The EU AI Act gives employers until December 2027 to explain an AI hiring decision

Explaining a hiring decision is a records problem before it is a legal one. Boylan's suggested exercise is to take one real AI-assisted hiring recommendation and follow it through the entire process, beginning with what data the system used ^[16]. His benchmark list for employers has four parts: know which systems influence employment decisions, put trained people in charge with authority to override them, keep records, and tell candidates and employees when AI is part of the process ^[14].

The readiness figures come from the company making the case. Avature published the AI Impact Report 2026, and Dimitri Boylan founded Avature and is its chief executive ^[10]. On the report's own split, 51% piloting against 5% strategic, there are about ten organizations in pilots for every one using AI strategically ^[18].

In this account, the reason a non-EU employer should care is operational. Boylan wrote that large organizations do not build different cybersecurity programs or financial controls for each market, and that AI governance is likely to follow the same pattern ^[12]. The piece is headlined "Why The EU AI Act Applies To You, Even Outside Europe" ^[23], and the text does not set out the jurisdictional test that brings a US-headquartered employer within the Act, or the penalties for missing the date ^[20]. For US employers he describes a fragmented environment of federal employment law, privacy obligations and state-level rules affecting automated decision-making ^[13]. "The lesson is not to copy Europe, but to build a model strong enough to hold up across jurisdictions," Boylan wrote ^[15].

Fourteen months separate the piece's publication on September 23, 2026 from the December 2027 date ^[19]. The harder part, by Boylan's account, is a stack assembled from multiple vendors and smaller AI tools that are not easy to inspect end to end ^[22]. An employer that cannot trace a ranking back through that chain faces a scaling decision well before the requirement lands. Boylan wrote: "if the system cannot be explained, it should not be scaled" ^[9].

Documentation has to describe the system as it actually ran, and a screening tool that filters or ranks applications is doing that every day until someone changes it ^[5]. A pilot instrumented now produces its own record; one already in production has to be reconstructed from whatever the vendors kept.