Contents
- Responsibility stays with the fiduciary, agent or not
- What the FADP says about automated decisions
- FINMA's review points work as a yardstick for fiduciaries too
- Six building blocks of an approval that proves something
- Approval as a click, approval as evidence
- Three mistakes that break approval in daily work
- Test your own approval step in one month
- Frequently asked questions
Key points
- Anyone using AI on a client engagement remains liable for diligent and faithful performance under Article 398 of the Code of Obligations. An error by the tool is no excuse.
- An agent is not a second pair of eyes. It reads everything and prepares; a named person must decide, together with a second person for material cases.
- An approval only proves something if it rests on rules in plain language, shows the source for every proposal, allows rejection and logs who decided, when and under which rule.
- Under Article 21 FADP, people can ask for an automated individual decision to be reviewed by a human. If a person reviews and decides first, as we understand it, no such decision arises.
- Three mistakes make approval worthless: approval fatigue, approving without ever rejecting and missing thresholds. All three show up in three numbers: proposals, rejections and time per approval.
Responsibility stays with the fiduciary, agent or not
Much fiduciary work counts in law as a mandate under Articles 394 ff. of the Code of Obligations. Under Article 398(2) CO, the mandatee is liable to the mandator for the diligent and faithful performance of the business entrusted. The Swiss fiduciary association TREUHAND|SUISSE (in German) spells out what that means for AI: a fiduciary generally cannot escape liability towards a client by pointing out that a wrong answer, calculation or assessment came from an AI system. The tool does not change the standard.
That leads to a sober division of labour. The agent reads, sorts and proposes. A person decides and is liable. The closer a proposal comes to the core professional judgement, the more closely that person has to look; the association puts it the same way. What remains open is whether that looking can be proven later.
An agent is not a second pair of eyes in this arrangement. It takes over the groundwork, and control stays with people. With AI, the four-eyes principle comes down to this: the agent reads everything, a named person checks every flag, and for material cases a second person looks as well.
What the FADP says about automated decisions
Article 21 FADP covers the automated individual decision: a decision based exclusively on automated processing that has a legal consequence for the person concerned or a considerable adverse effect on them. The person must be informed. On request, they may state their point of view and ask for the decision to be reviewed by a human being. The FDPIC stresses that this right applies to AI as well.
Most of what an agent proposes in a fiduciary office is no such decision. A booking proposal affects nobody. Steps that act directly on the outside world come closer: a payment reminder sent without anyone looking at it, or, in property management, a rejection sent to a prospective tenant. The FADP protects natural persons. A limited company as a client is not covered; its employees and your private clients are.
As we understand it, no automated individual decision arises if a person reviews the substance and decides before anything takes effect. Whether a click without review is enough is a legal question; settle it for your case with your legal adviser. In practice the answer is simpler. An approval that checks nothing proves nothing.
FINMA's review points work as a yardstick for fiduciaries too
FINMA Guidance 08/2024 describes seven areas in which FINMA assessed how supervised institutions use AI, from governance to explainability. The last is independent review. FINMA observed that the development of AI applications and their independent review were not always clearly separated.
Two further observations apply directly to approval. FINMA assessed whether institutions had defined thresholds or other validation methods to keep the quality of outputs in check over time. And it treats as part of monitoring the analysis of cases in which users ignored or changed an output, because such corrections point to weaknesses. That makes every rejection a data point.
The guidance is addressed to supervised financial institutions; a fiduciary office without a licence is not bound by it. It still works as a yardstick, because a client in a dispute wants to know the same thing: who checked, and how that can be seen. What the guidance contains in detail is covered in the article on FINMA Guidance 08/2024.
Six building blocks of an approval that proves something
- 01Named personFor every process, it is fixed by name who approves and who deputises. Both can judge the proposal on its merits.
- 02Rules in plain languageWhat the agent checks against is written in sentences a new colleague would understand. Whoever approves can read every rule.
- 03Read everythingThe agent reads every document and every line instead of a sample. The person checks every flag that results.
- 04Source referenceEvery proposal names the document, page or line and the rule that applied. Without a reference, the person can only trust the agent.
- 05Approve or rejectThe person approves, rejects with a reason or passes the case to a second person. Nothing is silently corrected.
- 06LogThe log records document, rule, proposal, decision, person and time, plus the reason for any rejection. At Grow Amplified this is called the replay log.
Remove any one of these and approval becomes a ritual. Without a reference, the person checks against a feeling. If rejection is not an option, they can only confirm. And with no log, nobody knows a year later who saw what.
Approval as a click, approval as evidence
| Feature | Approval as a click | Approval as evidence |
|---|---|---|
| Who | Whoever has time | A named person with a deputy |
| Basis | A feeling that it is right | Rules in plain language, the same for every file |
| Scope | A sample, or nothing read at all | The agent reads everything, the person checks every flag |
| Proposal | A result without a source | A result with source reference and rule |
| Options | Confirm | Approve, reject with a reason, pass to a second person |
| Trail | A tick without a date | A log with person, time, rule and document |
| A year later | Nobody remembers why | The decision can be replayed |
The right-hand column costs more work to set up and more time per approval. In return, it answers the question that matters in a dispute: what the person had in front of them when they decided. How long such a log is kept and how it stays verifiable is covered in the article on the replay log and record keeping.
Three mistakes that break approval in daily work
Approval fatigue
Whoever approves a hundred proposals in a row reads the eightieth differently from the first. TREUHAND|SUISSE calls this automation bias: people place too much trust in automatically generated recommendations. Worked example: 300 proposals a month. Confirming each in 20 seconds takes 100 minutes; opening every source reference takes closer to 2 minutes per proposal, so 10 hours. If the log shows an average of 20 seconds, someone clicked. The remedy: clear matches under a fixed rule are approved as a batch, with the rule in the log. Every exception crosses the desk on its own.
Approving everything
A rejection rate of 0% over three months rarely means the agent works without errors. Usually it means nobody rejects any more. FINMA counts the analysis of changed and ignored outputs as part of ongoing monitoring. The remedy: count and read rejections every month. If a reason keeps recurring, it becomes a new rule in plain language.
No thresholds
Without thresholds, an expense receipt for CHF 40.- gets the same click as a provision of CHF 40'000.-. Thresholds set when a second person approves: above an amount, for a type of transaction, for every borderline case. HSLU puts it as a rule of thumb in its real estate blog (in German): high impact means a human must decide. A threshold turns that rule of thumb into a rule that appears in the log.
Test your own approval step in one month
Take one process in which someone already approves today, with or without an agent. Write on a single page who approves, by which rules and above which threshold a second person joins. Then count three things for a month: proposals, rejections and time per approval. Those three numbers show whether your approval checks or only confirms.
The Flat-Fee Leak Check offers a trial run by hand: every line of a time-tracking export gets a mark, and your initials beside extra work and borderline cases are the approval. That is the principle in miniature, with source and trail. How Grow Amplified builds approval and the replay log into every review lane, and who owns them afterwards, is set out under Approach. Which data an agent may see for this is covered in the article on client data and AI.
Frequently asked questions
Is the AI provider liable if the agent gets it wrong?
Towards your client, you are liable in the first place under Article 398 of the Code of Obligations. Whether and how far you can turn to the provider depends on your contract with it. Read the liability clause before the first client file runs through the tool, and clarify details with your legal adviser.
Is a sample enough, as in an audit?
Samples exist because people cannot read everything. An agent can. The question then becomes whether the person looks at every flag the agent sets while reading in full. FINMA lists samples as one possible way to check outputs, alongside backtesting and predefined test cases.
Who should approve: the employee who runs the engagement or the partner?
The person who can judge on the merits whether the proposal is right, and who is named for it. Above a threshold, a second person joins. What matters is that both appear by name in the log.
What belongs in the log of an approval?
At least: which document, which rule, which proposal, who approved or rejected it, when, and the reason for any rejection. How long you keep it depends on what the log belongs to; more on that in the article on the replay log and record keeping.
Does the FADP rule on automated decisions also apply to companies as clients?
The FADP protects natural persons. Article 21 FADP concerns decisions about people, such as private clients, tenants or a client's employees, and not about a limited company as such. The duty of care under the mandate applies to every client.
15 years in banking and consulting, in transformation and finance operations. Today he builds review lanes with Grow Amplified: agents read and reconcile, a person approves, and a replay log records who decided what.