AI vendor terms for high-risk use: what do Anthropic, OpenAI, Google and Microsoft demand?

On 8 October 2026 Anthropic published an updated version of its Usage Policy, which applies from 12 November 2026. The policy requires a qualified person to review recommendations that the model makes about individuals in areas such as employment, credit, insurance, housing and healthcare, and requires the affected person to be told that AI was used. The requirements resemble the obligations of deployers of high-risk systems under the AI Act and the rules on automated decision-making in the General Data Protection Regulation. The AI Act has, however, not been incorporated into the EEA Agreement, and its high-risk obligations will not apply in the EU before 2 December 2027. For Norwegian businesses that use AI services in HR, credit assessment or customer handling, vendor terms are therefore the first binding source of such requirements. This article examines what the terms require, how they relate to legislation, and what contractual consequences a breach or a unilateral change of terms may have.

What changes in Anthropic’s Usage Policy from 12 November 2026?

The most significant change for businesses is that the high-risk requirements are now tied to the purpose for which the AI recommendation is used, and no longer to whether the service is directed at consumers. The previous version, which applied from 15 September 2025, described the requirements as measures for consumer-facing use cases with elevated risk. The new Usage Policy applies to anyone who uses the services to provide a high-risk recommendation, whether the recipient is a consumer, an employee or a business. Anthropic states that the requirements are unchanged in substance and that the section has been rewritten to identify more clearly which recommendations are covered. In our assessment, the new wording nevertheless means that internal HR tools and credit modules in B2B services fall more clearly within scope than before.

The update also contains other changes. The rules on fake accounts and influence operations have been brought together in a new part on deceptive campaigns, which applies to both political and commercial activity. Tracking individuals without consent is prohibited, both in real time and through analysis of data already collected. Where the model controls equipment capable of physical actions, a qualified operator must be able to observe and stop the equipment, and the equipment must be able to enter a safe state if the connection to the model is lost. According to the announcement, use is also prohibited for companies incorporated or headquartered in a region that Anthropic does not support, and for companies majority-owned or controlled from such regions.

What do AI vendor terms require for high-risk use?

Anthropic’s Usage Policy imposes two cumulative requirements where the model provides a high-risk recommendation. The first is that a qualified person must review the recommendation before it is given to the affected person or used to make a decision about that person. The policy expresses the requirement as a duty to review in the following terms.

“meaningfully review a High-risk AI Recommendation, with the authority to change or modify it”

Under the policy, a qualified person is someone with training or experience that enables them to assess the output in the relevant field, and with authorisation where the law requires it. That person retains responsibility for the accuracy and appropriateness of what is delivered or decided. The second requirement is that the person receiving the advice or subject to the decision must be clearly informed that AI was used. There is no need to name the vendor or the model.

The policy lists eleven high-risk areas, including legal matters, medicine, finance, credit, insurance, housing, employment, education and credentials, access to healthcare and public benefits. The employment area covers, among other things, screening and ranking of candidates, hiring and promotion, setting working hours and pay, and assessing an employee’s performance or conduct. The credit area covers granting or refusing credit and setting the amount, interest rate or limit of loans and credit, as well as credit scores used for such decisions. Exclusions include general information, internal drafting, research, summarisation and analysis that is not the final recommendation, and implementation of a decision already made. Where the law permits, no human review is required before the business acts on a decision that is entirely favourable to the affected person.

Do OpenAI, Google and Microsoft have equivalent AI vendor terms?

The three other major vendors all have rules on human oversight of AI-based decisions that affect individuals, although the wording varies. Anthropic and Microsoft additionally impose an express requirement to disclose the use of AI. The table is based solely on the vendors’ own policy documents as they stood on 16 October 2026.

Google’s policy applies to the Gemini API through the Gemini API Additional Terms of Service, which require the customer to comply with it. Microsoft’s code of conduct applies alongside the product terms and states that it has been designed to align more closely with forthcoming AI regulation, citing the AI Act as an example. OpenAI’s rules also prohibit tailored advice that requires a licence, such as legal or medical advice, without appropriate involvement of a licensed professional. None of the vendors defines in further detail what constitutes sufficient human oversight, beyond Anthropic’s requirement of meaningful review and authority to change the output.

How do the terms relate to AI Act Articles 14 and 26?

The vendor terms largely mirror the obligations of deployers of high-risk systems under the AI Act. Under AI Act Art. 26(2) the deployer must assign human oversight to natural persons who have the necessary competence, training and authority. Under AI Act Art. 14(4)(d) the system must enable the person exercising oversight to disregard, override or reverse the output. Under AI Act Art. 26(11) deployers of high-risk systems listed in Annex III that make or assist in making decisions about natural persons must inform those persons that they are subject to the use of the system. Point 4 of Annex III covers systems for recruitment, selection and decisions in employment relationships, and points 5(b) and (c) cover creditworthiness assessment and risk assessment and pricing in life and health insurance.

The difference lies in scope and timing. The obligations under the AI Act apply only to systems classified as high-risk, whereas Anthropic’s list of areas is broader and also includes, for example, housing and legal advice. Following the amendment of AI Act Art. 113 by Regulation (EU) 2026/1744, the obligations for high-risk systems under Annex III apply in the EU from 2 December 2027, and for systems under Annex I from 2 August 2028. The omnibus did not amend Article 14 or Article 26. In Norway the regulation does not yet apply. The Government stated on 4 August 2026 that the draft AI Act would be sent for a new public consultation in the autumn of 2026, with a view to a bill in spring 2027. According to the Government, negotiations on EEA adaptations are ongoing. The vendor terms thus introduce similar obligations in contractual form several years before Norwegian authorities can enforce them.

How do the terms relate to GDPR Article 22?

The General Data Protection Regulation already applies in full and currently imposes requirements that overlap with the vendor terms. Under GDPR Art. 22(1) the data subject has the right not to be subject to a decision based solely on automated processing which produces legal effects concerning them or similarly significantly affects them. Where an exception applies, GDPR Art. 22(3) requires, among other things, a right to obtain human intervention and a right to contest the decision. In C-634/21 SCHUFA, the Court of Justice of the EU held that a credit score may constitute an automated decision where the lender draws strongly on it. The judgment is discussed further in the article on automated decisions and the Uber fine.

The vendor terms go further than Article 22 in two respects. The human review requirement in Anthropic’s policy also applies where a human formally makes the decision, provided the AI recommendation forms its basis, and it is not limited to decisions with legal or similarly significant effects. The information duty under GDPR Art. 13(2)(f) and GDPR Art. 14(2)(g) applies only to automated decisions covered by Article 22, whereas Anthropic’s requirement applies to every high-risk recommendation. A business that complies with the vendor terms will nevertheless often have laid a sound basis for compliance with Article 22, because a genuine human assessment will normally take the decision outside the prohibition.

What happens in the event of a breach of an AI vendor’s usage rules?

A breach of the usage rules is a breach of contract. Under Section D.2 of Anthropic’s Commercial Terms of Service, the Usage Policy is incorporated into the agreement by reference, and the customer and its users may use the services only in accordance with it. Under Section I.3, Anthropic may suspend all or part of the service if it reasonably believes that the use breaches the rules, and the vendor is then not liable for loss caused by the suspension. In the event of a material breach, the agreement may be terminated on 30 days’ written notice if the breach is not remedied within that period (Section I.2.b). Under Section K.2 the customer must also indemnify Anthropic against third-party claims arising from use in breach of the Usage Policy.

The practical risk is business interruption. A business that has built case handling, customer service or recruitment on a single model may lose access without notice and without any claim for compensation against the vendor. For customers in the EEA, the agreement is concluded with Anthropic Ireland Limited and is governed by Irish law, with the Irish courts having jurisdiction (Section M.7). Norwegian default rules on breach of contract will therefore not automatically govern the relationship with the vendor. Liability towards employees, customers and applicants, by contrast, follows from Norwegian law and the General Data Protection Regulation, and the business cannot contract out of it with the vendor. The allocation of liability when an AI agent causes harm is examined in the article on AI agent liability.

Can the AI vendor change the terms unilaterally?

The vendor may change the terms unilaterally within the limits set by the agreement. Under Section M.3 of Anthropic’s commercial terms, the vendor may change the terms at any time, with effect 30 days after the change is published or notified. Changes required by new legislation take effect immediately, and changes do not apply retroactively. The Usage Policy is incorporated by reference, and the update of 8 October taking effect on 12 November falls within that notice period. Microsoft expressly reserves the right to update its code of conduct from time to time and places responsibility for complying with the current version on the customer, while OpenAI states that its rules are updated on an ongoing basis.

Unilateral changes to usage rules raise the same questions as unilateral price increases. The scope of the right to make changes depends on the wording of the agreement, and the customer’s most important protection is a genuine right to terminate and time to adapt. This is discussed further in the article on SaaS price increases. Businesses that have negotiated their own framework agreements with an AI vendor should check whether the agreement allows the vendor to change the usage rules without consent, and whether notice and a right to terminate apply to changes that require processes to be redesigned.

What should businesses do before 12 November 2026?

The steps below apply to businesses that use AI models directly or through system vendors in HR, credit assessment, insurance or customer handling. They are also relevant to the use of OpenAI, Google and Microsoft, because those vendors already impose corresponding requirements in their current terms.

1. Map the processes in which an AI model makes recommendations about individuals, and identify which vendor and which usage rules apply to each of them.
2. Assess for each process whether it falls within a high-risk area under the vendor’s rules, and whether it is also an automated decision under GDPR Art. 22.
3. Designate the persons who will review the recommendations, and ensure that they have the professional competence, the time and the formal authority to change or reject the output.
4. Document the review, so that the business can demonstrate to the vendor and to the Norwegian Data Protection Authority (Datatilsynet) that the control is genuine and not a formality.
5. Include information about the use of AI in rejection letters, replies to applicants and customer communications, and align the wording with the privacy notice.
6. Update the internal AI policy for employees, so that those who use the tools know the requirements for review and disclosure.
7. Ask system vendors that build on an underlying model to confirm which usage rules have been passed on in the agreement, and how the business will be notified of changes.
8. Review the agreements with AI vendors and consider negotiating notice before suspension, a right to remedy and a right to terminate if the usage rules change.

The work should be organised so that it can later be reused when the AI Act’s obligations for deployers become applicable in Norway. Further material on the regulatory framework is available on the topic page on artificial intelligence. Businesses that have not introduced human review and disclosure of AI use by 12 November 2026 risk Anthropic restricting or suspending their access under Section I.3 of the commercial terms.

Questions and answers

Next legal review: 12 November 2026