Skip to content
Download

An AI policy for a three-person firm

5 min read

In brief

  • Nobody reads a twenty-page policy, and an unread rule is not an organisational measure.
  • Permitted tools are listed by name and by plan: free tier and work plan differ on training terms.
  • The last section is the one that counts: who has read it, with a date and a signature.

The AI policies going around are written for companies with an in-house legal department: twenty pages, definitions, risk matrices, roles. In a three-person firm a document like that does not get read, and a rule nobody read is not an organisational measure: it is a file.

What follows is the minimum structure that works. It fits on a page, and its quality depends almost entirely on the last section.

0. Before writing: know what is already there

The step everybody skips comes before the first one, and without it the policy describes a firm that does not exist. Before deciding what is allowed you have to know what is there, and the answer hardly ever matches what the partners believe.

Four places to look, in half an hour. Personal subscriptions: almost every firm has two or three, activated on a credit card with nobody asking, and that is the line that produces the surprises. Browser extensions, which have permission to read open pages and therefore also the practice system. Add-ins in the word processor and the spreadsheet. And new features in the software you already use: the “summarise” button that appeared in an update to the practice system is an artificial intelligence tool in every sense, and nobody chose it.

This list does two jobs. It is the starting point of the policy, and it is also the first of the measuresArticle 4 of the AI Actopens in a new tab asks for under the name AI literacy: knowing which tools come in and who uses them. Write it once and keep it beside the policy rather than inside it: it changes more often than the policy does.

1. What is allowed, by name

A list of tools, by name and by plan: “assistant X, business plan, firm account”. Not “AI tools approved by the partners”, which points at a decision that does not exist. If someone wants to add one, they ask, and the list is updated.

Names matter because contractual terms differ between the free version and the work version, and that difference is exactly what decides whether your documents end up in training.

2. What never leaves

One line, and it is worth drawing it wide: no document concerning a client leaves through a tool that does not run on the firm’s computers, other than after identifying data has been substituted. Then the exceptions, if there are any, written down.

3. What always gets verified

Case numbers, statutory articles, dates of entry into force, threshold amounts: anything with a number gets opened at the source before it goes into a filing. It is the rule the Court of Florence restated, in an order of 14 March 2025, in a case about invented judgments.

4. Who gets told, and how

Article 13 of Italian Law 132/2025opens in a new tab requires informing the client in clear, plain and complete terms. The policy says where the text lives and who hands it over, it does not repeat it: the text belongs in the retainer or the notice, and what it has to say is written elsewhere.

5. What gets written down when something happens

Two lines on what to do if a document went out by mistake: who to tell, by when, and what to record. This is not defensive paperwork: without a written procedure the natural reaction is to tell nobody, and the delay is the part that makes everything worse.

6. Who has read it

The last section is the only one that turns the document into a measure. Article 29 GDPRopens in a new tab says that anyone processing data under the controller’s authority does so on its instructions; the instruction has to have been given, and the controller has to be able to show it. A date and a signature at the bottom, for each person.

The three lines not to write

Three formulas turn up in almost every first draft and are better removed, because each produces the opposite of what it is reaching for.

“The use of artificial intelligence systems is prohibited.” It will not be followed: somebody with a deadline tomorrow will use the tool and tell nobody, so the ban will have produced invisible use rather than no use. It is also out of step with the law: Italian Law 132/2025 regulates that use, it does not ban it.

“Tools are to be used with due diligence and with respect for confidentiality.” True, and not a rule: it tells nobody what to do at any particular moment. You can recognise a rule because, faced with an open document and a deadline, it says yes or no. This one says “you decide” at greater length.

“Every use must be authorised in advance by the partners.” It creates a bottleneck on something that happens twenty times a day, and bottlenecks get routed around, usually on day two. Prior authorisation makes sense for adding a tool to the list, not for using the ones already on it.

The part that makes half the document pointless

A policy holds for as long as following it costs less than going round it. If anonymising by hand takes twenty minutes, section 2 breaks on the first evening before a filing, and it will not break out of ill will: the arithmetic is here.

A tool that substitutes in thirty seconds, on the computer where the document already is, is what keeps section 2 standing without anyone having to be reminded. That is the work Omissia does.

Where this goes next

If the question left is how it works in practice, the four steps and the part that travels over the network are in “How it works”.

How to redact

Frequently asked questions

How long should an AI policy be?

One page. Twenty-page policies are written for companies with an in-house legal department, and in a three-person firm they do not get read: a rule nobody read is not an organisational measure, it is a file.

What cannot be missing?

Six things: the permitted tools by name and plan, what never leaves, what always gets verified, who informs the client and where the text lives, what gets written down if a document goes out by mistake, and who has read it, with a date and a signature.

Why do the tool’s name and plan matter?

Because free-tier and work-plan terms differ on training and retention. “Tools approved by the partners” points at a decision that does not exist; a list with names can be checked.

Does writing “the use of AI is prohibited” work?

No, for two reasons. It will not be followed: somebody with a deadline tomorrow will use the tool and not mention it, so the ban produces invisible use rather than no use. And since 10 October 2025 Italian law regulates the use of these tools in the intellectual professions rather than banning it, so the prohibition is not even the cautious alignment it looks like.

Sources

On the same topic

Try it on a case you have already closed.

Three days with every feature. The filing stays on your computer, and at the end of the trial you decide.

We use technical cookies needed for the site to work and, only with your consent, analytics cookies to monitor that everything works properly and to improve the browsing experience. Cookie policy