In mid-September 2026, Italy's Court of Cassation presented the results of a July-August pilot to its section presidents. The tax-law section used Microsoft Copilot, through the Justice Ministry's institutional account, to read appeal files and prepare summary sheets. The sheets covered formal requirements, procedural history and the parties' arguments. Staff and magistrates checked every output before it entered the system. Iusletter's account says the project was led by section president Angela Maria Perrino, with counselor Gianluca Grasso building and training the agent. It also reports that data stays in Italy and is not used for model training. First President Pasquale D'Ascola is weighing an extension to other sections, and the preliminary results were called "encouraging".
The design is deliberately modest, and that is why it is worth taking seriously.
The case for caution
The strongest argument against any AI in courtrooms is not about hallucinated citations, which are fixable. It is that a court's authority rests on a reasoned decision a human can be held to account for. Summarising the parties' arguments is already a form of judgment. A summary that frames an appeal one way can steer the magistrate who reads it, and a tired reviewer may skim a fluent summary rather than check it against the file. The EU's AI Act reflects this worry. It lists AI used by or for judicial authorities to assist in researching and interpreting facts and law among its high-risk categories (Annex III, point 8(a)). Regulators who treat this area carefully are not being paranoid.
Where Italy's rules draw the line
Italy has written that caution into statute. Article 15 of Law 132 of 23 September 2025 provides that when AI is used in judicial activity, every decision on interpreting and applying the law, assessing the facts and adopting measures is always reserved to the magistrate. The same article gives the Justice Ministry authority over AI systems used for court organisation and ancillary administrative work. It also requires ministerial authorisation for experimentation in ordinary courts until the EU regulation is fully applicable, and it mandates training for judges and court staff. The law itself was published in the Gazzetta Ufficiale on 25 September 2025 and took effect on 10 October 2025.
The Superior Council of the Judiciary (CSM) layered operational guidance on top. According to Edotto's summary of the council's 22 July 2026 update, Copilot may now be used for analytical and reconstructive tasks: summarising judicial acts, classifying documents, linguistic revision and assisted drafting of standardised documents. Drafting rulings and the dispositif remain off limits. ECNews reports that Copilot was extended to all magistrates through Microsoft 365 from 1 January 2026, and that the council fixed a later review date. The Cassation pilot is therefore not a rogue experiment. It sits inside a statutory reservation, a ministerial framework and a council-approved use list.
Why this is the right shape of experiment
Three features of the pilot deserve to become the default for courts elsewhere.
- The task is bounded and checkable. A summary sheet of procedural history can be compared line by line with the file. Unlike a legal conclusion, it has a ground truth. Errors are detectable by a competent reader in minutes.
- Human review is mandatory, not optional. Iusletter quotes the project's own framing: the machine prepares, and the human verifies and answers for the result. Accountability stays with a named person.
- Data governance is settled before scale. Italian data residency, tenant isolation from other Microsoft 365 environments and a ban on training use address the confidentiality concerns that sink most public-sector AI projects. These are contractual guarantees from the vendor, so they need audits to be more than promises.
The benefit is concrete. Italy's tax chamber at the Cassation has long carried a heavy backlog, and the 2022 reform that created a dedicated section (Law 130/2022) was meant to speed it up. A court that spends less magistrate time on intake, and more on reasoning, delivers faster and more consistent justice. For litigants, and for businesses waiting years on a tax dispute, that matters more than any debate over machine judges.
What the pilot does not yet prove
It would be a mistake to treat two months in one section as proof. We have not seen published error rates, time savings or the share of summaries that reviewers corrected. Without those numbers, "encouraging" is an impression, not a finding. Before the Court extends the tool, it should publish three things: a sample-based accuracy audit against the underlying files, a log of the corrections reviewers made, and an assessment of whether summaries anchor the magistrate's framing of the case. The anchoring risk is the real one. A summary that is 98% right can still tilt the reading of the other 2%.
There is also a boundary problem. The line between "summarising the parties' arguments" and "evaluating them" is thin, and the CSM bars AI from evaluating evidence. Courts should resist drift, where each tool upgrade moves a bit further from reconstruction toward assessment, without anyone deciding that it should.
A proportionate model
The pro-innovation reading is not "let courts use AI freely". It is that proportionate rules, a hard statutory reservation of judgment, a tight use list and mandatory human verification let a court capture real efficiency without eroding legitimacy. Blanket bans push magistrates toward unsanctioned consumer tools, which is the worst outcome for confidentiality. Italy has instead chosen an approved tool with local data, narrow tasks and accountable reviewers. If the Cassation publishes its evaluation data before scaling, that choice will be a template for other civil-law courts.