Start a trial

Operational notes Scenarios

AI for legality: what it takes for an automated analysis to hold up

6 min read

Paper accounting ledger open across two pages, pre-printed columns and handwritten entries, in black and white
A ledger is worth exactly as much as the line you can read back: who wrote what, when, and against which document.

On 27 July 2026 a memorandum of understanding was signed in Rome between Italy’s Ministry of the Interior, the Marche Region, the National Anti-Corruption Authority (ANAC) and the National Research Council (CNR). It is announced, in largely overlapping wording, by the ANAC statement and by the Region’s. It renews and widens a 2023 agreement between the Interior Ministry, Marche and ANAC; the declared novelty is the CNR joining. The Authority’s statement specifies that the text runs to eleven articles.

One point first, for anyone reading this for work: the articles have not been published. Neither page attaches the document. We know what the signatories say they want to do with it; we do not know what those eleven articles provide. The statements give no duration, no governing body, no timetable and no figure for spending. What follows therefore keeps two planes apart: what is already in service, and what is a declared intention.

What is already running

One thing only, but a real one. TrasparenzAI is an open-source portal developed by the CNR and — both statements say so — “already used by Anac for the automated analysis and monitoring of transparency obligations”, able to analyse the “Amministrazione trasparente” sections of more than 23,000 administrations. The statutory perimeter is that of Legislative Decree 33 of 14 March 2013: the platform checks that the sections public websites must publish are actually there.

The code can be read. The components sit on GitHub under the TrasparenzAI organisation, in a microservices architecture, and those published carry the GNU AGPL v3 licence. One detail is worth more than many announcements: the component that integrates the AI part is described in its own documentation as an interfacing service with Ollama — an engine that runs models locally rather than a remote service. Anyone who works on data sovereignty knows how much that difference weighs.

What is a declared intention instead

The rest is in the future tense, and the statements write it so: “the trialling of artificial-intelligence tools to improve the accessibility, interoperability and reuse of information”. In particular CLARA, described as the AI system built into TrasparenzAI, “will be tested on concrete administrative cases” so that citizens can query complex documents with simple questions, “starting from the status of public works planned, designed and under way”.

This needs saying precisely, because the headlines moved fast: what is described is an aid to consulting acts that are already public, not an engine that detects wrongdoing. The declared field of application is public contracts — integrity culture, prevention of corruption, support to administrations in tender procedures and in the use of national databases, the execution phase, the qualification of contracting authorities, investments funded by the national recovery plan. In remarks reported by the regional statement, the Under-Secretary of State at the Interior Ministry places the model in the context of central Italy’s post-earthquake reconstruction: “around 8,000 square kilometres” of affected area and “more than 16,000 public and private sites”. As for the data that would be processed, the statements list none beyond what administrations already publish. On that, you do not infer: you wait for the text.

An indication is not a finding

Here begins the part that concerns anyone who, in a public body or a company, will have to make such a system work — or live with it. An automated analysis sorts, compares, flags a deviation. It proves nothing. It produces an indication: something somebody will have to look at, and which in most cases will explain itself. Between “the system flagged it” and “it was established” lie an investigation, a right of reply and a signature. This is not politeness: it is the difference between a usable tool and one that collapses at the first challenge.

Four requirements follow. They describe none of the administrations named above: a recurring pattern, declared as such.

  • Provenance of the data, field by field. An automated survey of websites measures the state of a page at a given moment. That is exactly what is needed to monitor publication duties; it is not a measure of a body’s conduct. A missing section may be an unmet obligation, a CMS migration or a broken link. For every field, origin, date and collection method must be recorded — and must stay readable downstream.
  • The threshold is settled beforehand. Deciding when a deviation becomes a flag is an administrative choice dressed up as a technical parameter: it decides how many cases reach the desk and what share of false positives the office agrees to work through. It must be written, reasoned and dated before go-live, not tuned afterwards to the results that read best.
  • Traceability of the indication. From which document, under which rule, with which version of the model. If a generative model sits downstream, the same question yields different answers across two versions: the log must say which one was running that day. An indication that cannot be reproduced cannot be verified, and what cannot be verified does not belong in a procedure.
  • Who answers for it. The signature stays with a person. For that person to sign, the system must hand over a file that can be produced: input, rule applied, outcome, who read it, what was decided and why.

Three pieces already on the table

These are not abstract principles: they are written elsewhere, in rules that bind.

How we deal with it

An analysis system serving a public control function holds up on one condition only: being inspectable. Readable code, data with its provenance attached, the rule written beforehand, logs available to whoever will answer for the decision — not to the supplier. That is why we deliver in two modes, never one: on-premise, with the AI installed in the client’s own environment, or on our dedicated cloud reserved to the single client, with a dedicated VPN, a data centre resident in Italy and premises we staff ourselves. In both cases the models stay inside the perimeter, the trail behind every indication stays producible, and the ledger is not held by somebody else. That is what it means for a public body before it is a product choice.

If you are assessing an automated analysis supporting a control function, half an hour with one of our experts is enough to establish what it will have to show, and to whom.

Sources