IMY explains which recruiting AI is high-risk under the AI Act

What IMY says

IMY, the Swedish data protection authority, has published a Q&A on the AI Act aimed at Swedish organisations ^[1]. The guidance clarifies that AI systems used to recruit or select people for employment are high-risk under point 4 of Annex III.

That includes systems that target job ads at specific groups, filter or rank applications and evaluate candidates. It does not include simple calendar tools or ordinary email.

Think of it like this: if the AI decides who sees the ad or who gets through to the next round, it's affecting people's chances of getting work, and that puts it in the high-risk category.

What high-risk means in practice

From 2 December 2027, providers of high-risk AI must have several things in place before they can sell the system ^[2]:

• Risk management: identify and reduce risks throughout the system's life.
• Data governance: training, validation and test data must be relevant, representative and documented, including where the data came from.
• Technical documentation: everything a notified body needs to assess the system.
• Automatic logging: the system must technically allow for automatic recording of events over its lifetime.
• Human oversight: measures so a person can understand the system's output and intervene.
• Quality management: procedures to keep the system meeting the requirements.
• Conformity assessment: the provider goes through an assessment by a notified body (or, in some cases, an internal assessment) and must affix a CE marking before placing the system on the market.

Providers who already have a quality management system under ISO 13485 or similar will recognise parts of this. The logging and data governance are newer ground for many.

What users must do

Organisations that use high-risk recruiting AI have duties too, under Article 26 ^[3].

Before taking the system into use, they must inform employee representatives and give employees information about the AI in a clear and accessible way. That means telling people the system is there, what it does and how it affects decisions about them.

They must also:

• use the system according to the instructions,
• provide human oversight,
• monitor how the system works in practice,
• keep the logs they control for at least six months,
• inform people affected by decisions that a high-risk AI system was involved.

The last point matters. If a candidate is turned down and a high-risk AI played a part, they have a right under Article 86 to get a clear explanation of the AI's role and the main elements of the decision ^[5].

The logging question

Article 12 asks high-risk systems for automatic logs of what happened, covering risk situations, post-market monitoring and the user's oversight ^[2]. It says what to log. It does not say how to protect the logs once they exist.

That is a separate question, and a practical one: when a decision is questioned months later, can you show the log is the one written on the day?

A log kept by the same people who could edit it is easier to challenge. Here, the thing to seal is each decision record as it is written: the candidate reference, the screening model's version, the score or result and the human reviewer. Sealing takes a fingerprint of the record, not the record itself, so the candidate's data never leaves your systems.

If the decision is disputed six months later, anyone you share the record with—an auditor, a supervisor or the candidate's representative—can check that the fingerprint still matches. Change a single character in the log and it will not match.

Auditors, supervisors and, if it comes to it, a court can check the record. The seal does not show that the decision was fair or lawful, or that the log is complete. It shows that the record is exactly what was written at that time.

You seal log records as they are written, with the REST API, the SDK or the MCP server that connects to WitniumChain. Your system sends a fingerprint, not the record, and gets a witness ID back. The first 2,000 witnesses a month are free. The documentation explains how it works.

How it would work

1. Your recruiting system writes a decision to the log and sends a fingerprint of that record.
2. The fingerprint is sealed, and you get a witness ID.
3. Months later, if the decision is questioned, you calculate the fingerprint again and check it against the sealed one.

What this means for you

If you provide or use AI in recruiting, here are four questions for this week:

• Does your system target ads, filter applications or evaluate candidates? If yes, it is likely high-risk.
• If you are a provider, do you have documented risk management, data governance, logging and a quality management system in place?
• If you are a user, have you informed employee representatives and made sure employees know the AI is there?
• Can you show, six months from now, that a logged decision has not been edited since the day it was made?

IMY's Q&A also covers other areas, including general-purpose AI, prohibited practices and transparency duties. It is worth reading in full if you work with AI and Swedish users or employees ^[1].

What to watch

The high-risk requirements for Annex III systems, including recruiting AI, apply from 2 December 2027 ^[4]. For systems under Annex I (components of critical infrastructure), the date is 2 August 2028.

The European Commission's AI Office is expected to publish further guidance on conformity assessment and notified bodies during 2025.