Compliance
NIS2 and Artificial Intelligence: Supply Chain and the CISO
Editorial team Delphis · Published on · 6 min read
NIS2 is Directive (EU) 2022/2555, the European rule that strengthens cybersecurity for a broad set of essential and important entities. It does not regulate artificial intelligence specifically, but it requires organizations to manage the risk of the systems and suppliers they depend on. Third-party AI tools, from generative assistants to AI features built into software, fall under that requirement like any other digital service.
This article is informational and is not legal advice.
What is NIS2 and how was it transposed in Italy?
The NIS2 Directive sets a common level of cybersecurity measures across the Union. It applies to entities in certain sectors and above certain sizes, with obligations on governance, risk management, incident notification and supply chain security.
In Italy the transposition is Legislative Decree no. 138 of 4 September 2024. Whether it applies to a specific organization depends on sector and size, and this should be checked with a legal adviser before any compliance plan is drawn up. Management bodies have a direct role in approving and overseeing risk management measures: it is not a topic that can be delegated entirely to IT.
Does NIS2 regulate AI?
No. In the text of NIS2, artificial intelligence is not the subject of the rules. The relationship between the two is indirect but real: AI is a technology that organizations adopt, and like any adopted technology it introduces dependencies, risk surfaces and suppliers.
Anyone claiming that "NIS2 requires this or that for AI" goes beyond the text. The correct point is different: if the organization is subject to NIS2, it must manage the risk of its systems and suppliers, and AI tools are, for all practical purposes, systems and suppliers. The AI literacy obligation belongs to another law, the EU AI Act (Regulation (EU) 2024/1689), covered in EU AI Act obligations for companies.
Why do third-party AI tools enter the supply chain?
When a company adopts a generative assistant, an AI service via API or an AI feature embedded in existing software, it creates a dependency on an external supplier. That dependency has characteristics a CISO will recognize:
- company data leaving the perimeter, as prompts, uploaded documents or context linked through connectors;
- constantly evolving services, with models and features that change on a schedule the company does not control;
- subcontracting chains, because an AI vendor may rely on infrastructure and models from other parties;
- integrations with other systems, which widen the exposed surface if an account or token is compromised;
- ungoverned use, when employees adopt tools without going through procurement and security processes. The phenomenon is described in shadow AI risks and how to manage it.
For an organization subject to NIS2, these elements belong inside the supply chain risk management process, not in a separate track "for AI".
What should you ask AI tool vendors?
A list of operational questions makes the assessment repeatable. Answers should be documented and reviewed at defined intervals.
- Where and how is data processed? Location, subcontractors involved, how long prompts and outputs are retained.
- Is company data used to train the models? What is the default setting and how is it configured for the business contract.
- What access controls are available? Authentication, role management, exportable access and usage logs.
- How are security incidents handled? Timing and method of notification to the customer, contacts, channels.
- What certifications and independent assessments exist? Ask for evidence, not just statements.
- How are vulnerabilities and updates handled? Reporting and remediation process, communication of significant changes.
- How do you exit the service? Data export and deletion, termination conditions.
- What are the contract terms on personal data processing? The link with the GDPR is covered in GDPR and generative AI at work.
These questions are not a list prescribed by NIS2: they are a practical application of the supplier risk management principle. If a vendor does not answer or answers evasively, that fact is itself relevant information for the assessment.
Example: what changes when an AI feature is built in
Suppose a vendor of an ERP system you already use turns on a new feature based on generative AI. Contractually the vendor is the same, but the risk profile may have changed: new data flows to third-party services, new subcontractors, new default settings.
A mature process treats a change like this as a trigger for reassessing the vendor, rather than letting it pass unnoticed. For this reason it is worth including in contracts and procedures an obligation to communicate significant changes, and a periodic review of the list of AI tools, even when they are embedded in existing products.
How do you fold AI tools into the existing process?
There is no need to build a new process. It is better to extend the one already used for critical suppliers, adding a few specific elements:
- a single entry point to request approval of a new AI tool, including features activated in products already purchased;
- a classification of the data that may be sent to each tool;
- a periodic review, at a frequency proportionate to criticality, of the list of tools and of vendor answers.
The goal is for the assessment of AI vendors to be routine and repeatable, not an exception handled from scratch every time.
What is the CISO's role?
The CISO is the person who can make this perimeter manageable, provided they have the means to do so. In practice, the contribution has four parts.
- Inventory. Obtain an up-to-date list of AI tools in use, official and unofficial. Without visibility there is no risk assessment.
- Requirements. Define the minimum security criteria an AI vendor must meet to be approved, consistent with those for other critical suppliers.
- Assessment and approval. Take part in tool selection together with IT, procurement, the DPO and business functions, so that the choice is not made on cost or features alone. For comparing the tools, see Copilot, ChatGPT and Claude: how to choose.
- Monitoring. Follow actual use, service changes and incidents over time, reporting to management bodies.
Working with the DPO is natural: GDPR and NIS2 look at the same tool from different angles, personal data processing on one side, security and resilience on the other. A company AI policy shared between the two functions avoids conflicting requirements.
Where Delphis fits
Delphis is a vendor-independent platform that connects spend, adoption, impact and regulatory obligations. For the CISO it offers a single view of the AI tools in use, the linked use cases and the obligations to verify for each, including the vendor-related risk profile, with data on people always aggregated by role or department. The mapping helps identify gaps but is not legal advice. Learn more: AI compliance with Delphis.
Sources
Frequently asked questions
- Does NIS2 regulate artificial intelligence?
- No, NIS2 has no AI-specific provisions. It covers the security of network and information systems and risk management, including supplier risk. Third-party AI tools fall within that scope like any other digital service.
- How does NIS2 apply to third-party AI tools?
- A generative assistant or an AI feature built into software is a supplier to assess within supply chain risk management: inventory it, evaluate its security guarantees, define contract terms and monitor it over time.
- Which Italian law transposed NIS2?
- Italy transposed NIS2 with Legislative Decree no. 138 of 4 September 2024. Whether an organization is in scope, and with which obligations, depends on sector and size and should be checked with a legal adviser.
- What is the CISO's role with AI tools?
- The CISO brings AI tools into the risk management process: requires an inventory, sets the security requirements to ask of vendors, takes part in their assessment and oversees monitoring and incident handling.
Want to see it on your own AI tools?
A demo on your use cases, with our team and no commitment.
Request your obligations map