps to know where to look, because a generated clause goes wrong in recurring ways. First, the mandatory declarations: the planning references required by Article 46 of the Italian building codeopens in a new tab and Article 40 of Law 47/1985opens in a new tab, and the cadastral conformity statement under Article 19(14) of Decree-Law 78/2010opens in a new tab. A model reproduces them in the right shape with the wrong references, or invents one that does not exist, and those are precisely the lines a deed dies on. Second, references to repealed provisions, which survive in training material for years after the reform that replaced them. Third, the tax wording: first-home relief, the price-value regime, rates, all of which change more often than any model does.
So the check that matters is not rereading the clause to see whether it “sounds right”: it will always sound right, that is the thing the tool does best of all. It is opening the legal references it contains, one by one, the way you would open a case citation.
What can be automated without argument is the mechanical part: finding the values to substitute, lining them up, replacing them once you have confirmed them, and putting them back afterwards. That is the work Omissia does, on the office computer.