Skip to content
Chat with AI Agent
Business & governance

Automated Decisions Swiss Clients Accept: Designing Self-Learning Agents Around FADP Art. 21

Orange ITS — AI engineering team 8 min read

A customer asks your support team why their request for a higher credit limit, a faster delivery slot or a policy renewal was declined. The answer sits in a system that learned from last quarter’s corrections, and nobody can say which rule fired. The honest reply is “the software decided”, which is a poor thing to say to a client in Zurich or Lugano. It also ends the conversation.

Buyers worth winning rarely object to automation itself. They want to know that a person can be reached, that the reason can be explained and that their side of the story is heard. Our local inference work keeps the agent and its decision records on infrastructure you control, which is where that evidence is easiest to protect.

Design the agent so that every consequential decision leaves an explanation trail and a path to a human, and the legal duty becomes a reason clients pick you. This article is a design guide. It is not a legal analysis.

Scope note: AI Agents and Swiss Data Protection covers residency, DPIAs and processor contracts. Self-Learning AI Systems explains what improves on its own and what needs people. This piece joins the two at the point where an agent decides something about a person. It is general information and not legal advice.


What Article 21 Asks For

Article 21 of the revised Federal Act on Data Protection (FADP, SR 235.1) is short. Paragraph 1 says the controller informs the person about “any decision that is based exclusively on automated processing” and that has “a legal consequence for or a considerable adverse effect on the data subject”. The act calls this an automated individual decision. Under paragraph 2 the controller, on request, lets the person express their point of view, and the person may request that a natural person review the decision.

Paragraph 3 removes both duties in two cases: the decision is directly connected with concluding or processing a contract and the person’s request is granted, or the person explicitly consented to the decision being automated. The act has applied since 1 September 2023, and the Federal Data Protection and Information Commissioner (FDPIC) states that it applies directly to AI-supported processing.

Two words in that text drive the design. “Exclusively” means a person who genuinely weighs the case sits outside the article. “Considerable adverse effect” means a mislabelled newsletter segment does not qualify, while a refused application usually does. Published Swiss case law and FDPIC guidance on Article 21 for agents that adapt over time is still sparse, so whether a given system is covered depends on its facts. Your counsel decides that. The rest of this piece is about building so the question is easy to answer.


Sort Agent Decisions Into Three Groups

Most agents in a company make dozens of small choices and a handful that matter to a person. Sort them before you design anything.

GroupExampleDesign consequence
Internal and reversibleTagging a ticket, drafting a reply, ordering a queueLog it. No special path needed
Recommendation with a real human decisionAgent proposes a refund, a clerk decidesKeep the human genuine and measure it
Consequential and automatedDeclining an order, rejecting a candidate, ending a serviceInform and hear the person, then offer review by a person

The middle row tempts teams most. If a clerk approves 400 agent proposals an hour, the human step is decoration. Either give the clerk the time and data to disagree or treat the row as automated and build the review path. A useful check is the override rate. A person who never overrides the agent cannot be pointed to when a client complains.

Keep the third group small on purpose. Every decision you move from automated to recommended is one fewer that needs a hearing and a review queue, and one more that a client can talk to a human about.


Why Self-Learning Makes Explanations Harder

A fixed rule engine gives the same answer on Tuesday and on Friday. A self-learning agent may not. As the self-learning guide describes, most production systems learn by writing lessons into memory files, rules and test cases outside the model, and sometimes by updating weights. Each correction changes what the agent does next.

That is useful for quality and awkward for accountability. When a client asks about a decision from three weeks ago, “the agent works like this today” is the wrong answer. You need to say how it worked then. Learning also needs a person in the loop: a lesson becomes an active rule only after someone approves it, so the version history has an author and a date.


The Explanation Trail: What to Log Per Decision

Timing matters. Build the record when the decision is made, because reconstructing it after a complaint is guesswork. For each consequential decision, store:

  • Inputs. The data fields the agent actually used, with their values as of that moment, and which it ignored.
  • Versions. The model identifier, the prompt version, the active rules and the memory entries that were retrieved.
  • Outcome and confidence. What was decided, what alternatives were considered and any threshold that applied.
  • Route. Whether the decision went straight out or was held for review, and who the named reviewer was.
  • Notice. When and how the person was informed.

Generate the plain-language explanation from this record. An explanation the model writes afterwards from its own impressions can sound plausible and still describe a different path than the one taken. A template filled from logged fields, such as “your request was declined because the delivery postcode is outside the covered zone, rule version 14”, can be checked by the reviewer and by the client.

Records of this kind contain personal data, so apply the same retention and access limits as the underlying case file. Keeping them in a system you operate is simpler than reconstructing them from a provider’s logs. The AI agent governance playbook covers audit logging and ownership in more depth.


The Right to Be Heard: Make It a Real Channel

Paragraph 2 gives the person the chance to express their point of view. In practice that is a reply path in the decision notice, a form or a mailbox, that creates a case rather than a ticket nobody reads.

Three design points make it work:

  1. The notice names the channel. It says that a person can be asked to review, and how, in the language the client uses.
  2. The statement enters the case. What the person writes is attached to the decision record and is visible to the reviewer. If it adds a fact the agent lacked, the decision is re-run with that fact.
  3. The answer comes from a named role. A reviewer with a name or a role, a target response time and the authority to change the result.

Illustratively, a firm might promise a reply within two working days and track whether it keeps that promise. The number matters less than having one, publishing it and measuring it.


Review by a Natural Person Without the Rubber Stamp

The reviewer needs four things to make the review real: the full decision record, the person’s statement, the authority to overturn and no penalty for doing so. A reviewer shown only the agent’s conclusion will tend to confirm it.

Track these measures per month:

  • how many consequential decisions were issued, and how many were challenged
  • how many reviews changed the outcome
  • how long a review took from request to answer
  • how many overturned cases came from a missing or wrong input

A challenge rate near zero can mean clients are content or that the notice is hard to find. An overturn rate of zero over hundreds of reviews deserves a question about whether reviewers have the means to disagree. A moderate overturn rate with fast reviews shows the path works. Show it to clients.


Overturned Decisions Are the Best Learning Data

Each time a person changes the agent’s outcome, you have an expert-labelled example of where the agent was wrong. That is the material a supervised improvement loop needs, and it ties this article to the rest of the series: feed these cases into your test set, check that the fix holds on past cases and approve the rule change through a named owner before it goes live.

The rule is simple. Self-learning here means a person-approved cycle of capture, test, approve and release. An agent that rewrote its own decision rules on every complaint would produce exactly the unexplained behaviour this article is trying to remove. Because every active rule has a version, a date and an approver, the explanation trail and the learning loop use the same records.


Turn the Duty Into a Sales Argument

Procurement teams in banking, insurance, healthcare, HR and the public sector often ask the same questions about automated systems: who can a person speak to, how is a decision explained, what is logged and where does it live. A supplier who answers with a one-page description of the review path, a sample decision record and the monthly measures above is easier to approve than one who answers with a policy PDF.

Three practical moves:

  1. Put the review path in the proposal. One diagram: decision, notice, statement, reviewer, outcome, logged change.
  2. Show a redacted decision record. It proves the trail exists.
  3. Offer the measures in your service report. Challenge rate, overturn rate and review time give a client something to hold you to.

EU customers add a parallel rule. Article 22 of the GDPR protects people from decisions based solely on automated processing, and the EU AI Act adds duties for high-risk systems. The Digital Omnibus, in force since 27 July 2026, moved the main high-risk date to December 2027, as we explain in The EU AI Act High-Risk Deadline Moved to December 2027. One design serves both audiences.


What Is Still Moving

Switzerland is preparing AI-specific rules. In February 2025 the Federal Council decided to ratify the Council of Europe AI Convention and to regulate by sector where possible, with general rules only in areas such as data protection. According to the Federal Office of Communications (BAKOM), the consultation on that proposal is scheduled for spring 2027. Earlier timelines pointed to the end of 2026, so the schedule has moved. Checked on 7 November 2026; we will revisit it when the draft appears.

Article 21 applies now, whatever the timetable. A review path, an explanation trail and a supervised learning loop will be needed under almost any version of the coming rules, and building them now costs less than retrofitting them after the first challenge.

Frequently asked questions

Does FADP Article 21 apply to every AI agent?

No. It covers a decision based exclusively on automated processing that has a legal consequence for the person or considerably affects them. An agent that drafts replies or suggests a priority, while a person decides, is outside that definition if the person genuinely decides. A refusal, a termination or a price set without a person is the typical case inside it.

What must a company give a person who receives an automated decision?

The controller must inform the person about the decision. On request it must let them express their point of view, and the person may ask that a natural person review the decision. Two exceptions exist in paragraph 3: a decision tied to a contract where the person's request is granted, and explicit consent to automation. Ask counsel how they apply to you.

Can a self-learning agent make automated decisions at all?

Yes, if you can explain and review them. The difficulty is that an agent whose rules change can give different answers to the same case on different days. Logging the rule and memory version with each decision, and approving new rules through a person, makes each decision reproducible and reviewable later.

Why would a human review path help win customers?

Buyers in finance, insurance, HR and the public sector often ask in procurement how a person can challenge an automated outcome. A documented review path with a named role, a response time and a decision record answers that question before it is asked, and it shortens the security and compliance review.

Is a human who clicks approve enough to avoid Article 21?

Not necessarily. The article speaks of decisions based exclusively on automated processing. A reviewer who only confirms the agent output, without the time, data or authority to disagree, may not count as a genuine human decision. Measure how often reviewers change the outcome. A rate of zero is a warning sign.

Insights

Put these ideas to work

A 30-minute call is enough to find out whether an AI agent fits your workflow — and what it would return.