AI Supplier Changes Its Model – Notice, Risk Review and Approval
Practical guide for businesses when an AI supplier replaces a model: change notice, contract review, GDPR, AI Act roles, testing, approval and reversal.
Your AI supplier plans to replace the language model or significantly update the model behind a service you already use. The change can affect accuracy, automated decisions, data handling, cost and reliability. A 'better model' is not automatically safe, but not every update requires new CE marking or a fixed statutory notice period.
Practical answer: secure the release notice, identify what has actually changed, compare the contract and privacy commitments, test against predetermined acceptance criteria and record approval, conditional approval or refusal. Separately consider whether the change may qualify as a substantial modification to a high-risk AI system. GO/NO-GO is a recommended internal method, not a universal EU legal process.
1. What exactly is the supplier changing?
Request the old and new model identifiers, intended purpose, release notes, functions, limits, data locations and planned activation date. Distinguish changing the foundation model from adding external tool access, changing filters or adjusting the interface. Identify which business processes rely on this service.
Ask for relevant test evidence, version traceability, incident procedures, rollback options and any new subcontractors. The precise evidence should fit your risk and contract; no universal document format is imposed on every AI SaaS update.
2. Does the supplier have to provide advance notice?
There is no universal EU rule requiring 30 days' notice for every AI model upgrade. Review your change-control clause, SLA, right to keep the former version, testing window, termination and objection rights. Contract wording can create stricter duties, and a material alteration to service performance may raise questions under the particular agreement.
Distinguish GDPR Article 28(2): where a processor adds or replaces a subprocessor, specific or general prior written authorisation is required. With general authorisation the controller must be informed of intended changes and given an opportunity to object. This concerns subprocessors, not every internal update to model weights.
3. What is a substantial modification under the AI Act?
Article 3(23) of the AI Act defines a substantial modification as a post-market or post-deployment change not foreseen in the initial conformity assessment that affects compliance with specified high-risk requirements or alters the assessed intended purpose. For high-risk AI, Article 25 may treat certain downstream actors as providers when they substantially modify a system or change the intended purpose so that a previously non-high-risk system becomes high-risk.
Assess classification and timing first. Review Article 6 and the annexes, and identify the actual provider/deployer/integrator. Following Regulation 2026/1744, the relevant high-risk requirements apply from 2 December 2027 for Article 6(2)/Annex III and 2 August 2028 for Article 6(1)/Annex I, with transitional and sector-specific conditions. Other contractual and GDPR duties may already apply.
4. Personal data and a changed DPIA
If the AI service processes personal data, check whether the replacement causes new data storage, access, training on customer inputs, purposes, transfers or effects on individuals. Review lawful basis, transparency, processor terms, subprocessor list and access controls. A new model does not automatically require a new DPIA, but under GDPR Article 35 and Swedish IMY guidance a controller must examine whether the changed processing is likely to pose a high risk, and may need to revisit an existing assessment.
5. Risk matrix before approval
| Check | Request evidence | Decision |
|---|---|---|
| Version and scope | Model ID, release notes, changed intended use | Map affected workflows and responsibilities |
| Accuracy and bias | Company-specific regression cases and failure patterns | Set acceptance limits before testing |
| Personal information | Data flows, location, training and subprocessors | Review GDPR, contract and DPIA need |
| Security and access | New connectors, permissions, audit logs | Check cybersecurity and access impact |
| AI Act role and risk | Purpose, prior classification, change details | Review Articles 6 and 25 and applicable dates |
| Contract and operations | Notice clause, SLA, fees, rights to terminate | Check notice and contractual remedies |
| Rollback and approval | Version freeze, export, test report and owner | GO / conditional GO / NO-GO |
6. Eight steps from notice to decision
- Record the change. Keep supplier notice, receipt date, old/new versions and intended go-live date.
- Map dependencies. List workflows, affected people, data, integrations and important decisions.
- Review the agreement. Check notice, SLA, training limits, subprocessors, exit and version-freeze rights.
- Determine legal scope. Classify the AI use case, high-risk duties, provider status and GDPR risk changes.
- Approve tests in advance. Select independent cases, known correct answers, sensitive scenarios, cost and security criteria.
- Compare versions. Use parallel or restricted tests; document regressions, new errors and uncertain answers.
- Record and communicate the decision. Sign GO, conditional GO or NO-GO, conditions, responsible person and any contractual objection.
- Monitor and revert if needed. Track incidents and drift, preserve logs and exercise rollback or contractual exit when necessary.
7. Approve, approve conditionally or reject
GO: predetermined tests and contractual checks pass and an authorised person signs the release. Conditional GO: restrict the use case and introduce verifiable extra controls, such as mandatory human review and a deadline for follow-up. NO-GO: essential data flow or rights are unclear, critical outputs regress, mandatory safety compliance remains unresolved or unacceptable risks cannot be reversed. An internal 'GO' never overrides a statutory obligation.
8. Two real-world scenarios
Customer-facing chatbot changes and starts making overconfident errors
The vendor introduces a faster model that confidently misstates prices and contractual terms. Test anonymised real questions against confirmed answers, add fact checks and withhold approval if critical error rates increase. A better benchmark score does not establish fitness for your use case.
Model update introduces a new subprocessor and storage location
The supplier moves processing to a new infrastructure partner. GDPR Article 28(2) authorisation and notice may then apply, as may transfer, security and DPIA considerations. That legal distinction is different from claiming that every model-weight update itself requires GDPR advance notice.
9. Frequently asked questions
Is 30 days' notice mandatory for every AI model update?
No. The actual service contract and applicable special rules determine what is required. GDPR Article 28(2) concerns changes in subprocessors under its stated conditions.
Do we become an AI provider by changing a model?
Not automatically. AI Act Article 25 addresses defined cases involving high-risk systems, substantial modifications or changes to intended purpose. Check classification, role and start dates.
Does every update require a new DPIA?
No. Review whether the changed processing is likely to create a high risk and whether the previous assessment needs revision.
Can we reject the update and stay on the old version?
That commonly depends on contract wording and technical availability. Negotiate notices, testing windows, ability to pin versions, export and exit before depending on the service.
10. Legal sources and related guidance
- EU AI Act 2024/1689 – consolidated regulation
- Regulation EU 2026/1744 – updated high-risk dates
- European Commission – AI Act Article 25
- GDPR EU 2016/679 – Articles 28 and 35
- Swedish IMY – when a DPIA is required
- European Commission – AI Act definitions Article 3
- Mallbutiken – assessing an AI supplier before purchase
- Mallbutiken – when is AI classified as high-risk?
For an industrial example, see our guide to AI vision systems in production cells, which covers test batches, failure cases and technical release decisions after a model change. This complements supplier change control but does not replace machinery safety assessment.
Legal-source check: 19 November 2026. Contractual notice and GO/NO-GO are practical recommendations, not universal legal requirements. High-risk AI and GDPR responsibilities require case-by-case assessment. No template purchase is required.
Explore the topic
Business guides – contracts, HR, AI & complianceAI Supplier Assessment – Checklist before the company approves an AI service
How companies vet an AI supplier before approval: use cases, AI Act role, data, security, subcontractors, changes, agreements, and exit.
Business guides – contracts, HR, AI & complianceAI Act risk classification – when is an AI system high-risk?
Practical guide to AI Act risk classification: prohibited practices, high-risk according to Article 6 and Annex I/III, transparency according to Ar...
Business guides – contracts, HR, AI & complianceAI Vision in a Production Cell – Model Change, Validation and Release
Practical guide to updating AI-based visual inspection in manufacturing: distinguish quality control from safety functions, test defects and challe...


