Direct comparison
OMB M-25-22 vs EU AI Model Contract Clauses
Two governments published reusable AI contract clause libraries. What M-25-22 binds US agencies to, and what the EU's MCC-AI actually says.
Written and maintained by CASRAI Editorial Board
Last updated
Ask CASRAI · free to try
Ask about OMB M-25-22 vs EU AI Model Contract Clauses
Ask your first 2 questions free below. Subscribers get 150 a day for $29 a month.
An AI assistant specialized in research administration. It cites the sources behind every answer, labels web answers and says when it can't answer.
Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.
Works on this site and inside Claude, Cursor and the AI tools you already use.
Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.
How do OMB M-25-22 (United States), EU MCC-AI (European Commission) compare side by side?
The table below compares OMB M-25-22 (United States), EU MCC-AI (European Commission) across 14 procurement-relevant dimensions, from what the document actually is through the most useful thing each one does that the other does not.
Side-by-side comparison
| Dimension | OMB M-25-22 (United States) | EU MCC-AI (European Commission) |
|---|---|---|
| What the document actually is | An OMB memorandum directing agencies on how to buy AI. It states requirements for what contract terms must address; it does not publish model clause text. Agencies write the words themselves. | Drafted contractual clauses, written as a schedule that forms an integral part of a wider agreement. Two documents (MCC-AI-High-Risk and MCC-AI-Light) plus a 15-page Commentary that walks the numbering. |
| Legal force | Binding on US executive agencies. It rescinds and replaces OMB Memorandum M-24-18 outright, so there is no parallel older regime to reconcile. | Voluntary. The clause documents carry an explicit disclaimer: they are “a working document in progress” that “does not reflect an official position of the European Commission,” and actors using them “carry full responsibility for its use in public procurement.” |
| Who is bound | Executive agencies. Explicitly excluded: AI acquired for use as a component of a National Security System, and the Intelligence Community as defined in 50 U.S.C. § 3003. | Nobody, by force of the document. They are offered to public organisations across the EU; the Commentary encourages users to report their use to Public Sector Tech Watch rather than to any regulator. |
| What triggers coverage | Any contract awarded under a solicitation issued 180 days or more after issuance, and any option to renew or extend exercised on an existing contract after that date. Agencies had 270 days to update internal acquisition procedures. Both windows closed in late 2025. | The buyer’s own choice to attach the schedule. Which version applies turns on whether the AI system is high-risk under the AI Act: MCC-AI-High-Risk for Annex-listed high-risk systems, MCC-AI-Light otherwise. |
| What falls outside scope | “Any common commercial product within which artificial intelligence is embedded,” tested two ways: (1) is the product widely available to the public for commercial use rather than specialised or customised for agency use, and (2) does it have substantial non-AI purposes, as opposed to products for which AI is a primary purpose. Word processing bought for its AI features is covered; word processing with AI spellcheck bolted on is not. Also excluded: regulatory actions, enforcement-driven assessments, public-facing test methods, and AI a contractor uses incidentally and was not directed to use. | The clauses cover matters the AI Act covers. They are drafted to sit inside a larger agreement and do not attempt to be one — price, delivery timing, term, governing law and general liability all live in the host contract. |
| Training on the buyer’s data | The strongest single term in the memo. Agency processes must ensure contracts “permanently prohibit the use of non-public inputted agency data and outputted results to further train publicly or commercially available AI algorithms” absent explicit agency consent. Note that it reaches outputs, not only inputs, and that the prohibition is stated as permanent rather than coterminous with the contract. | Article 15.2 of the high-risk version: the Supplier “is not entitled to use Public Organisation Data Sets for any purpose other than the performance of the Agreement, except as otherwise provided in Annex B.” Article 15.3 adds destruction on first request, with evidence of destruction if demanded. Annex B is then a table with a “Rights of use of the Supplier” column and a Yes/No destruction flag — so the default is restrictive and any training permission has to be written in by hand. |
| Vendor lock-in and exit | Addressed repeatedly and by name. Licensing must be scoped to the intended use “to avoid vendor lock-in”; components needed to operate and monitor the system must remain available “for as long as it may be necessary”; solicitations should seek knowledge transfer, data and model portability, clear licensing and pricing transparency; and there is a separate vendor lock-in protection step triggered as soon as a decision not to extend is made. | Handled as data handover rather than as lock-in. On first request the Supplier hands over the most recent Public Organisation Data Sets, and (subject to Annex B) the Supplier and Third-Party Data Sets too, “in a common file format to be designated by the Public Organisation.” An optional clause extends the buyer’s licence to further development of the system by itself or a third party. The phrase “vendor lock-in” does not appear. |
| IP ownership | Agencies must include terms that “clearly delineate the respective ownership and IP rights of the government and the contractor,” must revisit their existing process for data ownership and IP in AI procurements, and should standardise across contracts where possible. The Chief AI Officer Council periodically reviews those processes. | Allocated by data set, not by system. All rights including IP in Public Organisation Data Sets accrue to the buyer; all rights in Supplier and Third-Party Data Sets accrue to the Supplier, with a non-exclusive licence to the buyer sufficient to perform the agreement. Mutual indemnities for third-party IP claims follow. IP in the model itself is left to the host contract. |
| Documentation and transparency deliverables | Stated as a preference, not a specification: agencies are “encouraged, where appropriate, to prioritize obtaining documentation that facilitates transparency and explainability” and an adequate means of tracking performance. For high-impact use cases agencies must tell vendors in the solicitation what transparency and documentation will be required, including enough to complete an AI Impact Assessment. | Itemised. Delivery includes handover of technical documentation and instructions for use, each with its own annex of minimum contents (Annex C and Annex D), updated on every Substantial Modification. Separate articles cover record-keeping and automatic event logs, transparency measures (Annex E), human oversight measures (Annex F) and required accuracy levels (Annex G). |
| Risk-tier logic | Borrowed from a sibling memo. Agencies must make an initial determination of whether a system is likely to host “high-impact AI” use cases as defined by OMB Memorandum M-25-21 — output that is the primary basis for decisions with legal, material or binding effect on rights, benefits, critical services, human life or critical infrastructure. That determination then drives solicitation disclosures. | Inherited from the AI Act. High-risk classification selects the full version; everything else gets the light version. The Commentary is unusually candid that the light version is itself derived from the AI Act’s high-risk obligations, so “only parts of the MCC-AI-Light should be used” for a genuinely low-risk system — it expects you to cut, not just adopt. |
| Audit and inspection rights | Not a clause requirement. The memo’s testing emphasis sits pre-award: agencies must, to the greatest extent practicable, test proposed solutions during evaluation, and should consider standing up a testing environment on government-owned systems. | A dedicated section. The Supplier must cooperate in an audit by or on behalf of the buyer, including interviews with staff and access to Supplier locations; the buyer writes a report, can set a remedy deadline that puts the Supplier in default by operation of law if missed, may use an independent auditor, and may run one audit per calendar year or on any Substantial Modification. |
| Versions and languages available | One memorandum, 13 pages, English. GSA was directed to plan a publicly available acquisition guide within 100 days and to build an Executive-Branch-only repository of AI acquisition information — which “may include language for standard contract clauses” — within 200 days. Both deadlines have passed; CASRAI has not verified what was delivered or whether the clause repository is public. | Three English documents (Model Clauses High Risk, Model Clauses Light Version, Commentary), version-stamped February 2025 and posted on the Public Buyers Community on 5 March 2025, with the set announced as available in all 24 EU languages on 16 June 2025. The clauses ship as editable DOCX, with optional articles marked inline. |
| Reusability by a private buyer | High as logic, low as text. There is nothing to paste — but the five sub-requirements on IP and government data read cleanly as a checklist for a private buyer’s own term sheet, and the covered-AI two-part test is a usable way to decide which purchases deserve AI-specific terms at all. | High as text, with caveats. The DOCX files are editable and openly downloadable, and the drafting is generic enough that “Public Organisation” substitutes for any buyer. The caveats are that the obligations are AI Act obligations, so a non-EU buyer inherits a European compliance frame, and that the disclaimer means nobody stands behind the wording. |
| The most useful thing each one does that the other does not | It makes the anti-training term permanent and extends it to outputs. Most commercial “we won’t train on your data” language is scoped to the term of the agreement and speaks only about inputs. | It gives you a publishable audit. Article 20.4 entitles the buyer to publish the conclusions of its compliance audit report — a disclosure right, in a procurement contract, that the US memo has no counterpart for. |
Common questions
Common questions about OMB M-25-22 (United States) vs EU MCC-AI (European Commission)
Is either of these a standard contract I can just sign?
+
No. M-25-22 contains no clause text at all — it tells agencies what their terms must address and leaves the drafting to them. The EU MCC-AI does contain real clauses, but its own pilot documentation is explicit that they do not comprise a full contract: they attach as a schedule to an agreement that still has to supply price, delivery, term, governing law and general liability.
Which one has the stronger “do not train on our data” term?
+
M-25-22, on two specific dimensions. It requires the prohibition to be permanent rather than coterminous with the contract, and it covers outputted results as well as inputted data. The EU clause is restrictive by default — the Supplier may not use the buyer’s data sets for anything but performance of the agreement — but it points to Annex B, a fill-in table where any wider use is granted by hand, so its strength depends entirely on what the buyer writes there.
Do the M-25-22 deadlines still matter now that they have passed?
+
Yes, but as scope tests rather than countdowns. The 180-day mark is what determines whether a given solicitation or exercised option is covered at all, and the 270-day mark is when agency internal acquisition procedures were supposed to be current. Read against the memo’s 3 April 2025 issuance date, both windows closed in late 2025, so any new federal AI solicitation today falls inside scope unless one of the exclusions applies.
If I am a private company, is there any reason to prefer one library over the other?
+
It depends on which problem you have. If you are negotiating data rights and worried about lock-in, the US memo’s five sub-requirements are the better checklist. If you need to impose concrete engineering and documentation obligations — risk management system, data governance, logging, human oversight, accuracy thresholds, corrective actions — the EU high-risk clauses are already drafted and editable. Many buyers will end up taking the data-rights logic from one and the obligations from the other.
Does the EU set really say the Commission does not stand behind it?
+
It does, in the disclaimer on the first page of both clause documents: the text is “a working document in progress” that “does not reflect an official position of the European Commission,” and “actors that decide to make use of this working document carry full responsibility for its use in public procurement.” It also adds that the clauses are without prejudice to the requirements stemming from the AI Act — signing them does not discharge anyone’s statutory obligations.
Does NIKOLAI have anything that maps onto a procurement clause?
+
One element comes genuinely close: Publication Rights Clause, on NIKOLAI’s N8 transparency-and-review track, defined as contractual or policy terms governing what an external evaluator may publish about a review, including who may redact and whether the developer keeps approval rights over findings. That is the same species of term as MCC-AI article 20.4, which entitles the buying organisation to publish the conclusions of its audit report. NIKOLAI is CASRAI’s own independent dictionary and it is not endorsed by anyone — every row on that element’s crosswalk is a shadow mapping, CASRAI’s own reading of a published document, and neither OMB nor the European Commission has filed a Mapping Declaration about it.
Going deeper
