Vendor AI assurances usually look reassuring on paper: a security questionnaire ticked off, a privacy attestation signed, a compliance one-pager attached to the contract.
Having engaged in many conversations around these assessments, it is still amazing to hear senior people say, “They are doing this already for their customers, and they are more experienced than we are“.
“Why are we making this so hard? the vendor is a large organisation and are using this ai tool for their customers. they have spoken to me and cannot understand why we don’t accept their assurances”
Three 2026 examples show how thin that reassurance can be — and why the organisation, not the vendor, ends up carrying the risk when it breaks.
The 30-Second Take
Paper and verbal assurances from vendors who use AI are not the same as verified control.
Germany’s data protection regulator has fined Vodafone €45 million partly for failing to properly vet and monitor a third-party partner. A 2026 industry report found nearly two-thirds of AI vendors don’t disclose the subprocessors actually touching your data. And a German court has confirmed that when an AI vendor’s tool gets something wrong, the company that deployed it — not the vendor — is the one that answers for it.
If your AI risk process stops at “the vendor said it’s fine,” it stops too early.
Real-World Scenarios: When Vendor Assurances Weren’t Enough
Vodafone Germany — €45 million, and a lesson in outsourced oversight. In 2026 Germany’s Federal Commissioner for Data Protection and Freedom of Information (BfDI) fined Vodafone Germany €45 million across two findings. €15 million was for insufficiently vetting and monitoring third-party partner agencies, whose weak internal controls let staff alter contracts and access customer accounts without proper authorisation — a breach of Article 28 of the GDPR, the provision governing exactly the kind of vendor-processor relationship most organisations rely on for AI tools. The remaining €30 million covered separate authentication failures. Vodafone’s own assurance processes for its partners existed; the regulator found they weren’t rigorous enough to catch what was happening underneath them.
DataGrail’s 2026 Privacy and AI Trends Report — the subprocessor you didn’t know about. DataGrail analysed roughly 2,400 software vendors and found that nearly 64% of those advertising AI capabilities failed to disclose the third-party AI subprocessors actually processing customer data in their legal agreements. In practice, a vendor your organisation approved — and whose privacy assessment your team signed off on — may already be routing your data to an AI model nobody in risk or procurement ever evaluated. The report also found close to a third of AI systems in use engage in at least one high-risk activity, such as automated decision-making, compounding the exposure.
OLG Hamm, Germany — the court that closed the “the AI did it” defence. In May 2026 the Higher Regional Court of Hamm ruled that a company is liable for the erroneous answers its own AI chatbot gives customers, rejecting any suggestion that responsibility sits with the AI vendor or the model itself. The case involved a medical practice whose chatbot gave incorrect information about a practitioner’s qualifications — but the principle applies well beyond healthcare: if you deploy a vendor’s AI tool under your brand, the legal and reputational consequences of what it says are yours to own.
Questions to Ask Your Organisation
- What independent evidence — not just vendor-supplied documentation — backs every AI vendor’s privacy, security and model governance claims currently in use?
- Do your vendor contracts require disclosure of every subprocessor and AI model touching your data, and is that list actually being checked, not just filed?
- Who in your organisation owns the outcome if a vendor’s AI tool gives a customer wrong information under your brand?
- How would you know if a vendor quietly added AI capability to a product you already approved for non-AI use?
- When did your risk, legal, privacy and cyber teams last jointly review evidence for a live AI vendor relationship — not just at onboarding?
- If a regulator asked you to prove independent verification of a vendor’s AI assurances tomorrow, what would you produce?
Where to Start
None of this means walking away from AI vendors — it means treating their assurances as a starting point for your own verification, not the end of it.
Innovation of Risk has practical AI governance and risk tools to help you assess where your organisation’s AI vendor oversight actually stands.
Take a AI readiness snapshot to see where the gaps are before a regulator, a court, or a customer finds them for you.