Anyone integrating AI into existing processes or a web application quickly encounters two questions that cannot be answered with data sheets alone: what happens to the data that is fed into the system? And how reliable are the answers that come back? According to its own statements, the TÜV Association is planning a programme that aims to address exactly this: an audit and certification of AI systems that assesses security in handling sensitive data and the quality of outputs.
Details on scope, criteria and timetable have not yet been made public. What is interesting is the direction: until now, trust in AI components has been based primarily on providers' own declarations. An external test seal would at least make some of these statements verifiable.
Why this matters for decision-makers
Procurement processes currently lack a common standard. One provider describes its model quality with benchmark values, another with reference customers, a third with certificates that relate to the data centre, not to the AI system itself. Specialist departments have to derive a decision from these differing statements – often without the ability to verify them technically.
An established test seal would not completely close this gap, but it would create comparability. Above all, it shifts the burden of proof: instead of "our system is secure", the standard becomes "this has been tested according to a defined procedure". For companies with compliance requirements, for example in healthcare, the financial sector or public administration, this is a difference that counts in internal approval.
Two audit dimensions, two types of risk
The two aspects named in the programme address different risks – and should also be considered separately in projects.
- Handling of sensitive data: This concerns questions such as transmission channels, storage periods, use of inputs for further model training, access rights and tenant separation. These are classic information security topics that can be audited because they can be tied to technical and organisational measures.
- Reliability of answers: This dimension is more difficult. Generative systems produce outputs that sound plausible but can be wrong. An audit can measure error rates in defined scenarios, but it cannot provide a guarantee for every future use case.
This distinction is important so that a certificate is not overinterpreted. A tested system remains a statistical system. Anyone integrating it into a process still needs control points wherever an error becomes expensive.
What you can do now, regardless of certificates
Waiting for an audit programme is not a strategy. The criteria that an external audit would apply can already be used as a checklist in your own projects today.
- Data classification before tool selection. Define which data categories may be sent to an external AI system at all. This decision is a governance question, not a technical one.
- Obtain contractual commitments. Clarify in writing whether inputs are used for training, where processing takes place and how long data is retained. Marketing statements on product pages are not a commitment.
- Build your own test cases. Collect 30 to 50 realistic requests from your specialist department with known correct answers. This allows you to measure whether a model change makes your application better or worse – independently of general benchmarks.
- Plan approval steps. Where an incorrect output has legal or financial consequences, a human belongs in the process. This applies to proposal texts as well as to automated information provided to customers.
- Ensure replaceability. Encapsulate access to the AI system in a separate layer of your application. This allows you to switch providers if audit results, prices or terms change.
Limits of a seal
AI systems are continuously updated. A model that has been audited today may behave differently after the next update. Audit programmes must therefore find an answer to versioning and repeat audits – similar to other certifications that are linked to surveillance audits. How this will be solved in concrete terms is still open for the planned programme.
In addition: a certificate assesses a product, not your integration. Whether personal data is correctly masked in a form, whether access rights are properly set in your content management system, whether logs are evaluated – that remains your responsibility and part of your project quality.
Assessment for web and software projects
A verifiable quality standard would turn the selection of AI components from a question of trust into a question of requirements – this makes tenders, internal approvals and the argumentation towards data protection officers easier. Until then, your own documented audit routine with clear test cases and contractual commitments replaces any seal. Anyone who plans their AI connection as a replaceable layer in the architecture from the outset can later react to new audit results without having to rebuild the application.