A general audit right in AI contracts gives no certainty: specify the inspection evidence concretely

An audit clause that only promises access to general documentation does not guarantee that you can verify a concrete AI outcome. So specify per contract which inspection evidence the supplier provides, within which deadline, with which inspection powers and which remedies apply on refusal or incompleteness.

The European Commission has announced that its Expert Group on AI Contracting will start work in July 2026 and help develop horizontal model contract terms for AI contracts. The aim is to reduce practical risks in AI contracts. That makes the contractual design of audit and verification rights a concrete design question now, and no longer a theoretical wish.

What changes now that the European Commission is preparing model contract terms for AI contracts?

The Commission positions AI contracts as a separate subject for standardisation. The announced model terms and user-oriented guidance are intended to give purchasers a foothold for arrangements on access, information and assistance. That is a policy prompt, not a ready-made audit clause: the Commission provides building blocks, not a universal clause that every purchaser can adopt blindly.

At the same time, the AI Act already lays a foundation. Under Article 25 of the AI Act, the provider of a high-risk AI system and the third party supplying it with a model, tool, service, component or process must, by written agreement, specify the necessary information, capabilities, technical access and assistance, in so far as this is needed to comply with the regulation. The guidance on Article 25 on responsibilities along the AI value chain underlines that this is an obligation to specify access, not an unlimited standalone right to inspect everything.

Why is a general right to information not enough for an enforceable audit right?

Because a general compliance statement does not prove that you can reconstruct a specific AI event afterwards. A peer-reviewed study in Frontiers in Artificial Intelligence on the contractual vacuum in AI procurement describes how AI-as-a-Service can deprive a purchaser of the telemetry needed to explain and correct automated actions.

The study therefore discusses transitive audit rights: when several suppliers contribute to one output, the audit right must run through to the underlying supplier that carried out the disputed step. In our assessment this is a significant blind spot in audit arrangements: where an audit right stops at the direct supplier, the records held by its sub-processors may remain out of reach, even though that is precisely where the evidence can lie. In its guidance on the AI Act, the Commission points out that high-risk systems must be traceable and auditable and that appropriate documentation must be maintained. In our analysis, this means that a contract should specify the relevant records, versions and retention periods so that the purchaser can test traceability in practice. In our analysis, a contract that does not enforce this makes that traceability depend on the supplier's goodwill.

Which concrete inspection evidence must an AI supplier be able to provide?

An enforceable clause names not just an "audit right" but the records and access the supplier must actually produce. A practical guide from the Global AI Certification Council on an ISO 42001 audit checklist makes concrete which evidence an auditor links to criteria, sampling and period, and which supplier artefacts are relevant. So build out your clause along five elements:

• Object and scope: the model, the application, retrieval, tools, datasets, sub-processors and relevant locations that fall under the audit.
• Inspection evidence: risk assessments, validation and bias tests, model and dataset versions, configuration and change logs, incident records, access and tool-usage records, human interventions and the decision state.
• Procedure: audit trigger, notification, deadlines, an independent auditor, interviews with staff, remote and on-site inspection, re-audit after a substantial change and access in the event of incidents or disputed outcomes.
• Enforceability: cure period, suspension, step-in, cost recovery, contractual sanctions and termination on refusal or non-reconstructability.
• Confidentiality: controlled access, non-disclosure, secure rooms and limited disclosure rather than a general exception for trade secrets that fully rules out verification.

Also test the completeness of the evidence, through sampling, period, version history and traceable records, and set out how missing evidence is recorded. That distinction between a contractual right to inspect and the concrete records that must be available determines whether you can turn a technical AI explanation into usable decision evidence.

What does this mean for directors, lawyers and CISOs who sign AI contracts?

Our analysis: because a general information request does not guarantee that a specific outcome is reconstructable, the purchaser runs the risk of being unable to convincingly explain or check an AI decision afterwards; our recommendation is therefore that the lawyer now include an inspection-evidence list in every new contract rather than an open audit sentence. Because Article 25 of the AI Act requires written specification of information, technical access and assistance for the value chain in question, but does not in itself create an unlimited audit right, in our assessment it remains for the parties to set out the concrete powers themselves; our recommendation is therefore that the CISO determine before signing which logs, model versions and tool-usage records are retrievable and with what retention period. Because, with AI-as-a-Service, several suppliers contribute to one output, in our analysis the evidence may become inaccessible as soon as the audit right stops at the direct supplier; our recommendation is therefore that the contract manager negotiate a transitive audit right covering relevant sub-processors and examine AI components in the supply chain. Because suppliers can fend off verification by invoking trade secrets, in our assessment an audit right without a remedy is just paper; our recommendation is therefore that the director have suspension, step-in and termination on non-reconstructability included as hard sanctions.

Together, these developments do not yet produce one universal audit clause. Model terms, the AI Act and the academic framework must be aligned with the risk, the sector, the applicable legislation and the actual system architecture. Anyone wishing to understand how these requirements fit into a broader accountability framework will find more context in our topic hub on AI governance and accountability frameworks and in the analysis of when audit trail software suggests control without proving it.

Sources: The article draws on the European Commission (model contract terms and AI Act guidance), the AI Act Service Desk on Article 25, a study in Frontiers in Artificial Intelligence and an audit checklist from the Global AI Certification Council.