What the data protection officer asks in the second meeting

The first meeting usually goes well. The exposure is real, the operational case is clear, the product demonstrates convincingly.

The second meeting is different, because the data protection officer is in the room.

What employee data is being processed. On what lawful basis. Whether the works council has been consulted. What happens to a risk score if it ever surfaces in a dismissal.

The answers get improvised, and the programme goes quiet.

It is tempting to read that as institutional caution. It is not. Those questions are correct, legally grounded and frequently decisive, and a programme that cannot answer them is not a programme yet.

What changed in 2026

Anyone building a case on European employment data needs the current position, because a good deal of collateral in this market is now out of date.

The EU AI Act's high-risk obligations for employment and worker management systems were due to apply from August 2026. They no longer do. The Digital Omnibus on AI entered into force on 27 July 2026 and deferred them to 2 December 2027.

Two things follow:

A case resting on an August 2026 deadline has expired, and a well-briefed compliance officer will know it.
The deferral reflects unreadiness in standards and conformity assessment rather than any withdrawal of the obligations, with preparation expected in the interim. That leaves a defined window in which human oversight and documentation can be designed into a system rather than retrofitted onto one.

Data protection moved in the opposite direction. The broader Omnibus proposed narrowing what counts as personal data once pseudonymised. In February 2026 the European Data Protection Board and the European Data Protection Supervisor objected, stating the proposal went well beyond a targeted amendment, did not reflect the case law, and would significantly narrow the concept of personal data.

The practical reading is straightforward. Hashing is a strong safeguard. It is not an exit from the GDPR, and any supplier implying otherwise is one question away from an uncomfortable meeting.

The design decisions that survive the second meeting

What separates a programme that deploys from one that stalls is rarely the strength of the argument made for it. It is what was constrained before the argument was needed.

Anonymisation at ingestion, not at reporting. There is a meaningful difference between processing identified records and applying access controls to the output, and removing identifiers before data reaches the analysis layer at all. Identity sits in an encrypted vault the client alone controls, and de-anonymisation runs through a documented workflow with an audit record. When HR asks who can see what, the answer is architectural rather than a policy undertaking.
Structural exclusion of categories. A system that could technically ingest medical data, social media or off-duty activity but undertakes not to is one configuration change from a serious problem. A system where those are out of scope by design is a different conversation. What remains is objective operational data.
A human always decides. A risk score is an investigative prompt, not a verdict. Human review on every escalation, with reasoning recorded, keeps the design inside GDPR Article 22, and it is also the oversight capability the AI Act will require by December 2027.
A stated retention period. A configured maximum turns an open-ended anxiety into a specific fact.
A transparency notice for the workforce. Legal basis, data categories, retention, and the Article 21 right to object. This is usually what the works council cares about most, and it is the difference between a programme people know about and one they discover.

The argument security teams tend to leave out

There is a fairness case here that HR finds more persuasive than anything about threat.

Population thresholds are blunt. Set against an average, they flag everyone whose working life legitimately differs from the norm: permanent nights, part-time, adjusted hours for caring responsibilities, roles that require unusual movement. Those flags cluster in predictable demographic patterns, and defending them at tribunal is unpleasant.

Baselining each person against their own established pattern inverts that. The signal becomes deviation from self rather than deviation from average. Fewer alerts, better ones, and materially less exposure on fairness grounds.

It is also the more honest position. The organisation is not asserting what normal looks like. It is observing what normal looks like for each person, and noticing when that changes.

Programmes that arrive at the second meeting with a completed Legitimate Interests Assessment, a DPIA framework, a transparency notice, documented access controls and a stated retention period are having a conversation about implementation.

Programmes that arrive with a demonstration and an undertaking to sort the paperwork afterwards are having a conversation about principle.

Those take considerably longer, and a good number of them never finish.

Iveron. Predict. Detect. Protect. From insider threat.