--- title: "AI-assisted decisions and effective human oversight: rules, dates and training" url: https://zerofive.ai/en/blog/compliance/ai-assisted-decisions-effective-human-oversight canonical: https://zerofive.ai/en/blog/compliance/ai-assisted-decisions-effective-human-oversight language: en published: 2026-10-20 updated: 2026-10-19 author: "ZeroFive.AI" tags: Legislative Decree 179/2026, human oversight, AI Act, AI training, AI Academy abstract: "From 23 October, Article 41 of Italy's Decree 179/2026 voids employment decisions taken solely by AI. Rules, dates and training for effective oversight." --- # AI-assisted decisions and effective human oversight: rules, dates and training **In short.** From 23 October 2026, Article 41 of Italian Legislative Decree 179/2026 makes employment decisions void when they are taken solely on the basis of automated processing, and requires the decision to rest with a person holding real and autonomous power. Read together with Legislative Decree 160/2026 and Articles 14 and 26 of the AI Act, the rule shifts the weight from the choice of systems to the preparation of the people who oversee them, and that preparation is the focus of the legal framework, the calendar and the AI Academy module described below. From 23 October 2026, a dismissal in Italy can be declared void because of the way it was decided, before any question about the reasons behind it. This is set out in Article 41 of Legislative Decree no. 179 of 7 October 2026, published in the Official Gazette no. 234 of 8 October, which aligns Italian law with the AI Act and establishes that final decisions on hiring, changing or terminating an employment relationship, including disciplinary measures and performance evaluations, cannot be taken solely on the basis of automated processing. The decision belongs to a natural person with real and autonomous power, the employee's consent is not enough to waive this, and paragraph 4 makes the dismissal and any other measure adopted in breach of the rule void. Reading the text, we asked ourselves a question that applies to us too, since we use language models every day to analyse documents and prepare material for clients: how many times, rereading an output, have we examined it with a concrete possibility of contradicting it, and how many times have we simply approved it? The law answers with one adjective, effective, which moves attention from the choice of software to the way an organisation prepares its people to decide, and therefore to training. ## The rules in force from 23 October Decree 179 has 54 articles and implements the delegation contained in Law no. 132 of 23 September 2025. For anyone managing people, the core sits in Chapter III. Article 41 leaves candidate search and pre-selection outside the prohibition, even when an algorithm excludes someone from later stages, while the final decision to hire is reserved to a person. Before switching on an automated decision-making or monitoring system, the employer must inform workers under Article 1-bis of Legislative Decree 152/1997, explaining how the system works, its main parameters, the control measures and its level of accuracy, and must reply in writing within thirty days to requests for clarification, including those sent through trade union representatives. Employees are also entitled to an intelligible explanation of any decision concerning them, given by a natural person, stating whether and to what extent the system influenced the outcome. Article 42 brings artificial intelligence into occupational health and safety: when a system redistributes shifts, priorities or work pace, the risk assessment must account for it, and workers must be informed and trained on the specific risks and the measures adopted. Article 47 requires professional bodies to include AI in initial and continuing training, with six months to update their rules. On supervision, the National Cybersecurity Agency (ACN) is the general authority, Banca d'Italia, Consob and Ivass act in their own sectors, and the Italian Data Protection Authority keeps its powers, with fines for prohibited practices reaching 35 million euros or 7% of worldwide turnover. Three weeks earlier, on 30 September, [Legislative Decree no. 160 of 9 September 2026](/en/blog/compliance/ai-documentation-as-evidence-decree-160-2026) came into force, changing the weight of documentation in civil proceedings. Anyone who believes they have been harmed by an AI system can ask the court to order the disclosure of operating logs, risk management documents and information on human oversight; if the order is ignored without justification, the court may treat the alleged facts as admitted, and when the harm results from a breach of AI Act obligations, the causal link is presumed unless proven otherwise. The same decree introduced Article 437-bis into the Italian Criminal Code, punishing with one to five years' imprisonment the failure to apply safety and human oversight measures to high-risk systems when this endangers people's life or physical safety. At European level, the AI Act requires high-risk systems to be designed for effective oversight (Article 14) and obliges those who use them to assign oversight to natural persons with the necessary competence, training and authority, as well as the necessary support (Article 26(2)). These obligations apply to Annex III systems, including those used for employment and recruitment, from 2 December 2027, following the postponement adopted with Regulation (EU) 2026/1744, the Digital Omnibus. The same Omnibus rewrote [Article 4 on AI literacy](/en/blog/compliance/mandatory-ai-course-not-enough), which has applied since 2 February 2025: it now requires providers and deployers to take measures to support the competence of their staff, without requiring any specific level to be guaranteed for any individual. ## Effective oversight inside the organisation A person who mechanically approves the software's proposal, without being able to examine and contradict it, does not provide the oversight Article 41 requires, and the same logic runs through Article 14 of the AI Act, which describes what the overseer must be able to do: understand the system's capabilities and limits, be aware of the tendency to over-rely on its results (automation bias), interpret the output correctly, decide not to use it or override it, and stop the system when needed. It is a list of capabilities, and capabilities are built. Automation bias grows precisely when the system works well, because the more the output proves correct in the first weeks, the fewer reasons the overseer finds to challenge it, until review shrinks to a quick confirmation that nobody records and nobody rereads. If we look at how HR processes are organised in most companies today, effective oversight depends on three conditions that are rarely designed together: the person deciding must be able to read the output and recognise its limits, must be able to decide differently without asking permission and without paying a price for it, and must leave a record of what they assessed. The second condition is the most fragile. If departing from a score requires a manager's approval, or if the deviation lowers a productivity indicator for the function, the person formally has the power to decide and in practice does not use it, which is exactly the situation Article 26 aims to prevent when it places authority next to competence. The third condition has also become a litigation issue. With Decree 160, the record of the human decision enters court proceedings as evidence, and anyone unable to show what the signatory actually assessed ends up defending a measure without documents. In 2026 the European Data Protection Supervisor published a checklist on human intervention in automated decision-making that turns this principle into concrete control points, from scenario-based training to decision logs, and one of its indications concerns deviations: if almost nobody departs from the output in a high-impact process, the figure may point to an excellent system or to excessive reliance, and only someone who knows the process can tell the two apart. ## Training as a requirement of oversight AI literacy under Article 4 and the training of those who oversee a decision answer different questions. The first concerns everyone who uses AI on behalf of the company and serves to understand what these systems are and what risks they carry; after the Omnibus it is an obligation of means, and a good general programme is enough to meet it. The second concerns a few people, each facing a specific system and a specific decision, and Article 26 ties it to the competence and authority of whoever exercises oversight, together with the support the organisation must provide. According to the Artificial Intelligence Observatory of the Politecnico di Milano (February 2026), 71% of large Italian companies have active AI projects and only 9% have structured governance. In our experience the gap is clearly visible in training, because most programmes so far have focused on tools and individual productivity, useful and necessary, while it is rare to find people trained to challenge an output that affects a colleague's career. The Italian legislator has added two obligations pointing the same way. Article 42 requires training workers on the specific risks of systems that change how work is organised, with content that clarifies the limits of the output, escalation procedures and verification responsibilities, and Article 47 brings AI into the continuing education of regulated professions, where a professional's signature has always carried legal weight and must now carry it even when a model sits upstream. Whoever oversees a decision needs to know how that particular system works, how to read its output, how to recognise an error or anomaly, how to override a result and how to stop the system, and must have experienced first-hand how easy it is to let a plausible error slip through. A prompting course teaches none of this. ## The AI Academy module In the ZeroFive AI Academy the topic becomes a dedicated module, "AI-assisted decisions and effective human oversight", which sits alongside literacy programmes and is organised by role, because the same rule asks different things depending on where a person sits in the process. - **Board and senior management.** Where AI already enters decisions about people, which measures risk being void, what documentation is needed to support a decision before a court or an authority, how the authority to depart from the output is assigned and protected. - **HR and line managers.** Critical reading of evaluations, rankings and suggestions, writing an intelligible explanation that states how much the system weighed, limits on using general-purpose assistants for evaluations and disciplinary letters. - **People overseeing a specific system.** How that system works and where it fails, exercises on outputs with planted errors, override, stop and escalation procedures. - **Health and safety.** How a system that changes shifts and pace enters the risk assessment document, and what information workers must receive. The method works on cases built from the company's real processes, with outputs in which we plant plausible errors: the exercise measures how many the overseer finds and how long it takes, and opens the discussion on why some errors pass unnoticed. At the end, participants take away tools to check for themselves, in their daily work, whether oversight is working: a map of the decisions in their role where AI is involved, and a log format that records assessments and deviations. The module works best after an [AI Rating](/en/start-ai-rating), which measures in its Risk dimension how the company handles human oversight, and before any new system that affects people is introduced. It does not replace an employment lawyer's opinion on specific cases, and we say so in the room. ## Key dates 23 October triggers the Italian employment rules, while most of the [European obligations on high-risk systems](/en/blog/compliance/ai-act-december-2-2027-fourteen-months) arrive more than a year later, and confusing the two leads companies to underestimate the first and postpone the second. | Date | What happens | |---|---| | 2 February 2025 | Article 4 AI Act on AI literacy applies, rewritten by the Omnibus from 27 July 2026 | | 30 September 2026 | Legislative Decree 160/2026 in force: disclosure of evidence, presumption of causation, Article 437-bis Criminal Code | | 23 October 2026 | Legislative Decree 179/2026 in force: Articles 41, 42 and 47, supervisory authorities and complaints | | April 2027 | Six-month deadline for professional bodies to update their training | | 23 October 2027 | Deadline to update professional fee parameters and fair compensation rules | | 2 December 2027 | Obligations for Annex III high-risk systems apply, including employment and recruitment | ## The cost of oversight A well-designed training programme has an effect few people account for: employees start contradicting the system more often, and for a few months some efficiency indicators get worse. That is the sign human oversight has started to work, and the law asks for it explicitly when it speaks of authority as well as competence. The question we would bring to a board is whether the organisation is willing to accept that cost, and to reward those who, with good reasons, decided differently from the algorithm. To find out where to start, you can [start an AI Rating](/en/start-ai-rating), [book an assessment call](https://calendly.com/fabiolalli/zerofive) or write to us at [hello@zerofive.ai](mailto:hello@zerofive.ai). *This article describes the legal text and its organisational implications and does not constitute legal advice.* **Sources.** Italian Legislative Decree no. 179 of 7 October 2026, Official Gazette, General Series no. 234 of 8 October 2026, Articles 41, 42 and 47; Italian Legislative Decree no. 160 of 9 September 2026, Official Gazette, General Series no. 214 of 15 September 2026; Italian Legislative Decree no. 152 of 26 May 1997, Article 1-bis; Italian Law no. 132 of 23 September 2025; Regulation (EU) 2024/1689 (AI Act) as amended by Regulation (EU) 2026/1744, Articles 4, 14 and 26; Artificial Intelligence Observatory, Politecnico di Milano, February 2026.