Twenty months of the AI Act, from a builder's seat

The AI Act entered into force on 1 August 2024. We are now twenty months in. The prohibition tier has been live since February 2025, the GPAI obligations since August 2025, and the high-risk obligations are due to land in August. From inside a lab whose customers care about this regime, here is what has and has not changed.

Procurement language changed. RFPs that two years ago asked for "GDPR compliance" now ask, in addition, where the inference happens, which providers see the prompt, what the data-flow diagram looks like, and — for high-risk uses — what the post-market monitoring plan is. The questions are not always well-posed. Some confuse provider obligations with deployer obligations. Some treat GPAI as a synonym for any LLM-shaped model regardless of compute threshold. Some ask about a fundamental-rights impact assessment when no fundamental-rights-implicating use is on the table. But the questions are present, where two years ago they were absent, and that is the larger fact.

The cost of the "we'll figure compliance out later" posture changed. In 2024 it was possible to ship an LLM-flavoured product to an EU customer and treat the regulatory layer as a follow-up. The customer's internal review processes had not caught up. By April 2026, they have. We see deals stall not because the technology is wrong but because the documentation isn't there — the model card, the data governance description, the conformity-assessment readiness for high-risk deployments, the chain of evidence that connects a customer's policy to a vendor's behaviour. Vendors who started producing those artifacts in 2024 are clearing review faster than vendors producing them now under deadline pressure. The competence-gap is now an artifact-gap.

What did not change, despite confident predictions: the number of high-risk deployments did not collapse. The expectation, in some early commentary, was that the threat of conformity assessments and post-market monitoring would push organisations toward we'll just say it isn't high-risk workarounds at a scale that emptied the high-risk tier of real systems. In our customer base, that has not been the dominant move. The dominant move has been to scope the high-risk system narrowly within a larger product so that the obligation set is bounded. The system that scores a job applicant remains in scope; the system that suggests interview questions to a human recruiter often sits in a quieter neighbourhood. This is sometimes legitimate (the second system is genuinely a different system) and sometimes a workaround (it is the first system with a human stapled to the front). The regulation will be tested on this boundary. We expect the case law to be slow and the guidance to be faster.

What did not change: the strongest closed models are still mostly outside the EU. The argument that the AI Act would chill EU model development relative to the US has neither been confirmed nor rebutted in the form commentators expected. EU GPAI providers exist — Mistral chief among them, with Regolo.ai consequential at the inference and privacy-mode layer — and their open-weight models are competitive at the tier where most operational EU deployments actually sit. The closed frontier models are a different procurement conversation about a different problem, mostly run by organisations who would, on independent grounds, not have used those models for sensitive workloads anyway. The EU did not "lose" the frontier; it largely opted out of the frontier-as-procurement-target. That is a substantive choice, not a failure, and the framing in 2024 that treated it as the latter has aged badly.

What changed inside the lab. We stopped writing AI Act compliance as a feature and started writing it as a substrate. The compliance layer of our products is no longer a checkbox bolted onto the front; it is the shape of the schemas, the structure of the audit records, the contract between the gateway and the calling application. None of this is "AI Act features." It is how the products work. The regulation is to our software what GDPR was to a relational database in 2018: not a layer, a constraint that shapes the schema.

What we still don't know. How the AI Office and the national competent authorities will exercise enforcement once the high-risk obligations land in August. The first wave of interventions has been low-key — guidance, a few information requests, a handful of correspondence-level matters. The high-risk obligations will be the first real test of whether the conformity-assessment apparatus has the bench depth to keep up. We're cautiously optimistic without being confident. The right reference is the slow build of GDPR enforcement after 2018, not the dramatic intervention. DPAs took the better part of a decade to settle into the GDPR posture they have today. The AI Office is unlikely to be faster than that in absolute terms, and the systems it must reason about are harder.

What we'd say to a builder reading this and wondering whether the regime is worth taking seriously. Yes, and not because the fines are scary. Because the customers — the ones whose buy decisions are tied to compliance review — already are.