On 23 July 2026 Italy's data protection authority, the Garante, adopted its inspection plan for August to December: at least 35 checks, also through the Guardia di Finanza. The only line on artificial intelligence covers municipalities and local health authorities (ASL). For an SME that uses chatbots, assistants or model APIs on customer and employee data, the most concrete references are two decisions already taken in 2026, on Myndoor and on Character.AI. From them we draw the questions to ask before switching on an AI tool.
What the plan covers
Deliberation no. 545 of 23 July 2026 covers August to December and follows no. 797 of 30 December 2025, which ran from January to July. It provides for at least 35 checks, also through the Guardia di Finanza, on every area below except the last. For the Garante an inspection is an on-site check, on subjects not necessarily chosen from complaints or reports.
- geomarketing, neuromarketing and dynamic pricing;
- electronic health records and medical apps;
- data breaches in major public databases, and technical checks on breaches notified to the Garante, in the public and private sectors;
- DPO service providers to public bodies, municipal social-assistance services and their software suppliers, the Visa Information System, the Customs Information System;
- application of the 17 April 2026 guidelines on tracking pixels in email;
- AI systems used by municipalities and ASL;
- ex officio inspections for urgent cases, reports or complaints.
In the January to July plan, with at least 40 checks, the AI line instead covered tools used in schools. Marketing, tracking pixels and dynamic pricing name no sector, so the text does not limit them to public bodies.
What the Garante decided on AI in 2026
On 14 May 2026 the Garante issued a warning to Myndoor S.r.l., an Italian company that develops a plug-in for Slack and Teams. The plug-in uses AI to analyse the messages of employees who choose to use it, to estimate their stress, and can give the employer a weekly aggregated report. The employer, the Garante notes, cannot access either the messages or the individual results. The Garante found no evidence that client companies had processed employees' data and did not fine the company. The warning concerns the risk that the report could later allow people to be identified, a risk that according to the Garante also depends on the size and staff of the client company.
In the warning, the Garante writes that information on employees' emotional sphere, meaning well-being or psychological stress, falls under Article 113 of the Privacy Code: the employer cannot know it. It also cites the AI Act ban on using AI systems to infer emotions in the workplace (Article 5(1)(f)), with exceptions for medical or safety reasons. It warns that systems based on semantic analysis can give results that are hard to explain and verify, with a risk of errors and discrimination, and it calls for effective human control.
On 3 July 2026 the Garante fined Character Technologies Inc., a US company that runs Character.AI, EUR 158,000. Character.AI is a generative AI service in which users, minors included, chat with virtual characters. The Garante found five infringements: privacy notices that were not clear and complete; inadequate information on the use of Italian users' data to train the models; a data protection impact assessment (DPIA) adopted late; the EU representative designated late; and inadequate age-verification measures for minors, in 2024 and for Italian users under 16 until 8 April 2025. The DPIA, the Garante writes, should have been done before launch, in September 2022; the first version is dated 14 November 2024.
What it means for your company
| Situation | What the Garante has said |
|---|---|
| Chatbot or AI service open to the public | Character.AI case: complete privacy notice, DPIA before launch |
| Tool that analyses employees' messages or mood | Myndoor case: the employer cannot know their emotional state |
| Software or services for municipalities and ASL | In the plan: AI use by municipalities and ASL, social-assistance software, DPO services |
These are the steps we recommend, not obligations read in the plan:
- List the AI tools that touch personal data: chatbots, assistants, meeting transcription, model APIs wired into your management software, even those employees use without telling you.
- For each one, write down what data goes in, where it goes, who the supplier is and whether a data processing agreement exists. Ask in writing where the data is processed and whether it is used to train models. If you cannot answer for the last tool you switched on, start there.
- Check that the notice to customers and employees names these tools and says whether the data is used to train models.
- Ask your DPO or a privacy consultant whether a DPIA is needed, and do it before switching the tool on. The Garante treats it as mandatory for AI systems when at least one other criterion of the European Data Protection Board (EDPB) applies, for example large-scale data.
- Do not switch on features that infer employees' mood, stress or emotions, and do not take delivery of reports on that data, even aggregated.
Where we start
Our AI automation and chatbot service includes control over costs and answer quality. If you ask us to look at an AI feature in your software, we start from a page written before it is switched on: which personal data goes in, which system it stays in, which supplier receives it and under what agreement, how long it is kept, who can read it. It is the same list as steps 1 and 2 above.
The documentation of Kmemo, our open-source semantic cache for model calls (Apache-2.0 licence), says that the cache keeps questions and answers in the store you choose, and that every question reaches the embedding provider before the cache decides anything: the rule that forbids saving it acts afterwards. Whoever must not send certain data to a provider therefore filters it in their own code, before calling the cache. Kdrant, our Kotlin client for Qdrant, talks to a Qdrant server over the network and never holds an index itself.
One limit: the impact assessment is the controller's job. We put the technical path of the data in writing, which feeds that assessment but does not replace it; we do not give legal advice.
What we do not know yet
- Which subjects will be inspected and what will be challenged: the plan lists areas, not names.
- Whether the AI checks will reach companies: the plan names only municipalities and ASL.
- Whether and how the Garante will apply the AI Act ban on inferring emotions at work: in the Myndoor case it cites it, but the warning lists only provisions of the GDPR and the Privacy Code.
If you want us to look at an AI feature in your software before it goes live, write to us through the contact page or start from our AI automation and chatbot service.
Sources
- Garante privacy: Deliberazione del 23 luglio 2026: attività ispettiva agosto-dicembre 2026
- Garante privacy: Deliberazione del 30 dicembre 2025: attività ispettiva gennaio-luglio 2026
- Garante privacy: Provvedimento del 14 maggio 2026 (avvertimento a Myndoor S.r.l.)
- Garante privacy: IA e lavoro: il Garante Privacy avverte una start-up italiana (comunicato stampa) (28 May 2026)
- Garante privacy: Provvedimento del 3 luglio 2026 (ordinanza ingiunzione a Character Technologies, Inc.)
- Garante privacy: Intelligenza artificiale: il Garante privacy sanziona Character.AI (comunicato stampa) (9 July 2026)
- EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5 (13 June 2024)
- NaCode Studios: Kmemo, threat model (library documentation)
- NaCode Studios: Kdrant, README (library documentation)
