Special regime for high-risk AI legacy systems: What companies need to do now

Which cut-off date applies?

For high-risk AI systems under Annex III — including AI systems used to assess the creditworthiness of natural persons (Annex III No. 5(b)) — the relevant cut-off date is currently 2 August 2026.

As part of the Digital Omnibus package, the European Commission has proposed postponing this date to 2 December 2027. However, the legislative procedure has not yet been completed, so 2 August 2026 remains binding for the time being.

In addition, there is ongoing discussion as to whether a possible extension would affect only certain penalty provisions, while the substantive requirements might still apply earlier. Companies should therefore continue to take the current cut-off date seriously and align their planning accordingly.

Which AI systems are affected?

The AI Act provides for two independent connecting factors under which a system may qualify as high-risk AI. First, where it constitutes a safety component of a product that is already subject to Union harmonisation legislation — such as medical devices, vehicles or machinery (Art. 6(1) in conjunction with Annexes I and II AI Act). Second, where it is expressly designated in Annex III of the AI Act, which covers certain use cases irrespective of the product context. For companies in the private sector, the following Annex III categories are particularly relevant in practice:

• Employment and personnel management (No. 4): AI used for applicant selection, performance monitoring, or decisions on promotion and termination — relevant for almost any company using AI in HR.
• Access to essential private services (No. 5(b) and (c)): Assessment of the creditworthiness of natural persons and credit scoring (relevant for banks, financial services providers and fintechs), as well as AI-supported risk assessment and pricing in life and health insurance. Pure fraud detection systems are exempt.
• Biometrics (No. 1): Remote biometric identification, categorisation based on sensitive characteristics, and emotion recognition — for example in customer interactions or in the workplace. Many of these systems may also fall under the prohibited practices in Art. 5 AI Act and require separate legal assessment.
• Critical infrastructure (No. 2): Also relevant to private energy suppliers and operators of digital infrastructure where AI is used as a safety component in ongoing operations.
• Education and vocational training (No. 3): Private educational institutions and e-learning platforms should examine whether their systems perform admission, evaluation, or exam monitoring functions.
• The areas of law enforcement (No. 6), migration (No. 7), and administration of justice and democratic processes (No. 8) are primarily aimed at public authorities. One exception is No. 8(b), which may also cover private actors in the field of political advertising technology.

What does special regime mean for AI systems?

The EU AI Act contains an important exception for existing systems. High-risk AI systems that were already placed on the market or put into service before the relevant cut-off date may, in principle, continue to be operated.

The key point is this: the extensive obligations for high-risk AI apply to such legacy systems only if their design is substantially modified after the cut-off date.

If the original design remains unchanged, these systems may continue to be used — without the need to implement the full high-risk requirements retroactively.

When does special regime cease to apply?

Special regime does not apply indefinitely in every case. It ceases to apply where there is a substantial modification of the system design after the cut-off date.

What exactly qualifies as substantial has not yet been fully clarified in legal terms. What is clear, however, is that changes to core technical components may jeopardise special regime.

Typical examples of potentially substantial modifications include:

• adjustments to the operating software
• changes to the system architecture
• fundamental technical redesigns
• structural changes to the way the system functions

Whether a change is in fact substantial must always be assessed on a case-by-case basis.

Distinguishing this from other changes under the AI Act

The assessment of a “substantial modification” in connection with special regime is not identical to other concepts of change used in the AI Act.

While other provisions primarily address alterations or modifications to systems already on the market, special regime specifically hinges on the design of the original system.

The relevant criteria are therefore not fully identical. In practice, existing standards may provide useful guidance — but they cannot simply be applied one-to-one.

The most important step before the cut-off date: documenting the system design

In order for companies to rely on special regime later on, proper documentation of the current system design is essential.

In particular, it is advisable to document:

• the categories of data used during training and in production
• the system’s typical input and output
• relevant decision thresholds
• the specific business purposes of the system
• the technical setup
• training methods, parameters, and training procedures
• anonymised sample inputs to describe the system’s behaviour

This documentation creates clarity as to exactly which design is protected.

After the cut-off date: control changes carefully

After the cut-off date, every system change should be reviewed in a structured manner.

Since it has not yet been conclusively clarified which changes qualify as substantial, a systematic change management process for AI systems is recommended.

Such a process should in particular:

• capture planned changes to the system
• document changes
• provide for internal review steps for technical adjustments
• also take seemingly minor changes into account

This allows companies to identify at an early stage whether an adjustment could jeopardise special regime.

Why the cut-off date should still remain on your radar

The currently applicable cut-off date for high-risk AI systems is 2 August 2026.

Although a postponement of this date has been proposed as part of a European legislative package, the corresponding procedure has not yet been completed.

In addition, there is discussion as to whether a possible extension would affect only certain penalty provisions, while the substantive requirements might still apply earlier.

For companies, this means: the cut-off date should continue to be taken seriously and factored into planning.

Conclusion

Special regmine for high-risk AI may provide important relief for many companies. Systems that are already in use may, in principle, continue to be operated — provided that their original design is not substantially modified.

What matters, however, is early preparation: companies should document their existing system design in good time and establish a structured change management process. Only then can special regime be safeguarded in the long term — regardless of whether the cut-off date ultimately remains 2 August 2026 or is postponed to 2 December 2027.