Buying an AI tool: five contract questions before your team pastes data

An AI demo usually shows the best answer the product can produce. Procurement needs to ask what happens to the inputs, who can access them, and what happens when the answer is wrong. Swiss data-protection law applies to AI-supported processing as it does to other technologies. Do not wait for a hypothetical special AI rule before assessing a real deployment. Start with the intended use case and the categories of information employees would enter.

Map data, purpose and responsibility

Identify personal data, business secrets and material covered by customer confidentiality obligations. Determine whether the vendor processes on your instructions or uses information for its own purposes. A data-processing agreement must reflect the actual service rather than a generic label. For high-risk processing, assess whether a data-protection impact assessment is required. Example: drafting a public product description and screening job applicants have very different consequences, even if the same chat interface handles both.

Ask what training and retention really mean

Read the terms for model training, service improvement, human review, abuse monitoring and deletion. An opt-out from training does not necessarily mean no retention or no third-party access. Specify settings and commitments for the product tier you buy. Check subprocessors, data locations and cross-border safeguards. Require a useful process for changes, security incidents and assistance with individuals' rights. Verify the contract against the account settings before allowing sensitive information into the tool.

Allocate output risk honestly

Clarify permitted use, output rights, third-party claims, warranties and liability limits. A vendor saying that you own output does not guarantee that every answer is accurate, unique or free of third-party rights. Decide which outputs need human review and who is accountable for the final action. Do not let a tool automatically make consequential decisions merely because an integration makes that technically easy. Record the limits that employees need to understand in their actual workflow.

Plan the exit before the pilot becomes infrastructure

Ask how prompts, knowledge files, logs and configurations can be exported or deleted. Check renewal, price changes and what happens when a model is retired. Run a bounded pilot with synthetic or appropriately cleared data, then compare observed behaviour with contractual promises. The surprising fact is that switching off model training answers only one procurement question. A useful approval combines legal terms, security settings, user guidance and a named owner who revisits the arrangement when the product or use case changes.