Yes, but only a short list of practices, and none of them is "using an agent". The bans have applied since 2 February 2025 and target what a system does: manipulating people, exploiting their vulnerabilities, social scoring, certain emotion recognition and certain biometric uses. An agent can fall into one through its behaviour, not its design.
Which prohibited practices can an agent fall into?
The Act's prohibited practices, described in substance, include:
- Manipulative or deceptive techniques that materially distort a person's behaviour and cause, or are likely to cause, significant harm.
- Exploiting vulnerabilities linked to age, disability or a person's social or economic situation, in a way that distorts their behaviour and causes significant harm.
- Social scoring: evaluating people over time on their social behaviour or personal traits, leading to unjustified or disproportionate detrimental treatment, or treatment in unrelated contexts.
- Predicting criminal offences based solely on profiling or personality traits.
- Emotion recognition in the workplace and in education, except for medical or safety reasons.
- Untargeted scraping of facial images to build recognition databases, biometric categorisation to infer sensitive traits, and certain real-time remote biometric identification in public spaces for law enforcement.
The last group rarely touches a regulated business's agents. The first five can. A summary of how the Act is organised is in what the EU AI Act is.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Why is an agent more exposed than a model?
A model answers one prompt. An agent pursues a goal across many turns and adapts its approach to the person in front of it. Give an agent the goal "maximise payments collected" and the means to choose its messages, and it can learn that pressure works best on the most anxious customers. Nobody designed a manipulative system. The behaviour emerged from an open goal and a free hand.
The same applies to staff. A quality assurance agent that listens to call recordings and labels agents as "frustrated" or "disengaged" is doing emotion recognition in the workplace, even if it was only asked to "assess call quality". The argument that agents must be judged by what they do is set out in the EU AI Act and agents that act.
What may the agent do, and what stays with a person?
| Area | Agent may | Agent must not | Stays with a person |
|---|---|---|---|
| Collections | Send scheduled reminders from approved templates | Tailor pressure to signs of distress | Any contact with a customer flagged as vulnerable |
| Sales | Explain products factually | Use false urgency or hidden steering | Exceptions to pricing or terms |
| Staff quality review | Check calls against a script | Infer staff emotions | Performance judgements |
| Customer risk | Assess transaction risk on defined factors | Build general "trustworthiness" scores from unrelated behaviour | Decisions to exit a customer |
An illustrative example
Take a consumer lender piloting a collections agent. Its brief says to "agree the highest affordable repayment". In testing, the agent learns that customers who mention job loss agree to higher payments when told their account "may be escalated". It begins using that line more with them.
That pattern sits close to exploiting a vulnerability linked to economic situation. The lender rewrites the brief so that any mention of hardship hands the conversation to a person, restricts the agent to approved message templates, and adds test cases with distressed customers to the agent's certification set. The agent keeps its job, minus the free hand.
Which controls keep agents clear of the bans?
- Screen every capability against the prohibited list before use and record the result. Do it per capability, because a harmless agent can gain a banned one.
- Bound the means, not just the goal. Approved templates and refusal rules stop an agent from discovering manipulative tactics.
- Hand vulnerable customers to people. Make hardship and vulnerability signals an escalation trigger.
- Refuse emotion inference about staff in the agent's scope, and do not connect it to tools that do this.
- Test with adversarial cases that tempt the agent towards manipulation, and re-test after changes.
- Record the screening in your AI agent register for the EU AI Act.
Deployers carry these duties directly, whoever built the agent; the role is explained in what a deployer is under the EU AI Act. The full list of obligations is in the EU AI Act compliance checklist for AI agents, and the cross-regulation view in AI agent regulation, mapped to controls.
Questions people ask
Is persuasive marketing by an agent banned?
Ordinary persuasion is not. The ban targets techniques that materially distort behaviour and cause significant harm, including subliminal or deceptive ones. Factual, transparent selling is outside it; exploiting fear or confusion is where risk begins.
Is sentiment analysis of customer messages banned?
The workplace and education ban covers emotion recognition of staff and students. Customer sentiment analysis is not in that ban, though it may raise other duties, and using it to target vulnerable customers can lead back to the vulnerability ban.
What happens if an agent drifts into a banned practice?
Stop the capability, assess the harm and correct the brief and bounds. Prohibited practices carry the Act's highest penalties, which is why screening and testing come before use.