The EU AI Act's Cost to American Innovation: What the ITIF Critique Gets Right and What It Leaves Out

On 22 September 2026, the Information Technology and Innovation Foundation published a blog arguing that the EU AI Act's compliance costs are a drag on American AI innovation. The piece is a policy argument, not a report of a specific AI system breaking. But it lands on a real operational problem that governance teams face every week: the cost of proving you comply, not the cost of complying.

What the ITIF piece actually argues

The core claim is straightforward. American AI firms selling into Europe face a stack of obligations — technical documentation, risk management systems, human oversight design, post-market monitoring, incident reporting — and the fixed cost of building that machinery does not scale down for smaller companies. A startup with one model in one market pays roughly the same to build an evidence trail as a large firm with fifty models. ITIF's point is that this fixed cost is a tax on entry.

That is a fair observation. It is also not new. The same argument was made about the General Data Protection Regulation in 2016, about medical device rules, and about financial services conduct rules before that. The question for a governance team is not whether the argument is persuasive in Brussels. It is whether your firm is paying more than it needs to.

The failure mode this represents

The pattern here is not a broken model. It is a governance program built as a project rather than a system. Firms read the EU AI Act once, produce a documentation pack, file it, and move on. Then the model gets retrained, the use case shifts, a new country goes live, and the pack is stale. When a regulator or a customer asks for evidence, the team scrambles for weeks.

That scramble is the real cost. It is not the cost of the rules. It is the cost of not having a live map of which systems are in scope, which obligations attach, and where the evidence sits. The Netherlands fraud-risk system case in 2021 is the extreme version: an automated risk scorer wrongly flagged thousands of families, the cabinet resigned, and the failure was not the algorithm alone. It was the absence of a fundamental rights impact assessment and real human oversight before deployment. Those are exactly the controls the EU AI Act now requires for high-risk systems.

The obligations that actually apply

If you sell AI into the EU, the transparency rules under Article 50 have been in force since 2 August 2026. Providers must mark synthetic content in a machine-readable way, and systems already on the market need that marking retrofitted by 2 December 2026. That deadline is close. If you have a generative feature live in Europe and no marking plan, you are behind.

The high-risk rules apply later but the preparation does not start later. Annex III high-risk obligations apply from 2 December 2027, and embedded or Annex I systems from 2 August 2028. Those dates were set by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026. For high-risk systems, the law requires a risk management system, technical documentation, logging, human oversight, accuracy and robustness testing, and a fundamental rights impact assessment for certain deployers.

That is a lot. But most of it is the same evidence a serious internal risk review would produce anyway. The firms that struggle are the ones doing it twice — once for the regulator and once for themselves.

Where the cost actually comes from

Three things drive the bill.

• Scope ambiguity. Teams do not know which of their systems are high-risk, so they either over-document everything or miss something and fix it under pressure.
• Stale evidence. Documentation is written once and never updated when the model or use case changes.
• Slow evidence assembly. When a customer or regulator asks for the pack, it takes weeks to pull together because it lives in five places.

The ITIF argument focuses on the first. The other two are self-inflicted and cheaper to fix.

What a lean program looks like

A workable program does three things well. It keeps a current map of every AI system, its jurisdiction, and its risk tier. It attaches obligations to that map so you can see, at a glance, what each system owes and by when. And it keeps evidence in one place so a complete pack can be produced on request rather than reconstructed.

None of that requires a large team. It requires treating governance as a living register, not a filing cabinet. A governance platform can help here, but the discipline matters more than the tool.

What to do

• Build a current inventory of every AI system you run or ship into the EU, with its risk tier and the obligations attached. If you cannot produce this in a day, that is your first gap.
• Check your synthetic content marking against the 2 December 2026 retrofit deadline. If you have generative features live in Europe, confirm the marking is in place or on a dated plan.
• Set a buffer for the 2027 and 2028 high-risk deadlines. Work backwards from the date and give yourself at least a quarter of slack.
• Practice producing a full evidence pack for one system. Time it. If it takes more than a few days, fix the storage and ownership before the next deadline, not after.
• Brief your board on the fixed-cost problem in plain terms: the risk is not the fine, it is the scramble. Budget for the register, not just the filing.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.