Every vendor in this category now has an AI Act paragraph. Most of them imply that a scoring model is regulated machinery requiring conformity assessment. It is worth reading the annex, because the answer is narrower than the marketing, and the thing that does apply is being ignored while everyone looks at the wrong clause.
What Annex III actually lists
High-risk classification is not a judgement call about how consequential your model feels. It is a list. Annex III names eight areas:
| # | Area |
|---|---|
| 1 | Biometrics |
| 2 | Critical infrastructure |
| 3 | Education and vocational training |
| 4 | Employment, workers management and access to self-employment |
| 5 | Access to essential private and public services and benefits |
| 6 | Law enforcement |
| 7 | Migration, asylum and border control |
| 8 | Administration of justice and democratic processes |
Sales prospecting is not there. The closest thing is inside area five, which covers systems used “to evaluate the creditworthiness of natural persons or establish their credit score” — an assessment that decides whether a person gets a loan, not whether a rep makes a call.
The difference is the consequence borne by the person scored. Being declined credit is a closed door. Being ranked 34th on a prospect list is a phone call that may not come. The annex draws the line there, and a B2B lead score sits on the safe side of it.
The obligation that does land
Systems that interact with people carry a transparency duty: the person has to be aware they are dealing with AI, and synthetic content has to be identifiable as generated. That is a real requirement, it is not burdensome, and in a go-to-market stack it points at exactly one component — the chat widget.
It is worth noticing how low that bar is, and how many widgets still clear it only by accident: a human first name, a typing indicator and no disclosure anywhere is a design that was chosen, and the transparency duty is the law catching up with why it was chosen.
The regulated surface is the one that talks to people, not the one that ranks them.
The law that has applied since 2018
While the AI Act was being drafted, GDPR Article 22 already governed decisions taken solely by automated means that produce legal effects or similarly significantly affect someone. Alongside it, Articles 13 to 15 require you to tell people that such decision-making exists and give meaningful information about the logic involved.
For scoring companies, this mostly does not bite: a company is not a natural person. The boundary is crossed the moment the record carries people — and every serious GTM record does:
- The named buying committee is a list of identifiable individuals.
- An email-confidence score is an assessment attached to a person.
- Behavioural history from your own site is a record of what one person did.
- Any inference about seniority, authority or intent is an inference about a human being.
None of that is prohibited. It does mean the honest answer to “what do you do with my data” has to be available, which is the same requirement we wrote about in what legitimate interest actually covers — and, again, it is answerable only if you recorded the source at the time you collected it.
What we do about it
Two things, both of which predate the Act and neither of which is generous.
The chat refuses questions its corpus does not cover and cites the page it answered from, which we wrote about in the best thing a chat widget can say. A widget that cites its source is transparent about being a machine by construction, not by disclaimer.
And the score is kept as three separate numbers — fit, intent and confidence — rather than one composite. A single number is unexplainable by design; three tell you which part is weak, which is what “meaningful information about the logic” means when a person asks. That is set out in the evidence standard.