New AI Rules for Businesses 2026: What Changes from 23 October

What Italian Decree 179/2026 provides on artificial intelligence

Decree 179/2026 does not emerge from nowhere: it implements the Italian AI law, No. 132 of September 2025, and transposes the European AI Act into the national legal framework. Compared with the framework law, the decree goes into detail: who oversees compliance, how oversight works, and what is at stake if the rules are not followed.

The relevant authorities are AgID (the Agency for Digital Italy) and ACN (the National Cybersecurity Agency). They are responsible for enforcement and for managing the sandbox. Penalties can reach up to €35 million or 7% of total annual worldwide turnover, whichever is higher — a threshold that can be material even for a mid-sized SME.

For professional firms, there is an additional element: professional bodies and associations have six months from the decree's entry into force — that is, until 23 April 2027 — to update their training regulations and include content on AI, covering its technical, legal and professional-conduct dimensions. Those who today use AI in their practice without any formal training on the subject will need to complete programmes recognised by their professional body.

AI and employment: when automated dismissal is void

Article 41 of the decree sets out a clear rule for anyone using AI to manage human resources: final decisions on hiring, contract changes, dismissal and disciplinary measures cannot be taken exclusively on the basis of automated processing. The decision-maker must be a real person with genuine, autonomous authority — not someone who simply clicks "approve" on an algorithm's output.

The most immediate consequence concerns the validity of such acts: a dismissal made in breach of this prohibition is null and void. Not merely voidable, not contestable through ordinary proceedings: null and void. And the employee's consent cannot cure the defect.

It is worth clarifying what is not prohibited: using AI to analyse performance, process data, suggest evaluations or support HR managers is entirely lawful. The problem arises when the algorithm replaces human judgement rather than informing it. The distinction between "support" and "delegation" is the heart of the provision — and also the point at which many systems currently in use risk falling foul of it.

A further obligation concerns transparency: before initiating the processing covered by the rule, the employer must fulfil the information obligations already required by existing legislation. Employees may request an intelligible explanation of any decision concerning them, including an indication of how AI systems influenced the process.

Workplace safety and AI: the Risk Assessment Document must be updated

Article 42 of the decree introduces an obligation that directly affects businesses using AI to organise work: if an AI system influences production rhythms, the way tasks are performed, or decision-making processes that are relevant to safety, the associated risks must be assessed and included in the Risk Assessment Document (DVR), as required by Article 28 of Italian Legislative Decree 81/2008.

This is not a blanket obligation covering any software. The rule applies to AI systems that concretely change the conditions in which people work: an algorithm that automatically adjusts workloads, assigns tasks to operators, or takes decisions that are relevant to the safety of a facility is precisely the kind of tool that falls within the scope of Article 42.

Alongside this, employees have the right to receive adequate information and training on the specific risks associated with the AI systems used by the organisation, and on the preventive measures in place. This is not a minor detail: in workplaces where AI tools are in use without employees having been informed, there is exposure both on the regulatory and reputational fronts.

For professional firms using AI to manage deadlines, assign matters or monitor the workloads of their staff, now is the time to check whether these tools fall within the scope of the provision — and to document the assessment, either in the DVR or in an equivalent document.

Is your firm or SME ready for Article 42?

• ✕
  Mapping of AI systems in useDo you have an up-to-date list of all AI tools that influence your work processes?
• ✕
  AI risk assessment in the DVRHave the risks associated with these tools been assessed and included in the Risk Assessment Document?
• ✕
  Staff trainingHave staff been informed about how the AI systems work and the associated risks?
• ✕
  Updated employee information noticesHave the information obligations towards employees who interact with AI systems been fulfilled?
• ✕
  Periodic reviewIs there a procedure for updating the assessment whenever new AI tools are adopted?

What Nadella means when he says AI models are 'compromised'

On 10 October 2026, Satya Nadella — CEO of Microsoft — published a post on X that sparked considerable debate. Not because of alarmist rhetoric, but because of the clarity with which he described the problem. The central point, in his own words: «We cannot treat superintelligence as a set of nested black boxes and simply accept or reject its recommendations, responses and actions».

Nadella is not saying that AI is useless. He is saying that it must be contained by design: separating the model from the system that orchestrates its work, externalising controls, and documenting every significant action with "human-readable and tamper-proof evidence". And — the point that sets his contribution apart from similar ones — he states that «we must assume that a model is compromised and keep it under control from the very start. Think of it as an emergency brake».

For a professional firm or SME, the practical meaning is this: using an AI model without being able to stop it, without tracking what it has done, and without a human responsible for validating its actions is not a "cautious but limiting" choice. It is an exposure — regulatory, today, under Decree 179/2026; reputational, always, with clients who expect your tools to be under control.

This position is not an isolated one. As we have already explored in our article on controlling AI agents in business, AI governance is not a luxury reserved for large organisations: it is the minimum requirement for using AI in a professional and defensible way.

Note: If you cannot stop an AI model while it is working, do not know what it has done, and cannot demonstrate this to an inspector or a client, you are already breaching the spirit — and in some cases the letter — of Decree 179/2026 and the AI Act. The question is not «do you use AI?» but «can you show how you control it?»

What the Italian AI sandbox is and who can use it

Articles 26 to 31 of the decree establish the Italian AI Experimentation Space, a regulatory sandbox in which companies developing AI solutions will be able to test them under the supervision of AgID and ACN before placing them on the market. The stated aim is to reduce regulatory uncertainty during development, not after.

It is important to be clear straight away about who can access it and who cannot. The sandbox is reserved for providers and prospective providers of AI systems — that is, for those who develop and intend to bring AI technologies to market. It is not a tool for professional firms or SMEs that use third-party software. If you use ChatGPT, a CRM with AI features or a management system that integrates language models, the sandbox does not directly concern you.

Those who develop their own solutions — or are a start-up or SME building an AI product — will find a fast track in the sandbox: the decree gives priority access to SMEs and start-ups, according to criteria to be defined by subsequent measures.

A practical caveat: the decree formally establishes the sandbox, but its actual opening to operators will require further measures from AgID and ACN. There is no operational access point yet. At the European level, the AI Act as updated by the Digital Omnibus requires each Member State to have at least one operational experimentation space by 2 August 2027.

How to demonstrate to clients and inspectors that you use AI safely

The most practical question a firm owner or SME manager should be asking today is: if an inspector or a client asked me tomorrow to demonstrate how I use AI, what would I show them? Decree 179/2026 does not ask you to stop using AI. It asks that its use be controlled, documented and explainable.

Three elements make the difference. The first is traceability: knowing what every AI tool has done, on which data, at what moment and with what output. It is not enough to know that "the AI helped draft the document": you need to know which model, with what instructions, and who reviewed the result. The second is effective human oversight: not a rubber stamp, but a person who understands the process and can intervene. The third is up-to-date documentation: client disclosures, employee notices, internal procedures describing how AI is used and controlled.

On the privacy front, the answer to the question «do I need to update my privacy notice if I use AI tools?» is almost always yes. If clients' or employees' data is processed by AI systems — even just to categorise emails or generate document drafts — the notice must reflect this reality. This obligation already flows from the GDPR and is reinforced by the AI Act and Decree 179/2026 framework.

A practical approach is to start with a simple mapping exercise: which AI tools you use, what data they process, who oversees them, and how the output is documented. This framework — which Ovia helps you build — is not just about compliance: it is also the most direct way to communicate to clients that their firm operates to professional standards. As we explained in our article on professional liability and AI hallucinations, whoever signs is responsible — and demonstrating control is the only real defence.

Can you demonstrate safe AI use?

• ✕
  Register of AI tools in useDo you have a list of all active AI systems in your firm or business, including the type of data they process?
• ✕
  Updated client privacy noticeDoes the privacy notice mention the use of AI tools in the processing of client data?
• ✕
  Updated employee noticeHave staff been informed about the use of AI in the processes that affect them?
• ✕
  Human supervisor identifiedFor every critical AI process, is there a named person responsible for validating the output?
• ✕
  AI activity logIs there a system — even a simple one — to track what each AI tool has produced and who has reviewed it?
• ✕
  Shutdown procedureDo you know how to stop or disable an AI system if it behaves abnormally?