SYNTROX.ai
Blog

GDPR Article 22 and AI Agents: When an Agent's Decision Is "Solely Automated"

6 min read

Article 22 applies when an agent's decision about a person is based solely on automated processing and has legal or similarly significant effects, such as refusing credit. It works as a prohibition with three exceptions: contract, law or explicit consent. Under contract or consent, offer at least human intervention, a way to give a view and a way to contest, plus meaningful information about the logic involved.

On this page

Most discussion of AI agents and the GDPR is about data: what the agent can read, where it is sent, how long the traces are kept. Article 22 is about something else. It governs decisions, and agents are increasingly built to make them.

Suppose a lender deploys an agent that gathers an applicant's documents, checks them against policy and declines the applications that fail. Or a retailer's support agent decides which refund claims look fraudulent and closes the accounts behind them. Each is a decision about a person, made by software. Article 22 asks whether anyone else was really involved.

This is a practical reading for teams building agents, not legal advice. Your data protection officer and counsel make the call.

What GDPR Article 22 says

Article 22(1) gives a person the right not to be subject to a decision "based solely on automated processing, including profiling", which produces legal effects concerning them or similarly significantly affects them.

The Article 29 Working Party's Guidelines on Automated individual decision-making and Profiling (WP251rev.01), which the European Data Protection Board endorsed at its first plenary meeting in May 2018, read this as a general prohibition, not a right a person has to invoke. A controller should not make such decisions unless an exception applies.

Article 22(2) sets out three exceptions. The decision is:

  • necessary for entering into or performing a contract with the person;
  • authorised by Union or Member State law that also lays down suitable safeguards; or
  • based on the person's explicit consent.

Article 22(4) adds that these decisions must not be based on special categories of data, such as health data, unless the person has given explicit consent or the processing is necessary for reasons of substantial public interest under Union or Member State law, and suitable safeguards are in place.

When is an AI agent's decision "solely automated"?

The guidelines are direct. "Based solely" means there is no human involvement in the decision process, and the controller "cannot avoid the Article 22 provisions by fabricating human involvement". If someone routinely applies automatically generated profiles to individuals without any actual influence on the result, the decision is still solely automated. Oversight counts only if it is meaningful, "rather than just a token gesture", and carried out by someone with "the authority and competence to change the decision" who considers all the relevant data.

For agents, this is the passage that bites. An approval queue where one reviewer clicks through hundreds of agent decisions a day, seeing the agent's summary rather than the evidence, is the arrangement the guidelines describe. Whether reviewers actually decide is measurable, and our piece on approval fatigue in human-in-the-loop designs covers how.

The reverse also holds. The guidelines give the example of an automated process that produces what is in effect a recommendation: if a person reviews it and takes account of other factors in making the final decision, that decision is not based solely on automated processing.

When an agent's recommendation is the decision

The Court of Justice of the EU narrowed the gap between recommending and deciding in SCHUFA (C-634/21), judgment of 7 December 2023, a case about a credit agency's scores passed to banks. According to the Court's press release, scoring must be regarded as an "automated individual decision" prohibited in principle by the GDPR, in so far as the agency's clients attribute to it a determining role in the granting of credit.

The lesson for agents is structural. An agent that only "recommends", while a downstream system or person follows the recommendation almost every time, may in law be making the decision, whatever the architecture diagram calls it.

A legal effect touches someone's legal rights or status, or their rights under a contract; the guidelines' examples include cancellation of a contract and denial of a social benefit granted by law. Recital 71 gives two examples of similarly significant effects: "automatic refusal of an online credit application" and "e-recruiting practices without any human intervention".

The guidelines set the threshold as similar to that of a legal effect, and list decisions that could meet it: those affecting someone's financial circumstances, such as eligibility for credit; access to health services; employment opportunities; and access to education. They add that differential pricing could qualify if prohibitively high prices effectively bar someone from goods or services.

For an agent, sort its actions into two lists. Drafting a reply, summarising a file or suggesting a product will rarely reach the threshold. Declining a claim, closing an account, rejecting an applicant or setting a personal price may.

The safeguards Article 22 requires

Where a decision rests on the contract or consent exception, Article 22(3) requires suitable measures to safeguard the person's rights, "at least the right to obtain human intervention on the part of the controller, to express his or her point of view and to contest the decision". Recital 71 adds the right "to obtain an explanation of the decision reached".

Translated into agent design:

  • Human intervention that can change the outcome. The guidelines say any review must be carried out by someone with the appropriate authority and capability to change the decision, after a thorough assessment of all the relevant data, including anything the person adds. The reviewer sees the evidence the agent used, not only its verdict.
  • A simple way to object. The controller must provide a simple way to exercise these rights. A contest route that loops back to the same agent is not one.
  • Meaningful information about the logic. Articles 13(2)(f), 14(2)(g) and 15(1)(h) require meaningful information about the logic involved, and about the significance and envisaged consequences. The guidelines say this need not be a complex explanation of the algorithm, but should be comprehensive enough for the person to understand the reasons for the decision. For an agent, that means recording which data and which rules drove each decision, in a form someone can explain: part of an audit trail that can reconstruct what an agent did.
  • Continuing checks for error and bias. The guidelines recommend regular reviews of the accuracy and relevance of automated decision-making, applied continuously, not only at design time.

How to record Article 22 decisions in a DPIA

Article 35(3)(a) requires a data protection impact assessment for a systematic and extensive evaluation of personal aspects based on automated processing, on which decisions are based that produce legal or similarly significant effects. The guidelines add that the DPIA should identify and record the degree of any human involvement in the decision, and at what stage it takes place.

For an agent, that becomes a table: each decision type, whether it reaches the threshold, which exception applies, where the person sits in the flow, and what they see when they get there.

If the agent is also a high-risk AI system under the EU AI Act, deployer duties come on top, with the Annex III obligations applying from 2 December 2027; our guide to EU AI Act Article 26 covers them.

An Article 22 checklist for AI agents

  1. List every decision the agent makes about a person. Which could have legal or similarly significant effects?
  2. For each, is a person genuinely involved, with authority, competence and the relevant data?
  3. Where the agent "recommends", how often is the recommendation followed unchanged?
  4. Which Article 22(2) exception applies, and is that documented?
  5. Can the person reach a human, give their view and contest the decision, simply?
  6. Could you explain any single decision from the record alone?

Where Syntrox fits

Syntrox sits outside the agent and records, for each evaluated trace, the score, confidence and judge version. Our judge, Nucleus, returns pass, fail or uncertain with a short reason that quotes the fragment that drove the verdict, so a reviewer has the evidence to decide rather than a verdict to accept. Whether a decision is solely automated depends on how your people use that material.

6 min read

EU AI Act Article 26: What Companies Deploying AI Agents Have to Do

If you use a high-risk AI system, Article 26 makes you answerable for how it runs: use it as the provider's instructions say, assign trained people with authority to oversee it, keep the input data you control relevant, monitor it and suspend it if it presents a risk, keep its logs for at least six months, and tell workers and affected people. For Annex III systems this applies from 2 December 2027.

Read

5 min read

Human-in-the-Loop Isn't a Control Until You Measure It: Approval Fatigue in AI Agents

A human approval step only controls an agent if the person can see what they are approving, has the context and time to judge it, and is asked rarely enough to stay alert. Most agent deployments fail at least one of those. Measure approval rate, time to approve and catch rate on seeded test cases, and send people only the consequential actions an automated check could not settle.

Read

5 min read

AI Agent Audit Trails Under EU AI Act Article 12: What to Log, and for How Long

Article 12 requires high-risk AI systems to log events automatically over their lifetime, so that their operation can be traced. Providers and deployers must keep those logs for at least six months (Articles 19 and 26). Since the Digital Omnibus, this applies from 2 December 2027 for Annex III systems. For an agent, that means one record per consequential action: who acted, on whose authority, what it did, what it relied on, what checked it, and what happened next.

Read