Control, Responsibility And Governance Of Autonomous AI Agents In Swiss Fintech

As autonomous AI agents increasingly perform decision-making functions within financial institutions, Swiss fintechs must confront fundamental questions of control, accountability and governance. Existing regulatory obligations remain fully applicable: autonomy in the machine does not displace responsibility in the firm. The task is to build oversight frameworks that keep pace with technological independence rather than yield to it.

1. Introduction: the rise of autonomous AI agents in financial services

For most of the past decade, artificial intelligence in finance meant tools: models that scored a credit application, flagged an anomalous payment or drafted a client message, always handing the result back to a human who decided what to do with it. The technology now reaching Swiss financial institutions is different in kind. Autonomous AI agents do not merely recommend; they set sub-goals, call external systems, chain multiple steps together and take governed actions — opening a case, escalating an alert, adjusting a limit, generating and filing a report — with materially less human intervention at each step.

Adoption is no longer hypothetical. In its April 2025 survey of roughly 400 Swiss institutions, FINMA found that around half already use AI in day-to-day operations, that respondents had on average five applications in use and nine in development, and that the overwhelming majority of AI users rely on generative models — with a corresponding increase in dependence on a small number of large technology providers.¹

The governance discussion has become urgent precisely because the agentic shift compresses the distance between an algorithm's output and a real-world consequence for a client, a counterparty or the market. Where a human reviewer once stood between model and money, an agent may now act directly. That is an operational gain and a legal exposure at the same time. This article sets out, under Swiss law, where responsibility lies when an autonomous agent acts, and what a defensible governance and control framework looks like for a Swiss fintech.

2. What makes an AI agent "autonomous"?

"Autonomy" is a spectrum, not a binary. For governance purposes, four characteristics distinguish an agent from a conventional AI tool:

• Goal-directed decision-making — the system is given an objective and selects among possible actions to achieve it, rather than executing a fixed, pre-specified rule.
• Interaction with external systems — the agent can read from and write to other applications (ledgers, case-management tools, payment rails, communication channels), so its decisions produce direct effects.
• Self-learning and adaptation — behaviour may change over time as the model is updated or as the agent incorporates feedback, which means yesterday's validation does not guarantee tomorrow's conduct.
• Multi-agent operation — agents increasingly call other agents, creating chains in which responsibility for a given outcome is distributed across several components and, often, several vendors.

In a Swiss fintech these capabilities surface in recognisable forms: AI "compliance officers" that triage alerts and draft suspicious-activity reports; autonomous transaction-monitoring systems that continuously screen flows and act on thresholds; customer-support and onboarding agents that verify identity, run sanctions and PEP screening and approve or reject applications; and advisory agents that assemble and personalise investment proposals. The legally salient point is constant across all of them: the closer an agent comes to acting instead of a person rather than for a person, the more the institution's control and accountability arrangements must be deliberately engineered rather than assumed.

3. The Swiss regulatory landscape: existing rules apply even when AI acts

3.1 A technology-neutral framework, with no AI-specific financial statute

Switzerland has deliberately not enacted a horizontal "AI Act". On 12 February 2025 the Federal Council confirmed a sector-specific, technology-neutral approach: rather than a single cross-cutting law, it will ratify the Council of Europe Framework Convention on Artificial Intelligence (which Switzerland signed in March 2025), make targeted amendments to existing legislation, and confine general cross-sectoral rules to fundamental-rights areas such as data protection. The responsible departments are to deliver a consultation draft by the end of 2026.²³

For financial institutions this means the operative rules are the ones that already exist. FINMA applies them on a "same business, same risks, same rules" basis, and its Guidance 08/2024 of 18 December 2024 makes the supervisory expectation explicit: institutions must identify, assess, manage and monitor the specific risks of their AI use cases within the existing requirements on adequate business organisation, risk management and internal control. FINMA does not treat AI as inherently high-risk; the level of expectation scales with the materiality of the application.⁴

3.2 The instruments that already bind AI-driven conduct

The principal Swiss instruments engaged when an autonomous agent operates inside a financial institution include:

The decisive message is simple: an AI agent is not a regulated entity and has no legal personality. It can be neither licensed nor sanctioned. Supervisory and civil responsibility attach to the regulated firm and the natural persons who run it. Deploying an agent therefore reallocates work, not accountability.

4. The accountability challenge: who is responsible when an AI agent acts?

Suppose an autonomous agent makes a harmful or unlawful decision — it wrongly rejects an eligible client, executes an unsuitable transaction, fails to escalate a money-laundering red flag, or leaks personal data. Swiss law does not leave the question of responsibility open. Because the agent itself cannot bear duties, liability is channelled to the persons and entities behind it through several established routes.

4.1 Contractual liability: you cannot delegate liability to a tool

Where the institution uses an agent to perform an obligation owed to a client, Article 101 of the Code of Obligations treats the agent as an auxiliary: the principal answers for the conduct of those it employs to discharge its obligations as if it had acted itself. A firm cannot escape contractual liability by pointing out that "the system decided". Coupled with the general rule on breach (Article 97 CO), this means defective or non-conforming performance produced by an agent is, in substance, the institution's own breach.⁷

4.2 Tortious and organisational liability

Outside contract, Article 41 CO imposes liability for unlawful, culpable damage, and Article 55 CO holds the business owner liable for damage caused within the business unless it proves it took all due care in selection, instruction and supervision. Translated to AI, the "due care" defence becomes an organisational test: did the institution exercise diligence in choosing the model and vendor, in instructing and constraining the agent, and in monitoring its behaviour? A firm that deploys an autonomous agent without validation, guardrails or oversight will struggle to make that showing.

4.3 Corporate and directors' liability

At the top of the structure, Article 716a CO assigns the board a set of non-transferable and inalienable duties — the overall direction of the company, the organisation of its affairs, and the supervision of the persons entrusted with management. These duties cannot be handed to an algorithm. Article 754 CO then exposes board members and management to liability toward the company, its shareholders and creditors for damage caused by a breach of duty. An uncontrolled delegation of consequential decisions to autonomous systems is itself a potential organisational failing for which directors answer personally.⁸

4.4 Data-protection and product liability

Where an agent takes an automated decision producing legal or similarly significant effects on an individual, the FADP requires the data subject to be informed and given the opportunity to express a view and to request review by a natural person. Separately, where AI is embedded in a product placed on the market, the Product Liability Act can impose strict liability for damage caused by a defect — relevant to fintechs that supply, rather than merely operate, AI-enabled solutions.⁹

4.5 Delegation versus abdication

Across these routes a single principle recurs. Swiss law permits the delegation of functions to autonomous systems; it does not permit the abdication of responsibility for them. The realistic candidates who may bear consequences — the financial institution, its board and senior management, and, by contract, technology and cloud providers and third-party developers — are allocated their share through the firm's governance arrangements and its agreements. The better those arrangements, the more predictable the allocation. The weaker they are, the more liability concentrates, by default, on the regulated firm and the individuals at its head.

5. A governance framework for autonomous AI agents

A defensible framework rests on four pillars: calibrated human oversight, clear decision rights and escalation, risk classification, and documentation and explainability.

5.1 Human oversight — choosing the right model for the risk

Oversight is not a single setting. The appropriate model should be matched to the materiality of the use case:

5.2 Decision rights and escalation

Autonomy must be bounded by explicit authority limits, not left implicit in code:

• Authority limits — define, per use case, the actions an agent may take unaided and those it may not (e.g. monetary thresholds, transaction types, client categories).
• Approval thresholds — specify the points at which an action requires human sign-off before execution.
• Escalation mechanisms — ensure that uncertainty, anomalies, low-confidence outputs or out-of-bounds situations route to a named human owner, and that the agent can be paused or overridden at any time.

5.3 Risk classification and materiality

Consistent with FINMA's proportionate approach, each agent should be classified by risk, weighing factors such as the impact of failure, the number and profile of affected clients (retail or institutional), the importance of the process, and the complexity, predictability and explainability of the agent's behaviour. High-risk uses — those touching capital, customer outcomes or regulatory parameters — warrant the strongest controls; FINMA expressly invites institutions to engage with it in good time before deploying AI in critical processes.

5.4 Documentation and explainability

If a firm cannot explain what an agent did and why, it cannot supervise it, defend it to a client or auditor, or answer to FINMA. The minimum is a maintained inventory of AI agents and their risk classifications; model and system documentation describing purpose, data, limits and known weaknesses; and complete, tamper-resistant decision logging and audit trails capturing inputs, actions and the human interventions around them. Explainability is not an ethical nicety; it is the precondition of control.

6. Operational control requirements

Governance principles must be operationalised through concrete controls across the agent lifecycle:

• Model validation — independent pre-deployment testing for robustness, accuracy and bias, with re-validation after material model updates.
• Monitoring and testing — ongoing performance indicators, data-quality checks, stability and robustness testing, and — where appropriate — backtests, stress tests and adversarial tests.
• Incident management — defined detection, containment, escalation and remediation procedures for agent failures, including a reliable means to disable an agent quickly.
• Business continuity — fallback mechanisms and manual-mode procedures so that a withdrawn or failed agent does not interrupt critical services.
• Cybersecurity controls — recognising that an agent with system access and tool-use rights materially enlarges the attack surface; access must be least-privilege, segmented and continuously monitored.

Fintech-specific pressure points. The intensity of these controls should rise with the sensitivity of the process. Payment processing and trading demand the tightest limits and real-time monitoring given the speed and irreversibility of action; customer-facing decisions engage suitability, fairness and data-protection duties directly; and AML/KYC monitoring must preserve a defensible, auditable basis for every automated screening, escalation and filing decision.

7. Third-party and outsourcing risks

Most agentic capability in Swiss fintech is not built in-house. It rests on a thin layer of foundation-model providers and cloud platforms, which creates two linked exposures: dependence on a small number of external providers whose models can change beneath the firm, and concentration risk if many critical processes — or many institutions — rely on the same handful of vendors. FINMA has repeatedly flagged outsourcing of critical functions as a key source of operational risk, and its outsourcing and operational-resilience requirements apply in full to AI arrangements.

Sound governance of the third-party layer requires, at a minimum:

• Due diligence — on the provider's security, model governance, data handling and financial stability before onboarding.
• Audit and information rights — contractual access sufficient to let the institution, its auditors and FINMA verify the service.
• Service levels and change control — defined performance, availability and notification obligations, including notice of material model changes.
• Clear allocation of responsibility — explicit contractual treatment of liability, indemnities and responsibility for AI outputs; the institution cannot, however, contract away its supervisory responsibility toward FINMA and its clients.
• Exit strategy — a workable plan to substitute or repatriate the function, so that vendor dependence does not become a single point of failure.

8. Lessons from emerging international AI regulation

Although Switzerland has chosen a sectoral path, Swiss fintechs cannot treat international AI regulation as someone else's problem.

The EU AI Act. Regulation (EU) 2024/1689 entered into force on 1 August 2024 and introduces a risk-based framework with obligations for high-risk systems becoming applicable from 2 August 2026. It does not apply to Switzerland as such, but it reaches Swiss actors whose AI outputs are used in the EU — so a Swiss fintech serving EU clients, or supplying AI to EU-based institutions, may fall within its scope and its penalties.¹⁰

The Council of Europe Convention. The world's first binding international AI treaty — which Switzerland helped negotiate, signed in 2025, and is implementing — anchors transparency, accountability and risk-management principles across the AI lifecycle and is the backbone of Switzerland's forthcoming consultation draft.

Principles-based regimes and convergence. The United Kingdom's principles-based, regulator-led model and the broader international trend — visible in supervisory communications across major financial centres — point in a consistent direction: existing supervisory authority is being applied to AI, with growing emphasis on inventories, human oversight, third-party due diligence and auditability. For Swiss fintechs with cross-border operations, the practical implication is to design to the highest applicable standard, since regulatory convergence and extraterritorial effects make a fragmented, jurisdiction-by-jurisdiction approach both costly and fragile.

9. Building an effective AI governance model

Translating the above into practice, a Swiss fintech can follow a six-step roadmap:

1. Establish board-level AI oversight — place ultimate accountability with the board, consistent with its non-transferable supervisory duty.
2. Create AI governance policies — a clear policy framework covering permitted uses, oversight models, data and third-party rules.
3. Define accountability structures — a named owner for every agent and unambiguous roles across the three lines of defence.
4. Implement risk-based controls — calibrate validation, monitoring and human oversight to each agent's risk classification.
5. Maintain auditability and transparency — inventory, document and log, so that every consequential action can be reconstructed and explained.
6. Conduct regular reviews and testing — re-validate after changes and review the framework as capabilities and regulation evolve.

An illustrative governance model distributes responsibility across layers, each accountable to the one above:

10. Conclusion: autonomy requires stronger governance, not less

Autonomous AI agents can meaningfully enhance Swiss fintech operations — faster onboarding, sharper monitoring, leaner compliance, better service. But the legal centre of gravity does not move with the technology. Under Swiss law, responsibility for what an agent does remains with the regulated institution and, through the board's non-transferable duties, with the individuals who govern it. The supervisory expectation is already in force, the civil-liability routes are well established, and the forthcoming implementation of the Council of Europe Convention will reinforce rather than relax them.

The right conclusion is therefore the counter-intuitive one. Greater machine autonomy calls for more deliberate governance, not less: clear accountability, oversight calibrated to risk, defined decision rights, rigorous documentation and disciplined control of third parties. Swiss fintechs that treat AI governance as a board-level risk-management discipline — rather than a technology afterthought — will be the ones able to capture the efficiency of autonomous agents while keeping firmly within the bounds of Swiss law.

Footnotes

1. FINMA media release, "FINMA survey: artificial intelligence gaining traction at Swiss financial institutions", 24 April 2025 (survey of c. 400 institutions; around half use AI; the large majority of AI users deploy generative AI, with rising dependence on a small number of large technology providers).

2. Federal Council media release, "AI regulation: Federal Council to ratify Council of Europe Convention", 12 February 2025; Switzerland signed the Convention in March 2025; a consultation draft is expected by the end of 2026.

3. Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), adopted 17 May 2024.

4. FINMA, Guidance 08/2024, "Governance and risk management when using artificial intelligence", 18 December 2024.

5. Federal Act on Data Protection (FADP), in particular the provisions on automated individual decisions (information of the data subject; right to be heard; right to request review by a natural person).

6. FINMA Circular 2018/3 "Outsourcing – banks and insurance companies" and FINMA Circular 2023/1 "Operational risks and resilience – banks" apply to AI functions provided by, or dependent on, third parties.

7. Art. 101 Swiss Code of Obligations (CO / OR) (liability for auxiliary persons); cf. Art. 97 CO (liability for non-performance).

8. Art. 716a CO (non-transferable and inalienable duties of the board, including supervision of management) and Art. 754 CO (liability of board members and management); cf. Art. 41 and 55 CO on tortious and business-owner liability.

9. FADP (automated individual decisions) and the Product Liability Act (PrLA / PrHG) for AI embedded in products placed on the market.

10. Regulation (EU) 2024/1689 (EU AI Act), in force 1 August 2024; obligations for high-risk systems applicable from 2 August 2026; extraterritorial reach where AI outputs are used within the EU.