AI Act checklist: the internal audit in 10 verifications

On Tuesday we lined up what kicks in on 2 August and what slips; today we keep the promise of the operational piece. What follows is the checklist we use for the rapid internal audit, ten verifications executable in one week by a small group, compliance, IT and one process owner, with a rule that applies to all of them: the answer is never "yes", it is a document. Each verification names the evidence that closes it, because an audit without evidence is a survey, and surveys defend nobody on 2 August.

The foundations: knowing what you have

Verification 1, the inventory. Does a register of AI systems exist, complete with those embedded in supplier products and those that entered from below without approval? Evidence: the register with its last-update date and owner. The fastest completeness test remains the comparison with reality: a tour with IT through network logs and expense reports almost always finds systems the register ignores.

Verification 2, the classification. Does every system in the register carry a risk classification with a written rationale, not just a label? Evidence: the classification sheet per system. The rationale is what turns an opinion into a defensible position, and it is also what tells you which systems the postponement to 2027 applies to.

Verification 3, the prohibited practices. Has somebody formally verified that no system falls into the prohibited practices, from social scoring to emotion recognition at work? Evidence: the minutes of the verification, with scope and date. It is the one point of the list where the sanction ceiling is the highest and tolerance is zero.

The obligations kicking in now

Verification 4, transparency towards people. Do the systems interacting with customers or employees disclose themselves as AI? Evidence: interface screenshots, collected system by system. Verification 5, generated content. Is the AI-produced or AI-manipulated content you publish recognisable as such, with a plan for the machine-readable marking enforceable from December? Evidence: the content policy and marked examples. Verification 6, literacy. Does the Article 4 obligation have evidence: who was trained, on what, when, with what coverage of the people operating the systems? Evidence: the training plan and attendance records. These are the three verifications most visible from outside, and therefore the first ones an authority or a complainant can test without even entering the company.

The supply chain and endurance over time

Verification 7, the roles. For every system, is it written who answers for it, and does the chain reach a single point accountable to the top? Evidence: the responsibility matrix. Verification 8, the suppliers. For purchased systems, do you have the provider's documentation, its classification, the clauses you need for your own deployer obligations? Evidence: the supplier file for the main systems; the questions to start from are the ones we published on vendor evaluation. Verification 9, the data. Do the AI uses of personal data have a documented legal basis and, where needed, a DPIA? Evidence: the processing register updated with AI uses, because GDPR never waited for the AI Act and the 15 million euro fine to ChatGPT's provider (Italian DPA, December 2024) reminded everyone. Verification 10, the logs and the plan. Do the relevant systems keep traces of their operations, and for the Annex III ones does a written roadmap to December 2027 exist? Evidence: the logging policy and the plan with dates and owners.

How to read the score

Ten verifications, ten pieces of evidence: the count of the ones you hold is your photograph, and the reading scale we use is blunt. Below four, the problem is not the AI Act, it is that the organisation does not know what it owns, and you restart from the inventory. Between four and seven, governance exists in patches, and the weeks until 2 August suffice to close verifications 4, 5 and 6, the publicly exposed ones. From eight up, you are in the equipped minority, and the real work is verification 10's plan, because your risk is not the August deadline, it is the 2027 queue.

Honesty requires saying what this checklist is not: a certification of conformity, which no ten-point list can grant, and which we ourselves never promise. It is the instrument for knowing where you stand, which is the condition for deciding where to invest, and it is the same logic with which our AI Rating extends these ten verifications into a full gap analysis on the Risk dimension, with roadmap and priorities: calendly.com/fabiolalli/zerofive, or hello@zerofive.ai.

The week to execute it, between now and 2 August, is still there. How many of the ten pieces of evidence could you put on the table this morning?