On August 2, 2026 the EU AI Act reaches its general applicability date, and the first thing nearly every obligation assumes is something most companies do not have: a maintained, provable inventory of the AI systems in use, each with an accountable owner, an approved purpose, and a record of what data it touches.
What actually happens on August 2
The EU AI Act entered into force on August 1, 2024 and has been phasing in ever since. Prohibited practices became enforceable in February 2025. Obligations for general-purpose model providers followed in August 2025. August 2, 2026 is the Act's general applicability date, and it activates the Commission's enforcement powers. What it does not bring is the part most coverage still gets wrong.
The high-risk obligations that companies spent two years preparing for are no longer what arrives. The Digital Omnibus, adopted by the Parliament and Council in June 2026 and in force since July, moved them: standalone high-risk systems under Annex III now come due December 2, 2027, and AI embedded in regulated products under Annex I comes due August 2, 2028. What actually lands on August 2 is the Article 50 transparency regime. Tell people when they are interacting with an AI system. Mark synthetic content in machine-readable form, with a grace period to December 2, 2026 for systems already on the market. Disclose deepfakes. The penalty math is also more specific than most posts suggest: the seven percent top tier attaches to prohibited practices, which have been banned since February 2025, while the transparency duties arriving now sit in the three percent tier.
So a vendor telling you that your high-risk compliance is due Friday is working from a calendar that changed in June, and that is worth remembering when the same vendor offers to fix the problem. The companies genuinely exposed to the high-risk requirements now have until late 2027, and the ones that treat that as extra time to build a real inventory will be in better shape than the ones that treat it as permission to stop.
(Correction, July 31, 2026: an earlier version of this article said the high-risk requirements became applicable on August 2, 2026 and cited the top penalty tier alongside them. The Digital Omnibus moved the high-risk dates, and the seven percent tier attaches to prohibited practices. The text above is corrected.)
Who is actually in scope
The Act applies to providers and deployers that place AI systems on the EU market or whose system output is used in the EU. A US mid-market firm with no EU operations and no EU customers may have limited direct exposure today. That is the honest reading, and it is the one your counsel should confirm rather than a software company.
Three things still bring the Act to your door earlier than a strict scope analysis suggests. The output clause has extraterritorial reach, and modern SaaS makes it easy to serve EU users without deciding to. Enterprise customers increasingly push AI Act style questions down through vendor questionnaires regardless of where either party operates. And US frameworks are converging on the same expectations, from state AI laws to NIST AI RMF to ISO 42001. The GDPR pattern is repeating: the framework becomes the market's default expectation well before the law technically requires it of you.
Every framework asks the same first question
Strip away the differences between the EU AI Act, NIST AI RMF, ISO 42001, the state AI statutes, and an institutional investor's due diligence questionnaire, and they open with the same request. Show us the AI systems in use. For each one, tell us who is accountable, what business purpose it serves, and what data it can reach.
There is a structural reason this comes first. Risk classification, the mechanism at the center of the EU AI Act, is an operation performed on an inventory. You cannot classify a system you have not identified, and you cannot demonstrate that a system fell outside the high-risk categories without a record showing you evaluated it. The inventory is not one compliance artifact among many. It is the precondition for every judgment the frameworks ask you to make.
What a defensible AI inventory contains
A list of tool names in a spreadsheet is a starting point, not an inventory. A defensible entry answers the questions a reviewer will actually ask. Which system and which provider. Who is the named, accountable owner, a person rather than a department. What business purpose was approved, in language specific enough that an unapproved use is recognizable. Which population uses it, whether a team, a role, or the whole company. What classes of data it is permitted to touch, and which are prohibited. Which internal policies apply to it. Where it runs, including the tenant and deployment facts that determine your obligations. And when it will next be reviewed.
The review date is the field that separates an inventory from a snapshot. A snapshot describes a moment and starts aging immediately. An inventory entry with an owner and a review date is a commitment: someone is accountable for this system, and there is a date by which the record must be reaffirmed or corrected. Reviewers notice the difference, because one is evidence of a process and the other is evidence that someone once filled in a spreadsheet.
Declared inventories drift. Discovery is the correction.
Every declared inventory is wrong within a quarter. A department adopts a tool without asking. A trial that one person expensed becomes something a whole team relies on. An engineer wires a model API into an internal workflow. An AI feature switches on inside software you already licensed for something else. None of these arrive through the intake process your inventory depends on.
A maintained inventory therefore needs a second input: discovery from signals the company already owns, including identity provider logs, network data, expense records, and provider billing. Reconciling what was declared against what those signals show is the corrective loop, and the gap between the two lists is itself a finding. Companies running this comparison for the first time routinely discover their actual AI footprint is a multiple of the sanctioned one.
Honesty about method matters here as much as the result. A tool observed directly is a different quality of fact than one inferred from an expense line, and both differ from a corner of the company where you have no signal at all. An inventory that labels its evidence, including what it cannot see, will survive scrutiny that a confident but unlabeled list will not.
An inventory is evidence only if you can prove it
When a regulator, auditor, or enterprise customer asks for your AI inventory, they are not asking whether a document exists. They are asking whether the record is dated, owned, reviewed, and current, and whether you can show the history behind it. A spreadsheet assembled in a panic the week of a deadline is precisely the artifact that fails that test, because it proves the process did not exist until the request arrived.
This is the part of the problem Tallin was built for. Every sanctioned provider goes on the record with an owner, an approved purpose, a population, a policy scope, and a next review date. Discovery reconciles that declared inventory against real usage signals. Every fact carries a label for how it is known: observed, inferred, enforced, or not covered. And the record exports in the forms reviews actually take, from a customer questionnaire to a board report.
To be equally plain about the boundary: none of this makes anyone compliant with the EU AI Act. Classification decisions and legal conclusions belong to your counsel. What Tallin provides is the maintained, provable record those judgments rest on, which is the part no one can produce retroactively.
Key takeaways
- August 2, 2026 is the EU AI Act's general applicability date, bringing the Article 50 transparency duties and Commission enforcement powers. The Digital Omnibus moved high-risk obligations to December 2027 and August 2028.
- Many US mid-market firms have limited direct exposure today, but customer questionnaires and converging US frameworks impose the same expectations sooner than the law does.
- Every major AI framework opens with the same question: which systems are in use, who owns each one, and what data can it reach.
- A defensible inventory entry carries a named owner, an approved purpose, a population, permitted data classes, policy scope, deployment facts, and a next review date.
- Declared inventories drift. Reconciling declarations against discovery signals, with honest labels for how each fact is known, is what keeps the record true.
Keep reading