A customer support chatbot that does not inform the user that they are speaking to a machine, a workstation whose NPU remains disabled due to outdated drivers, a content generation tool that does not indicate the artificial origin of the content: since August 2, 2026, these situations are no longer considered negligence but regulatory non-compliance. Current IT trends are reflected as much in legal obligations as in hardware catalogs.
AI Act and Omnibus Regulation: What Changes for IT Teams
The European regulation on AI came into general application on August 2, 2026. We are no longer talking about regulatory monitoring: companies deploying AI systems in sensitive areas (human resources, credit, education, public services) must document their uses, ensure effective human oversight, and guarantee transparency to users.
The Digital Omnibus Regulation on AI (EU 2026/1744), which came into effect on July 27, 2026, has reshaped the timeline. The heaviest obligations regarding high-risk systems have been postponed to 2027 and 2028, but transparency requirements apply immediately.
To explore the entire IT universe of D’clic Lab, it is clear that these developments affect both the developer integrating a language model and the system administrator configuring an automated ticket sorting tool.
In practice, IT teams must adapt three things: internal documentation of automated workflows, legal notices on user interfaces, and audit procedures for deployed models. The effort is significant, but the Omnibus timeline allows some leeway for systems classified as high risk.
Chatbots and AI-Generated Content: Marking Obligations in 2026
Since August 2, 2026, any interactive AI system (chatbot, virtual assistant, support conversational robot) must explicitly inform the user that they are interacting with AI and not a person, unless this is already obvious from the context. This rule applies to websites, mobile applications, and messaging platforms.
Deepfakes and AI-generated or modified content must be clearly marked. We are talking about visible marking in the interface, not a mention buried in terms and conditions that no one reads.
For a tech team, this means reviewing user journeys at every automated touchpoint. Feedback varies on the best way to implement this marking (banner, icon, textual mention), but the principle remains non-negotiable: the user must know they are not speaking to a human.
Checklist for a Quick Audit
- Each conversational interface displays a visible AI mention before the first exchange, without the user needing to click or scroll.
- AI-generated content (texts, images, videos) carries a readable mark at the point of dissemination, not just in the metadata.
- An internal register lists the deployed AI systems, their purposes, and the associated risk level, to anticipate upcoming obligations in 2027.
NPU and Local Processing: When Hardware Dictates Software Practices
Since 2024, manufacturers like ASUS and Lenovo have been delivering machines equipped with NPUs dedicated to local AI processing. Microsoft’s Copilot+ range has popularized this approach. In practice, there is a gap between the available hardware and actual usage.
An NPU allows for tasks such as real-time translation, searching through a file history, or assisting with writing without sending each request to a remote server. Local processing reduces latency and limits data exposure.
The concrete problem: many workstations delivered with an NPU do not utilize it, due to outdated drivers or incompatible applications. Before renewing a fleet, it is wise to check that business software effectively takes advantage of this chip. An unused NPU is a hardware cost without operational benefit.
Selection Criteria for an AI Workstation
- Check the compatibility of business applications with the embedded NPU (the list of certified Copilot+ software is regularly updated).
- Evaluate the actual latency gain on daily tasks: if most work is done through a web browser, the NPU offers little benefit.
- Compare energy consumption for local AI processing versus a cloud call, especially for fleets of more than a few dozen workstations.
- Prefer a model from a manufacturer that regularly publishes NPU driver updates.
AI Compliance and Cybersecurity: Two Converging Projects
Security teams and compliance teams often worked in parallel. The AI Act forces a convergence. Documenting an AI model for compliance also produces a useful inventory for cybersecurity, as it identifies data flows, access to models, and potential failure points.
The Zero Trust approach, already adopted by many organizations for network access management, naturally extends to AI systems. Every call to a model, every transfer of training data, every weight update must go through an identity and rights verification.
This is not a theoretical project. When a generative AI tool is connected to a customer database to personalize responses, a flaw in the model exposes the same data as a classic network breach. The attack surface expands with each un-audited AI deployment.
The upcoming deadlines of the AI Act, in 2027 and 2028, will add requirements for high-risk systems. Teams that document and audit their deployments now will absorb these obligations without major redesign. Those that wait will have to catch up on several years of compliance in a few months, with the cost and stress that entails.