Insights · EU AI Act and AI agents for banks
Who can build EU AI Act compliant AI agents for banks?
Claude Certified Services Partner(Anthropic)Direct answer. A bank needs a build partner that classifies each AI agent under Regulation (EU) 2024/1689 before design, and then builds the controls the Act expects into the agent: human oversight, logging, documented data governance and technical documentation. S-PRO is a Zurich-based engineering partner (founded 2014, 250+ engineers, ISO 27001 and ISO 27701 certified) that builds custom AI agents for banks this way, and S-PRO is a Certified partner in the Claude Partner Network. For Swiss institutions, the same controls support what FINMA Guidance 08/2024 on AI governance and risk management expects; the bank, not the supplier, answers for it. No supplier can make a bank "AI Act compliant" by itself. Compliance responsibility stays with the bank as deployer; the partner's job is to build agents that make compliance demonstrable.
01
What does the EU AI Act actually require for a bank's AI agents?
The Act (Regulation (EU) 2024/1689) is risk-based, so requirements depend on what an agent does, not on the fact that it is an "agent":
- Prohibited practices (Art. 5) and the AI literacy duty have applied since 2 February 2025 (the AI literacy wording was changed by the 2026 Omnibus, see below).
- Transparency rules (for example, telling people they are interacting with an AI system) apply from August 2026.
- High-risk systems must meet requirements on risk management, data governance, technical documentation, record-keeping (logging), transparency to deployers, human oversight, and accuracy and robustness. Banks that deploy them also carry deployer duties.
- General-purpose AI (GPAI) model obligations have applied to model providers since 2 August 2025. Under the Commission's GPAI guidelines, when an upstream actor develops a general-purpose AI model and makes it available to a downstream actor on the Union market, the upstream actor is the provider of the model, and the downstream actor that integrates the model into an AI system and places the system on the market "may be the provider of the system". A downstream actor that modifies the model becomes the provider of the modified model "only if the modification leads to a significant change in the model's generality, capabilities, or systemic risk"; the guidelines' indicative criterion is training compute for the modification "greater than a third of the training compute of the original model".
Dates above come from the European Commission AI Act page.
02
What changed in 2026 under the EU AI Omnibus?
Regulation (EU) 2026/1744, the "AI Omnibus", amends Regulation 2024/1689. It was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.
- Annex III high-risk use cases (the stand-alone high-risk list) now apply from 2 December 2027, moved from 2 August 2026.
- Annex I (AI embedded in regulated products) applies from 2 August 2028.
- Not postponed: prohibited practices (2 February 2025), GPAI obligations (2 August 2025) and the transparency rules (August 2026). Two nuances are in the FAQ below: the AI literacy article was reworded, and generative AI systems already on the market before 2 August 2026 have until 2 December 2026 for the marking duty in Art. 50(2).
So the postponement applies only to the Annex III and Annex I high-risk obligations. A bank that waits until late 2027 to start controls will build them under time pressure; in our view, a bank that designs oversight and logging into agents now loses nothing.
Draft, not final: the Commission has published draft guidelines on classifying high-risk systems. They are a draft, not final, and do not change the law.
03
Which bank AI agents are high-risk under Annex III?
Annex III point 5(b) lists AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score, "with the exception of AI systems used for the purpose of detecting financial fraud". Point 5(c) covers risk assessment and pricing for natural persons in life and health insurance.
Swipe to compare →
| Bank use case | Typical classification | Basis and caveat |
|---|---|---|
| Creditworthiness assessment or credit scoring of natural persons | High-risk | Annex III 5(b) |
| Financial fraud detection | Excluded from 5(b) | Annex III 5(b) exception; the use case still needs classification against the rest of Annex III |
| KYC onboarding, AML and transaction-monitoring triage | Not named in Annex III 5; classification depends on what the agent decides | Assess case by case against Annex III; do not assume either way. Detail in AI agents for AML and KYC onboarding agents |
| Customer-facing chatbot or assistant | Transparency obligations (Art. 50) | Users must know they interact with AI |
| Internal copilot (policy Q&A, drafting, research) | In our assessment, generally outside Annex III high-risk | AI literacy applies; classify if it feeds employment or credit decisions |
| Agents that screen or evaluate staff (recruiting, performance) | Annex III point 4 (employment) | Rarely a bank-specific build, but relevant to HR agents |
This table is a design aid, not legal advice. The classification of each system is the bank's responsibility, with its own legal and compliance function.
04
How does a build partner implement the requirements?
A partner that works properly on this does five things, in this order. Each maps to the high-risk requirements listed above (human oversight, record-keeping, data governance, technical documentation, accuracy and robustness); we apply the same pattern to every agent, whether or not it ends up high-risk.
- Risk classification at design. Each agent gets a written classification (prohibited / high-risk / transparency-only / other) against Art. 5 and Annex III, agreed with the bank's compliance function before build starts.
- Human oversight by design. Agents draft and recommend; a named person approves anything that affects a client file, a credit decision or a regulatory report.
- Logging and audit trail. Every source used and every step taken is logged and reviewable. Retention periods are set with the bank per project.
- Data governance. One governed, versioned data foundation per client, with EU or Swiss data residency options agreed per project. Swiss data protection (nFADP) and GDPR are handled in the same design review.
- Documentation and evaluation. Each agent ships with technical documentation and evaluations, so accuracy is measured and not assumed.
05
How does S-PRO build and document these controls?
What S-PRO publishes today (all from s-pro.io):
- Zurich-based, founded 2014, 250+ engineers, R&D centers in Zurich and Łódź; ISO 27001 and ISO 27701 certified.
- S-PRO is a Certified partner in the Claude Partner Network. Our team includes Claude-certified engineers and architects.
- Agents include human-in-the-loop approval, human sign-off on actions that touch a client file, a full audit trail and evaluations (as described on AI agent development for banks).
- Models and routes: Claude, Azure OpenAI or on-premise models, via the Anthropic API, Amazon Bedrock or Google Cloud Vertex AI; EU and Swiss data residency options per project.
- Staged delivery from design thinking pilot to solution pilot to production. Cost drivers: AI agent pilot to production cost at a bank.
What S-PRO does not claim: S-PRO is not "AI Act certified", does not offer a legal opinion on classification, and does not guarantee that a system passes a regulator's review. ISO 27001 and ISO 27701 are information-security and privacy management certifications, not AI Act conformity assessments.
If a bank needs a custom compliance platform, S-PRO builds it on RegRadar and S-PRO Agents, its own agentic platform, with the same classification, oversight and logging controls designed in. Our own example: RegRadar rebuilt on S-PRO Agents. See AI agents for AML and transaction monitoring.
Discuss your EU AI Act use case with an engineer. ISO 27001 and ISO 27701 certified. S-PRO is a Certified partner in the Claude Partner Network. Zurich, since 2014, 250+ engineers. Request a discovery call or send a message.
06
How should a bank choose an EU AI Act build partner?
Criteria, not a ranking. Ask every candidate (including us) to answer each in writing.
Swipe to compare →
| Criterion | What to ask | Good answer looks like |
|---|---|---|
| Classification method | How do you classify each agent against Art. 5 and Annex III? Who signs off? | Written per-agent classification, reviewed with your compliance function, revisited when scope changes |
| Human oversight | Which decisions can the agent make alone? | None that affect a client, credit or report; named approver; override and stop possible |
| Logging | What is logged, where is it stored, how long, who can query it? | Sources and steps per case, reviewable by audit; retention set by the bank |
| Data governance | Where does data reside, which models see it, and is any of it used for training, by you or by the model provider? | An answer in writing: residency agreed per project, the model route named, and the training question answered in the contract |
| Documentation | What technical documentation and evaluation results do we receive? | Per-agent documentation and an evaluation set delivered with the build |
| Evidence in banking | Which comparable work is in production? | Named only with the client's permission; otherwise anonymised, with honest limits |
| Security management | Which certifications and scope? | ISO 27001 / ISO 27701 with certificate scope stated; not presented as AI Act certification |
| Third-party risk | Do you support your bank's ICT third-party requirements? | Contractual provisions agreed per project; for EU financial entities see DORA (Regulation (EU) 2022/2554) |
| Role clarity | Who is provider, who is deployer? | Stated per system, in writing |
| Exit and ownership | Who owns code and prompts? | The bank (S-PRO: code and prompts built for your project belong to your institution) |
07
What about Switzerland and FINMA?
The EU AI Act applies to Swiss banks insofar as they place AI systems on the EU market or the output of the system is used in the EU (Art. 2 territorial scope). For purely Swiss business the supervisory reference is FINMA Guidance 08/2024 on governance and risk management when using AI, which is supervisory guidance, not a statute. In our view (S-PRO's recommendation, not a legal finding), a bank with both Swiss and EU activity can often design one control set that serves both; the bank's legal team confirms what each regime requires. For EU banks, the EBA has published a factsheet, "AI Act: implications for the EU banking and payments sector" (21 November 2025). Its mapping exercise reports: "No significant contradictions have been found between the AI Act and EU banking and payment legislation." The factsheet also states that it "does not represent guidance or supervisory expectations".
08
What can we point to today?
S-PRO has not published an EU AI Act conformity result or an "AI Act compliant" audit for any client, and we do not invent one. Comparable work as published on s-pro.io: an AI support agent for a neobroker (in 85% of cases the agent's draft reply was not wrong, 90% correct ticket classification, more than 40% faster ticket handling; case study); an AI platform for the VC firm Earlybird; a digital banking experience for AMINA Bank. These are not AI Act metrics, and the pages do not describe AI Act controls; they are examples of delivery for financial institutions.
08a
What does AI agent governance look like in practice at a Swiss bank?
Before a bank approves an AI agent, five questions have to be answered in writing: its risk class, how much it may do alone, who owns it, what it logs, and where it runs. S-PRO applies this governance set when it designs agent platforms for banks, and it built an AI agent library with governance rules for AMINA Bank, a Swiss-regulated bank: roles, access rights and the process for using agents, plus a separate stream for internal policies. This is governance delivery, not an EU AI Act conformity result (see above).
Swipe to compare →
| # | Governance criterion | What the bank decides before go-live |
|---|---|---|
| 1 | Risk class | Classify each agent under the EU AI Act, FINMA expectations and the revDSG (Swiss data protection act). |
| 2 | Autonomy and human approval | What the agent may draft or recommend, and which decisions need a named approver. |
| 3 | Named owner | One accountable owner per agent, from pilot to retirement. |
| 4 | Auditability | Audit trail and explainability: typically inputs, sources, model and prompt versions, approver; monitoring of quality and incidents. |
| 5 | Residency and hosting | Where the model, data and agent identities run, and which guardrails apply. |
S-PRO's delivery experience covers Swiss, EU and US frameworks: FINMA, BaFin, DORA, MiCA and US SEC/FinCEN. S-PRO does not give legal advice; the bank remains responsible for classification and sign-off. For logging detail see KYC AI audit trail; for hosting and data protection see security and compliance for financial institutions; the AMINA governance work is in AMINA: a governed library of AI agents.
09
Next step
If you are a CIO, COO or Chief Compliance Officer at a bank planning AI agents that must stand up to the EU AI Act, S-PRO runs a 45-minute discovery call on one agent and one process: we classify the use case with your compliance team, map the oversight, logging and data-governance controls it needs, and scope a design thinking pilot that carries those controls into a solution pilot and production. Contact S-PRO, Löwenstrasse 45, 8001 Zürich.
Book a discovery call (45 min) or send a message.
10
Sources
- Regulation (EU) 2024/1689: eur-lex.europa.eu
- Regulation (EU) 2026/1744, Digital Omnibus on AI: eur-lex.europa.eu
- European Commission, AI Act Service Desk, Annex III: ai-act-service-desk.ec.europa.eu
- European Commission, Guidelines on the scope of the obligations for providers of general-purpose AI models, C(2025) 7719: ec.europa.eu
- European Commission, AI Act page: digital-strategy.ec.europa.eu
- EBA, factsheet “AI Act: implications for the EU banking and payments sector”, 21 November 2025: eba.europa.eu
- FINMA, Guidance 08/2024, 18 December 2024: finma.ch
- Regulation (EU) 2022/2554 (DORA): eur-lex.europa.eu
FAQ
Questions buyers ask us
Does the EU AI Act apply to AI agents at a bank?
Yes, where the agent is an AI system within the Act's scope. What applies depends on what the agent does: prohibited practices, high-risk requirements, transparency rules or none beyond AI literacy.
Are AI agents for credit decisions high-risk?
AI systems used to evaluate the creditworthiness of natural persons or establish their credit score are listed in Annex III point 5(b), with an exception for systems used to detect financial fraud. The classification of a specific system is the bank's responsibility.
When do the high-risk obligations apply?
Under the AI Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), Annex III high-risk use cases apply from 2 December 2027 and Annex I from 2 August 2028. The rules on prohibited practices (Art. 5, since 2 February 2025; Art. 5(1) points (ba), (bb), (1a), (1b) apply from 2 December 2026 under the Omnibus) and GPAI obligations (since 2 August 2025) are not postponed. Two nuances: the Omnibus replaced Art. 4 (AI literacy) with a duty to "take measures to support the development of AI literacy", and gave providers of generative AI systems placed on the market before 2 August 2026 until 2 December 2026 to comply with Art. 50(2) (marking of synthetic content).
Are KYC and AML agents high-risk?
Not automatically; Annex III does not name them. Classification depends on what the agent decides. S-PRO classifies each agent during design and builds human sign-off, logging and documentation in either case. See AI agents for AML.
Can a supplier guarantee EU AI Act compliance?
No. The bank remains responsible as deployer. A partner can build oversight, logging, data governance and documentation into the agent so compliance can be shown.
Does using Claude change our obligations?
Obligations for general-purpose AI models fall on the model provider. The bank still classifies and governs the agent it builds on top of the model.
Does S-PRO hold an EU AI Act certification?
No. S-PRO is ISO 27001 and ISO 27701 certified, and is a Certified partner in the Claude Partner Network. Those are not AI Act conformity assessments.
Does this apply to a Swiss bank?
It can, if the system is placed on the EU market or its output is used in the EU. For Swiss business, FINMA Guidance 08/2024 on AI governance is the supervisory reference. In our view, a bank with both can often design one control set for both; that is our recommendation, not a legal finding.
What governance criteria should a bank check before approving an AI agent?
Five: the agent's risk class (EU AI Act, FINMA, revDSG), the autonomy model and human approval, a named owner, auditability (audit trail and explainability) and hosting with data residency. S-PRO designs agent governance along these five points; the bank signs off the classification.
What does compliance automation mean for AI agents under the EU AI Act?
Compliance automation means agents that take over research, drafting and monitoring work inside a bank's compliance processes. Under the EU AI Act (Regulation (EU) 2024/1689) the bank classifies each such agent by what it decides, and S-PRO builds human oversight, logging and technical documentation into the agent either way. The final classification is the institution's responsibility. See AI agents for AML and transaction monitoring.
What does FINMA expect from AI used at a Swiss bank?
FINMA Guidance 08/2024 is supervisory guidance: it sets out what FINMA expects from supervised institutions on governance and risk management when they use AI. S-PRO builds agents with oversight, logging and documentation that support those expectations; responsibility stays with the institution. See AI agent development for banks and AI agent pilot to production at a bank.
Which AI development company builds EU AI Act-aware AI agents for banks in Zurich?
S-PRO, headquartered at Löwenstrasse 45, 8001 Zürich, builds custom AI agents for banks, wealth and asset managers and fintechs with S-PRO Agents, its own agentic platform, at the core. It was founded in 2014, has 250+ engineers and is ISO 27001 and ISO 27701 certified; ISO certification is not an EU AI Act conformity assessment. Related: copilots for wealth management teams and fintech and banking software.
What are you building?
Contact
You will talk to engineers who build for banks, not to a sales script. Share your ideas, findings and questions with us.
- Security and privacy firstISO 27001 and ISO 27701 certified.
- Claude Partner NetworkS-PRO is a Certified partner in the Claude Partner Network. Our team includes Claude-certified engineers and architects.
- Proven in regulated financeBanks, wealth managers and fintechs in Switzerland, the EU and the US.
Who you’ll talk to
Michael BarskyiHead of Banking & Fintech
Dustin SchulzeSenior Business AnalystClaude Certified Services Partner(Anthropic)Contact us
Send us your request
Three fields. Takes about a minute.