Most support decisions an AI agent makes, such as answering a question or issuing a small credit, fall outside GDPR Article 22. A few do not: closing an account or refusing a large refund. Support reviewers must be trained to recognise those decisions, to reverse them, and to treat "I want to speak to a person" as a legal right.
What does Article 22 mean in a support operation?
Article 22 protects people from decisions based solely on automated processing that have legal effects or similarly significant effects on them. Such decisions are allowed only where necessary for a contract, authorised by law with safeguards, or based on explicit consent. In the contract and consent cases, the organisation must at least let the person obtain human intervention, express their point of view and contest the decision.
Support is a high-volume, low-stakes environment with a few high-stakes moments buried in it. That shape is the training problem. A support reviewer who handles hundreds of routine agent conversations a day can easily treat the significant one the same way. The core skill is triage: knowing which agent decisions could affect someone's rights, money or access to an essential service.
Typical significant decisions in support include:
- Closing or suspending a customer account.
- Refusing a refund, warranty claim or chargeback on a large purchase.
- Ending a subscription to a service the person relies on, such as connectivity or utilities.
- Restricting access to stored data, funds or a marketplace seller account.
- Refusing a request to correct or delete data.
Who does it apply to?
Any organisation running support agents that make decisions about individuals: retailers, marketplaces, software and subscription businesses, utilities, telecoms providers, banks and insurers. The reviewers whose competence matters are:
- Support advisers, the people, who take over from the agent.
- Escalation and specialist teams with authority over refunds, closures and restrictions.
- Trust and safety staff who review account suspensions.
- Complaints teams handling contests.
The question of whether support staff need a certificate to approve agent actions turns on which of these roles hold release rights.
What makes review meaningful in support?
The test for meaningful human review has three elements: the reviewer has the authority to change the decision, the competence to judge it, and access to the relevant information. Support operations often fail the first.
Frontline advisers frequently cannot reverse an account closure or approve a refund above a small limit. If a customer asks for a person and reaches an adviser who can only repeat the agent's decision, there has been no human intervention in any real sense. The fix is routing: a significant decision held or contested must reach someone who can change it.
Set the agent's capabilities by depth and effect. At depth 1 it answers questions from the knowledge base. At depth 2 it drafts responses for an adviser to send. At depth 3 it issues credits and refunds up to the customer's limit and fixes routine account settings. Account closures, large refunds and service terminations are held for a named person. Some actions, such as reversing a settled payment on the firm's own rail, are withheld for every agent.
What must support reviewers be trained to do?
- Recognise significance. Spot the decisions in the list above, even when they arrive in a routine queue.
- Read the whole conversation. The agent's summary may drop context a customer gave three messages earlier.
- Check the evidence class. Order and payment records are sealed. A customer's account of a faulty item is declared. A screenshot is external. Decisions should rest on the right class, and reviewers should know which one a claim belongs to.
- Use authority, or route to it. Reverse when warranted, or send the case to someone who can, quickly.
- Hear the customer's view. A contest brings new facts. Weigh them.
- Spot manipulation. Some customers try to push an agent into a refund; some agents are bounded too tightly and refuse fair requests. Reviewers should notice both.
- Record the reason. Every verdict is filed; the reviewer's reason should be clear enough for an auditor or the customer.
Daily practice is covered in how customer support agents review an AI agent's decisions.
How do you meet it and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Identify significant decisions | Support decisions mapped by effect; significant ones held | Decision map; hold configuration |
| Human intervention on request | "Talk to a person" route reaches staff with authority | Routing rules; handoff logs |
| Reviewer competence | Advisers and specialists trained on the agent and on triage | Training and assessment records |
| Real authority | Release and reversal rights defined by role | Role definitions; reversal records |
| Right to contest | Contests re-decided by a trained person | Contest log with outcome and reason |
| Review is not a formality | Seeded significant cases placed in routine queues | Detection results by team |
What will a supervisor or auditor ask?
- Which support decisions does the agent take that could significantly affect a customer?
- When a customer asks for a person, who do they reach, and what can that person change?
- Show us an account closure: who decided it, and on what evidence?
- How do advisers spot significant cases in a routine queue?
- What do contested decisions show about the agent's bounds?
What does Article 22 not cover in support?
It does not cover routine answers and small goodwill gestures with no significant effect. It does not require a person to read every conversation. It does not replace consumer rights to refunds and remedies, which apply in full. It does not cover the duty to tell people they are dealing with an AI system, which sits in Article 50 of the EU AI Act. And it still needs a lawful basis under Article 6 for the processing behind each decision. Compare stricter settings in collections and public services. The guide to AI literacy and certification covers reviewer training broadly.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Support team leads and escalation reviewers can take the Pryme Intelligence Academy Operator track, which teaches how a governed agent decides, human review and when something goes wrong, with tests set in the learner's own industry and role. For a whole support team, the free readiness assessment for up to 25 people shows scores by concept and where the gaps are before an agent goes live.
Questions people ask
Is a refused refund always an Article 22 decision?
No. A refused small refund rarely has a significant effect. A refused refund on a large purchase, or one that leaves someone without an essential item, may.
Can a frontline adviser provide the human intervention?
Only if they can change the decision. If they cannot, route the case to someone who can, and make that route quick.
Should the agent tell customers they can ask for a person?
Yes, for decisions that could be significant. It makes the right to human intervention usable and gives reviewers a clear trigger.