New AI model: how to check after a change that everything still works

Why does a new model change the results?

Every AI model answers differently: in length and tone, in the choice of sources, in keeping to formats and in calling tools. A new model is usually better on average but may handle individual tasks worse. In 2023, researchers from Stanford and Berkeley compared the March and June versions of well-known language models: on individual tasks they differed markedly, sometimes for the better, sometimes for the worse.

For an organisation, this means that an app drafting a discharge letter or an administrative decision in a fixed structure may, after the change, swap sections, leave out details or cite sources differently, even though the instruction and the documents have stayed the same.

What should you check after a change?

Criterion	Question	Example
Content	Are the statements correct, is anything missing?	medication complete in the letter
Sources	Are the right passages cited?	guideline instead of an outdated version
Format	Does the answer keep to the structure?	outline of the administrative decision
Language	German, technical language, Easy Language as specified?	no stray English phrases
Tools	Are connected systems queried correctly?	right mailbox, right search
Limits	Does the AI refuse what it should refuse?	no diagnosis without a basis
Speed	Does the answer arrive in a usable time?	also with many people at the same time

How do you build a set of test cases?

1. Select important apps: the most used ones and those where errors have the greatest consequences.
2. Record cases: for each app, a handful of typical and difficult inputs, with the documents that belong to them.
3. Describe the expectation: What must be in the answer, and what must not? A sample answer helps.
4. Run before the change: save the results of the current model.
5. Compare after the change: specialists from the department assess the results, not only IT.
6. Document the result: date, model and version, cases checked, findings, decision.

If there are deviations, an adjusted instruction in the app often helps. If a problem remains, talk to the provider.

What does the EU AI Act say?

For most uses, such as drafts, summaries and search, an AI platform is not a high-risk AI system, and there is no explicit obligation to test after a model change. Testing is still good practice, because you remain responsible for the results.

High-risk systems are subject to stricter rules: deployers use them in accordance with the instructions for use and monitor their operation (Art. 26). Anyone who substantially modifies a high-risk system may become a provider themselves (Art. 25 with Art. 3 no. 23). A model change can be such a modification if the manufacturer has not provided for it. For systems under Annex III, these obligations apply from 2 December 2027 under the amending Regulation (EU) 2026/1744. Whether your use falls under this: Identifying high-risk AI. For medical devices, their rules apply in addition: Is AI a medical device?

How does this work with basebox?

basebox selects the AI model and tests new models before they are rolled out: for chat, long texts, document search, tool calls, German and load, with tasks from hospitals and public administration. A new model arrives with an update, also without an internet connection; apps, permissions and data stay the same. There is no silent change to a different model. Which model in which version gave an answer is documented, and a model register records the developer, version and licence of each model.

Every version comes with release notes, and you are informed of new versions by email. Nothing installs itself automatically on the server; without an operations contract you install updates yourself and can prepare your test cases beforehand. In the cloud, basebox installs the updates. If you want a particular model, including your own, we deploy it on request.