The AI Act's Quiet Second Act: What Article 6 Now Asks of Doha and Riyadh

There is a stage in every European file when the political theatre subsides and the implementing acts arrive. The AI Act is in that stage now. The plenary photographs were taken long ago, the Council compromise has been chewed over by lawyers in three languages, and what remains is the unglamorous, decisive question: who classifies what as high-risk, and on what evidence.

The operative provision worth re-reading this week is Article 6. It does not, contrary to a persistent misreading in the trade press, deem an AI system high-risk because someone in a ministry feels uneasy about it. It deems a system high-risk when it is intended to be used as a safety component of a product already covered by Union harmonisation legislation listed in Annex I, or when it falls within the use cases enumerated in Annex III. The distinction matters because the second category — biometric categorisation, critical infrastructure, education access, employment screening, essential private and public services, law enforcement, migration, and the administration of justice — is the one that now reaches into every procurement file written in Brussels, and, by extension, into the contracts that Gulf vendors are quietly redrafting.

The compromise that produced the current text was, as ever, a triangulation. Paris pushed late and hard for a narrower derogation on remote biometric identification, having lost on the foundation-model thresholds it wanted lower. Berlin, having spent capital protecting its industrial Mittelstand from the brunt of the conformity assessment regime, conceded ground on general-purpose model obligations in return for a longer transitional runway. Rome carried water for the law enforcement carve-outs. The smaller member states — and here the Dutch presidency files repay close reading — extracted clearer language on the role of national competent authorities, which is the sentence that determines whether enforcement is real or decorative.

That brings us to the test that matters. A regulation is what its enforcement is. The AI Office, sitting within the Commission, has formal competence over general-purpose models, but the high-risk regime is enforced through a patchwork of national market surveillance authorities, notified bodies, and the data protection authorities that have informally annexed adjacent territory. France's CNIL has already signalled it will read its mandate broadly. The German Länder-level supervisors will, predictably, not move in unison. Ireland's authority, which inherits a disproportionate share of the work by virtue of corporate domicile, is staffing up but remains the bottleneck through which most of the heaviest files will pass. The picture, in short, is one of asymmetric enforcement velocity. Companies will discover, file by file, which capital is the slow one and route their compliance posture accordingly.

For the Gulf, the transmission mechanism is no longer theoretical. Three things have changed in the last quarter. First, the model documentation obligations under the general-purpose tier mean that any Emirati or Saudi entity offering a model on the Union market — directly or through a European deployer — must produce the technical file and the training-data summary in a form a European regulator can read. Second, the high-risk classification reaches deployers, not only providers, which means a Doha-based bank running a European-built screening model on its Frankfurt branch is squarely within scope. Third, the sovereign AI investment vehicles in Abu Dhabi and Riyadh have begun writing AI Act conformity warranties into the term sheets of their European joint ventures. That is not a legal nicety. That is the regulation doing extraterritorial work without a single line of treaty text.