The Signal · Column 082

The EU AI Act Just Created a Hiring Problem

Andrei van Noordt · 13 October 2026 · 2 min read

The news this week is procedural rather than dramatic. The EU AI Act's compliance obligations and deadlines are now being spelled out for small and mid-sized companies, which is the part of the market that assumed this was a problem for the large platforms. It is not. If you build, buy, fine-tune or resell a model that touches hiring, credit, education or safety, you are inside the scope, and the calendar is not asking your opinion.

Most companies are reading this as a legal project. They will ask their counsel to produce a memo, their counsel will produce a very good memo, and the memo will describe obligations that nobody in the building can actually discharge. Documenting training data provenance, logging model decisions, explaining an output to a regulator, running a risk classification on a system you did not build: none of that is work a lawyer can do alone, and none of it is work an engineer can do without someone translating the statute. The gap between the memo and the implementation is where this gets expensive.

Compliance is no longer a legal function with a technical attachment. It is a technical function with legal consequences.

So the scarce person in this market is not the compliance officer and not the machine learning engineer. It is whoever sits between them. I am already seeing mandates that did not exist eighteen months ago, roles that want a data or ML background plus the patience to read regulation closely, and companies are struggling to brief them because the job title has not settled yet. Meanwhile the ML engineers who happen to have documentation discipline, who wrote model cards before anyone demanded them and kept their evaluation pipelines legible, are quietly becoming the most defensible hires on the team. Rigour used to be a personality trait. It is now a compliance asset.

There is a second-order effect worth naming. Once a company has to document how its systems make decisions, it can no longer tolerate the undocumented ones. That means the engineers who built the opaque internal tooling become load-bearing in a way nobody planned for, and the teams that let them go in a cost-cutting round will be reconstructing their own systems from memory, under a deadline, for an auditor.

For hiring leaders, the practical read is simple. Before you write another model-building job spec, find out who in your organisation can read the regulation and the architecture in the same sitting, and if the answer is nobody, hire that person now, while the role is still mispriced.