Arthur Keleti, a cyber‑secrecy futurist, IT security strategist and founder of the Informatikai Biztonság Napja (ITBN), warns that companies should not be complacent about the EU AI regulation. Although some deadlines were postponed, many obligations are already in force, and firms must quickly determine which AI tools they actually use, in which business processes and with what data.
Keleti will join the AI Talks 2026 conference panel “Breathing space but no pardon: what company leaders should know about the EU AI regulation” on November 18 at Bálna Budapest. Ahead of the event he explained what the recent postponements mean in practice and outlined concrete steps executives can take.
Which deadlines were postponed, and what already applies?
The AI Omnibus delayed the application of certain rules. According to Keleti, the postponements mainly affect high‑risk systems: for some uses such as certain credit scoring, HR processes and biometric identification systems, detailed rules were pushed to December 2, 2027, while high‑risk AI embedded in regulated physical products was deferred to August 2028.
However, several provisions already apply: rules on prohibited AI practices and on AI literacy have been in force since February 2025; from August 2026 additional transparency requirements took effect; and on December 2, 2026 prohibitions related to watermarking older generative systems, certain ‘‘nude‑creation’’ apps and child‑sexual content will become applicable. Keleti stresses that the delay is not a waiver—there remain immediate obligations well before the end of 2027.
How can a company tell if a system is high‑risk?
Keleti says the starting point should not be a technical definition but an inventory of uses: a company must list which AI tools it uses and for what purposes. HR is a clear example: recruitment, promotion, performance evaluation or dismissal decisions can quickly trigger high‑risk classification. An Applicant Tracking System (ATS) that was formerly administrative can become high‑risk once it starts ranking candidates with AI. The same applies to credit decisions, certain life and health insurance use cases and parts of the justice system.
For leaders, the key question is whether AI makes decisions about people or substantially affects such decisions. High‑risk systems require ongoing oversight: appoint responsible personnel, monitor input data relevance and system behaviour, keep incident logs (retained for at least six months) and file mandatory incident reports. Worker representatives and all affected parties must be informed. For credit or insurance uses, the AI Act requires a Fundamental Rights Impact Assessment (FRIA) for public authorities and private sector users of high‑risk systems. These obligations may seem straightforward but together form a significant organisational workload.
Keleti warns that executives cannot rely solely on AI vendors to ensure compliance—a vendor’s signature does not equal certification.
Shadow AI: when does leadership become responsible?
Shadow AI—employees using AI tools outside official corporate control—is a major challenge. Keleti argues leaders can only claim non‑responsibility while they truly could not have known; once aware, denial is indefensible. International and domestic studies show many employees use personal accounts or devices for work, so organisations that claim no shadow AI likely simply lack visibility.
Blanket bans are a poor strategy: the stricter the ban, the more shadow AI tends to appear. Companies should provide controlled corporate tools under clear contractual terms, implement data classification so employees know what may or may not be uploaded, and introduce monitoring.
Building AI literacy without turning it into another e‑learning box‑tick
The Omnibus from July 2026 requires organisations to support AI literacy but does not mandate guarantees. Keleti supports role‑based training: HR staff working with AI filters need different skills than a manager who acts on AI outputs. Training should show real errors and biases, teach proper data handling, and clarify what may or may not be uploaded.
Policies cannot be one‑off documents; they should be updated regularly. A 2024 AI policy, for example, might not even mention agents, which are relevant today. Annual or semi‑annual reviews are necessary.
What if an existing HR, CRM or finance product suddenly gains an AI feature?
Keleti calls this one of the most underestimated risks in 2026–2027. Software vendors face intense pressure to add AI functions, so a product can gain a capability that shifts it from low to high risk almost overnight — for instance, an HR system that was administrative yesterday may start ranking candidates automatically today.
He recommends that new AI features be disabled by default. Vendor contracts should specify which AI features exist in the product and their classification. Change management should explicitly verify whether an update affects AI functionality. Organizations should read release notes — and, ironically, may use AI to monitor those release notes and flag changes that could alter risk classification. The AI inventory must then be updated.
When is human‑in‑the‑loop real oversight?
A simple test is the time allowed for the human decision: if a task requires approval in twenty seconds but would normally take minutes or hours, the human is likely a figurehead. Also examine how often humans override the system; if that rate is effectively zero, it is suspicious. Can the human stop the process or only nudge it? Do they have access to information the system lacks? Do they understand why the suggestion was made?
Organisationally, what happens to a person who stops the process? Is the organisation protecting them for doing their job, or penalising them for slowing operations? If the latter, human‑in‑the‑loop exists only on paper. True oversight requires real decision power, sufficient information and time, and organisational empowerment.
Don’t treat regulations as separate islands
Alongside the AI Act, companies must consider the GDPR, NIS2, and in the financial sector DORA, plus other sectoral rules. Keleti says treating these separately is a mistake—data protection, cybersecurity, supplier and AI risks intersect. An AI incident can simultaneously affect personal data, information security, third‑party services and AI Act obligations.
Existing NIS2 processes for risk management, audits and incident response and DORA’s focus on third‑party tech risks in finance should be starting points for AI risk management, rather than building new, isolated compliance silos for each regulation.
Practical immediate steps for Hungarian executives
Keleti suggests that within a month many companies can take basic steps: designate a responsible person, run an anonymous, amnesty‑style survey to map which AI tools employees use (the goal is visibility, not blame), and ask how those tools are used. Compare procurement and contractual records with actual usage and check for potentially prohibited practices.
That yields an initial AI usage inventory and real situational picture. From there, companies can proceed to vendor due diligence, policy drafting, training and in‑depth reviews of higher‑risk systems.
What should we expect next?
Keleti is focused on defence: manual methods cannot effectively counter systems that operate orders of magnitude faster. He asks whether companies already have staff specifically responsible for AI risk and oversight, whether they maintain an AI usage inventory, and whether they understand the kinds of systems they face — particularly those capable of autonomous decisions. He warns that organisations are only beginning to grasp the security and organisational consequences of such systems.
Tickets for the 2026 HVG AI Talks conference are still available at aitalks.hu. The content was produced within the AI Talks event with HVG BrandLab’s cooperation; HVG weekly and hvg.hu editorial teams did not participate in its production or editing.
