EU AI procurement clauses: a governance schedule, not an industrial buying agent

External-source research synthesis, not an OfferSignal client result. Consolidates retained I39 and I51 research from 23 and 27 October 2026; source-specific dates and limitations remain below. This is not legal advice.

Buying an AI service is not the same as buying conventional software. A buyer needs to establish intended use, data permissions, measurable performance, human oversight and audit rights—not merely collect assurances. The European Commission-supported Model Contractual Clauses for AI (MCC-AI), published through the Public Buyers Community, offer a reusable contractual framework. They are not a supplier-search engine, quoting system or compliance certificate. The community does not itself buy, deploy or operate AI, and the retained evidence establishes no industrial customer result. 1

What the package provides

The updated package was published on 5 March 2025: full clauses for high-risk AI, a customisable light version for non-high-risk AI, and commentary explaining adaptation. A 16 June 2025 notice announced availability in 24 EU languages, also reported in a Commission newsletter. A related community resource was dated 23 January 2026; that resource date should not be read as a new clause revision. These are public working materials, not commercial software or an official Commission position. The adopting public organisation remains responsible; private organisations should not use them wholesale, but may adapt provisions after their own legal and operational assessment. 2 3 4 5 6 8

The workflow is contractual: describe the system and intended purpose, classify its use, identify public-organisation, supplier and third-party datasets, then specify documentation, evaluation and controls. The high-risk template addresses risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy, robustness, safety and cybersecurity. It also covers incidents, change control, corrective action, data handover and audit. 6 7

Annexes must make these requirements operational: metrics, test population, tolerances, abstention rules, drift thresholds, incident severity, response times, rollback and acceptance criteria. The template creates a place for those decisions; it does not supply the baselines. Delivery can be made conditional on testing, with versions monitored afterwards. The full clauses propose an audit once per calendar year or after a substantial modification, allowing an independent auditor. Exercise of that right depends on the signed agreement's scope, costs, confidentiality and remedies; the stated frequency is not evidence that an audit occurred. 7

An AI schedule still needs a complete contract

The commentary treats the clauses as AI-specific provisions, not a complete services agreement. Acceptance, payment, delivery deadlines, applicable law and liability remain outside them; independent legal analysis and a separate procurement guide provide additional context. There is no ready-made warranty for dimensional fitness, manufacturing capacity, lead time, quote validity or certification. Price, currency, tax, freight, delivery terms, Incoterms, customs, export controls and consequential loss still require appropriate commercial provisions. 6 9 10

The framework can support supplier pre-vetting or review of an AI service that searches project records. It does not find suppliers, issue an RFQ, calculate price, approve an offer, place or accept an order, manage logistics, invoice or inspect quality. Contractual disclosure cannot establish that equipment exists or that a proposed part is manufacturable.

Data rights do not confer decision authority

Annex B and Articles 15–17 offer tools for reducing supplier dependency. Public-organisation datasets accrue to the organisation or its designee by default; supplier use is limited to agreement performance unless the annex permits otherwise. Destruction may be requested with feasible evidence, and handover should use a common file format. Supplier and third-party datasets remain theirs by default; onward-development rights and handover must be selected. For general-purpose systems, training-data rights may not be obtainable, leaving the buyer with descriptions rather than the underlying data. 6 7

The package does not automatically transfer model weights, source code, supplier algorithms, manufacturing drawings or enterprise records. For proprietary drawings, computer-aided designs and customer records, the signed agreement still needs explicit rules for provenance, revision, access, residency, subcontractors, model training and deletion evidence. A data-rights provision is not proof that records are complete, current or accurate. 7

Human oversight means assigned, authorised people can understand limitations, monitor operation, detect anomalies, override or reverse outputs and stop the system safely. Their presence does not transfer commercial authority to AI, nor does approval automatically establish competence or independence. Named engineering and commercial owners must still approve drawing revisions, materials, feasibility, quotations, supplier appointments and quality acceptance. 7

An illustrative manufacturer adaptation

Consider a manufacturer buying an assistant to retrieve archived drawings, triage custom-work RFQs and prepare draft enquiries. This is illustrative, not a reported deployment. Selected clauses could form an attached schedule defining intended purpose, permitted sources, revision boundaries, system architecture and version, data provenance and preparation, validation metrics, limitations, logs, change notices, oversight, cybersecurity and an exit format. Outputs should remain drafts pending accountable review.

Industrial requirements could address materials, tolerances, compatibility, equipment constraints, current capacity, lead-time assumptions and certification evidence. The authoritative engineer would confirm revision, material, tolerances and manufacturability; the authorised commercial owner would release any binding quote. Enterprise, manufacturing-execution and product-lifecycle records should remain authoritative.

Acceptance tests should include stale or permission-incomplete drawings, near-duplicate revisions, mixed units, missing tolerances, incompatible grades or variants, outdated capacity, absent certificates, missing freight or tax assumptions and unsupported delivery promises. Suggested measures include authoritative-record retrieval, wrong-revision rate, unsupported-claim rate, review time, correction effort and rework. These are proposed evaluation measures: the retained research supplies no baseline, threshold or achieved improvement. A cleaner contract does not guarantee a correct recommendation.

Legal status, dates and the limits of the evidence

Regulation (EU) 2024/1689 is binding law; these clauses are voluntary drafting aids. Depending on the system and use case, legal duties concerning risk, data, documentation, human oversight, accuracy, robustness, cybersecurity, deployer monitoring and record-keeping still need assessment. The clauses do not replace the AI Act, the General Data Protection Regulation, public-procurement rules, worker obligations, sector safety law or national remedies. They provide neither certification nor legal advice. 11 12

The retained I51 research records the Commission FAQ as reporting extensions to 2 December 2027 for standalone high-risk systems and 2 August 2028 for AI embedded in regulated products. Those date-sensitive claims were not independently verified in the duplicate audit or this consolidation. The retained legal link is a consolidated text dated 27 July 2026; current obligations, implementation dates, warranties and conformity triggers must be checked against applicable law before signature, rather than inferred from the March 2025 template. 12 13

Two critiques remain distinct. A March 2025 independent critique described limited changes and under-specified accuracy and cybersecurity annexes. An October 2026 practitioner review reported no 2026 revision and questioned fit with later AI Act changes. Both are secondary analysis, not official Commission findings or independent legal verification. 14 15

The demonstrated outputs are public availability of three English artefacts, announced language access and the template's proposed audit right—not adoption or production performance. Through 27 October 2026, the retained research identified no public adoption denominator, named industrial implementation, awarded tender, completed audit, production incident, termination or dispute-resolution result. It found no quantified savings, accuracy lift, cycle-time or lead-time improvement, defect or incident reduction, quote-conversion improvement, quality or safety gain, or revenue outcome attributable to the updated MCC-AI. This is a bounded absence of evidence, not proof that none exists. The useful lesson is disciplined preparation and negotiation, with commercial authority and manufacturing validation still separate.

References