An AI initiative starts with a specific task, not with the choice of a model. Decide who will use the result, which decision it supports and which measurable baseline you will compare it with. How long a pilot takes, and whether it succeeds, depends on data, integration, review and approvals.
Define the use case and data
Describe inputs, outputs, users, the consequences of errors and where a person will review the result. Record data sources, purpose, access and retention. For sensitive or personal information, clarify the legal basis, transparency and access before you load data into a service. Swiss data protection law requires a data protection impact assessment when planned processing is likely to create a high risk to personality or fundamental rights. This does not automatically apply to every AI project; assess the specific context and seek legal advice where needed.
Run a pilot with measurable criteria
Compare the AI solution with the existing process. Use representative, approved test data and set criteria in advance for quality, error types, latency, cost and escalation. Test unusual or abusive inputs too. Have domain experts review the results and document when the system must not be used.
Scale and operate
Before broader use, you need owners for data, model or service configuration, access, support and incidents. Monitor the metrics that matter for the use case and review changes to the model, prompt, data or provider. Test again after every material change.
Microsoft Foundry provides evaluators for certain risks and tasks; such tools do not replace expert review or organisation-wide controls. Document limitations, allow human override and switch the solution off if it does not meet the agreed criteria.
Sources: FDPIC: Data protection impact assessment, Microsoft responsible AI for Azure workloads, Microsoft Foundry risk and safety evaluators.

