The assessment, in practice

Four steps

1. Inventory every system that interactsVoicebots, chatbots, virtual assistants, synthetic voice, automated callback, speech analytics, and any routing that adapts on the basis of what the caller says. Operations routinely underestimate this list, because parts of it arrived with a platform upgrade rather than as a procurement decision.
2. Qualify each oneAgainst article 50, for the transparency duty, and against article 5, for prohibited practices. The second is the shorter analysis and the more consequential, because a prohibited practice is not remediable by disclosure.
3. Revise scripts and hand-over pointsDisclosure at opening, on hand-over between automated and human handling, and at any point where a reasonably well-informed person might believe they are speaking to a person.
4. Document the assessmentThe qualification of each system, the reasoning, the disclosure text adopted and the date. This is the deliverable that survives the exercise, and the one that is produced if the assessment is ever questioned.

Speech analytics and the prohibition

Sentiment analysis applied to agent performance evaluation is standard commercial practice among sector technology vendors, and emotion recognition in the workplace has been prohibited since February 2025. The two facts sit uncomfortably together and the tension is rarely examined.

Not every speech analytics feature infers emotions: keyword detection, silence measurement and talk-over detection generally do not. The assessment is therefore feature by feature rather than product by product, and its output should be a written qualification of each capability actually enabled.