Three signs your AI tool may count as high-risk

Look beyond the model itself. Ask what decision follows, who is affected and whether a wrong output can block someone from work, credit, education or a public service.

The decisive detail often sits in the workflow. A model may only recommend, but if employees routinely accept that recommendation without checking it, the system has substantial influence in practice. Document actual behaviour, not only the process diagram.

When signals point to high risk, pause expansion and document the system. Early classification gives your team options. Late discovery removes them.

Pausing expansion is not the same as banning innovation. It creates room to involve the supplier, test performance for different groups and design genuine human oversight before more people depend on the outcome.

Ask what role the supplier believes it has under the AI Act, which intended uses the system was designed for and what evidence supports its performance claims. Then ask which changes, incidents or limitations it will report to you.

Do not accept a generic security page as the complete answer. You need information about the version you use, your configuration and the data flows in your contract. If the supplier cannot provide that, record the uncertainty as a risk rather than silently assuming it is safe.

The process owner understands the decision, IT understands integration and access, privacy understands personal data, and legal understands the regulatory test. None of them has the full picture alone.

Capture the conclusion in plain language, including assumptions and the date for reassessment. Revisit it when the purpose, model, data, user group or degree of automation changes. Classification is a living decision, not a sticker placed on a tool forever.