A bank should run security due diligence on an AI agent vendor as an outsourcing review first and an AI review second: classify the service, check the evidence behind the vendor’s certificates, fix the contract terms, and keep monitoring after go-live. For Swiss banks the anchor is FINMA Circular 2018/3 on outsourcing. DORA and the EU AI Act add checks only where they apply to your entity or your use case.

This guide is written for CISOs, vendor risk managers, procurement leads and the COO or CIO who sign off. It covers what to ask, what evidence answers it and where a certificate stops being enough.

S-PRO
s-pro.io
Put AI Security Controls Into Practice
S-PRO builds AI agents for banks with EU AI Act, FINMA and DORA requirements in mind, and holds ISO 27001 and ISO 27701 certifications.

Visit

Key Takeaways

  • Start by classifying the service: what data the agent sees, what it can do, and whether the outsourcing is significant under FINMA rules.
  • ISO 27001 and ISO 27701 show that a management system exists. They do not show how a specific agent, model provider or data flow behaves in your environment.
  • Contracts must settle data locations, subcontractors, audit access, incident notice and exit before the pilot starts, not after.
  • Logging, human oversight and reassessment decide whether an agent stays safe in production. Ask for them in the design, not as an add-on.
  • DORA applies to EU financial entities, and the EU AI Act to specific high-risk uses. Check both against your own entity structure and use case.

Table of Contents

Step 1: Classify the service before you assess the vendor

Classification comes first because it decides how deep every later check goes. An agent that drafts internal summaries from public data is a different risk from one that reads client files, screens names against sanctions lists or writes into a core system.

Record four things in a short scoping note:

  1. Data: which categories the agent reads or writes, including bank client identifying data and personal data under the Swiss Federal Act on Data Protection (FADP).
  2. Actions: whether the agent only recommends, or can trigger a payment, a client message or a record change.
  3. Role in the process: whether the function is critical, and whether the arrangement counts as significant outsourcing.
  4. Hosting and model path: where the application runs, where data is stored, and which model provider and subprocessors sit behind it.

FINMA Circular 2018/3 sets requirements for outsourcing by banks and securities firms. Its fuller duties, such as an inventory, due diligence on the provider, a written agreement and security requirements, apply to significant outsourcing. Under the circular the bank remains responsible for the outsourced function. Your note should therefore state plainly why the service is or is not significant, and who signed that view off.

What ISO 27001 and ISO 27701 show, and what they do not

An ISO 27001 certificate shows that an independent auditor assessed the vendor’s information security management system within a defined scope. ISO/IEC 27001 is the international standard for such systems. ISO/IEC 27701 extends that system to privacy information management.

That is useful evidence of discipline: risk assessment, access policy, incident handling, supplier management. It leaves several questions open, and your review has to close them:

  • Scope: does the certificate cover the teams, locations and systems that will run your agent?
  • Statement of Applicability: which controls were selected, and which were excluded and why?
  • Your data flows: a management system certificate does not describe where your prompts, documents and outputs go.
  • AI-specific risks: prompt injection, model provider changes and output quality are not settled by the certificate alone.
  • Regulatory fit: neither standard is a FINMA, DORA or EU AI Act assessment. Treat it as one input to your own.

S-PRO is ISO 27001 and ISO 27701 certified. Ask any vendor, including us, for the certificate, its scope and the Statement of Applicability, then map them to your own control catalogue.

Outsourcing control under FINMA Circular 2018/3

Under FINMA’s outsourcing circular, the bank keeps responsibility and control even when an AI vendor runs part of the process. For significant outsourcing, the circular expects the bank to select, instruct and monitor the provider with care and to document the arrangement.

Translate that into a working file for each AI vendor:

  • Inventory entry: service, owner, data categories, locations, subcontractors.
  • Selection record: the due diligence you ran and the evidence you reviewed.
  • Written agreement: security requirements, audit and information rights, and rules on subcontracting.
  • Ongoing monitoring: service levels, incidents, changes and periodic review.
  • Exit plan: how you take the function back or move it, without losing access to your data.

If client identifying data leaves Switzerland or becomes accessible to foreign staff, the review must also cover the confidentiality duties of the bank. Involve legal and compliance early, not after procurement has chosen a shortlist.

For AI specifically, FINMA published Guidance 08/2024 on governance and risk management when using artificial intelligence in December 2024. It is a supervisory guidance, not a circular, and it is useful as a checklist for governance, model inventory, data quality, testing and the handling of third-party AI. Use it to frame the AI-specific part of your questions.

The Swiss cloud and subcontractor angle

Most AI agents run on cloud infrastructure and call a hosted model, so the cloud and subcontractor chain is part of the vendor’s risk. The Swiss Bankers Association’s Cloud Guidelines (2025) say that data security, covering confidentiality, integrity, availability and traceability, should be a high priority in due diligence when choosing a cloud provider and its subcontractors.

Cooling plant outside a Swiss cloud data centre

They also say the institution should be told in advance about a change of significant subcontractor. They expect suitable precautions to bring outsourced functions and client data back in-house or to a new provider, including termination arrangements and a free choice of export interfaces and formats. These points map directly to contract clauses, covered below.

Checklist: the security evidence to request

Request evidence, not assurances. The table lists what to ask for and what a good answer looks like.

Area Evidence to request What to look for
Governance Security policy, named owner, ISO certificates with scope and Statement of Applicability Scope covers the delivery teams and systems that touch your data
Access control Access model, privileged access process, joiner and leaver routine Access limited to staff who need it; stronger controls and review for privileged accounts
Encryption Description of encryption in transit and at rest, key management, who holds keys Clear key ownership; options for bank-controlled keys
Vulnerability management Patch process, scanning cadence, penetration test summary Independent testing and a tracked remediation process
AI-specific testing Tests for prompt injection, data leakage, output errors; evaluation method Tests run on your use case, with results you can read
Incident response Incident plan, notification path, roles, past exercise results Defined notice to the bank and a named contact
Data residency Hosting regions, backup locations, support access from other countries Locations written into the contract, not just a website
Subprocessors Full list including the model provider, with their roles and locations Advance notice of changes and a right to object
Retention and deletion Retention periods, deletion method, deletion confirmation No use of your data for model training unless you agree in writing
Logging What is logged, how long, who can read it, export format Logs that let you reconstruct each agent action and decision

Access to systems that handle critical data should be regulated systematically, monitored and limited to authorised staff with a task need, with extra controls for privileged access. That principle runs through Swiss data protection practice and should be tested in the vendor’s real setup, not read from a policy.

Data residency, model providers and training

Data residency has three layers: where the application runs, where the data is stored, and where the model processes it. A vendor can host the application in Switzerland while the model call leaves the country, so ask for the whole path.

Put these questions to the vendor in writing:

  • Which model or models does the agent call, and who operates them?
  • Are prompts and outputs stored by the model provider, and for how long?
  • Is your data excluded from training, by contract?
  • Can the agent run with a model hosted in a region you approve?
  • Can staff in other countries see your data through support or maintenance access?

For personal data, the FADP applies in Switzerland. It contains rules on disclosure abroad and on data protection impact assessments for processing that may pose a high risk to the persons concerned. Your data protection officer should sign off the data flow map before the pilot begins.

What DORA adds for EU entities

DORA applies to EU financial entities, so it reaches a Swiss bank only through EU subsidiaries, branches or services it provides there. For those entities, Regulation (EU) 2022/2554 (DORA) sets rules on ICT third-party risk in Articles 28 to 30, and it has applied since 17 January 2025.

Article 30 lists the contract topics. For an AI vendor, these are the ones to test:

  • Locations: where services are provided and data is processed and stored, with notice of changes.
  • Subcontracting: conditions for subcontracting ICT services that support critical or important functions.
  • Audit and access: access, inspection and audit rights for the entity and the authorities.
  • Incidents: assistance when an ICT incident occurs.
  • Exit: termination rights and exit strategies, with a transition period.

If your group includes an EU entity, run one contract standard that satisfies both Swiss and EU expectations. It is cheaper than two negotiations.

The EU AI Act: check only if the use case is high-risk

The EU AI Act matters to you only when your entity falls under EU law and the use case is high-risk. Check Annex III against the use case; do not classify it from the vendor’s marketing. Article 26 sets the obligations of deployers of high-risk systems, including use according to instructions, human oversight, log retention and monitoring.

The dates are set by the AI Act as amended by Regulation (EU) 2026/1744 (the AI Omnibus), in force since 27 July 2026. Stand-alone high-risk (Annex III) obligations apply from 2 December 2027. High-risk AI embedded in products (Annex I) applies from 2 August 2028. The general application date of 2 August 2026 is not the date for high-risk obligations. Confirm current dates on the European Commission’s AI Act page.

Even where no high-risk duty applies, the same practices (human oversight, logs, documentation) are sound controls for any agent in a bank.

Contract terms that make due diligence enforceable

Due diligence only protects you if the findings end up in the contract. Settle these terms before the pilot, because a pilot that handles real data is already outsourcing in practice.

  1. Audit and information rights for the bank, its auditors and FINMA, including on subcontractors.
  2. Incident notification with a defined window, content and contact.
  3. Security changes: notice and approval for material changes to hosting, model provider or architecture.
  4. Service continuity: recovery objectives and a tested continuity plan.
  5. Data return and deletion: export in an open format, then deletion with written confirmation.
  6. Exit: transition support so the process can move to another provider or back in-house.
  7. Subprocessor list with advance notice and a right to object.

Talk to our engineers about a security review for your AI agent project

Controlling the vendor in pilot and production

Control continues after signature: a pilot is the cheapest moment to test logging, oversight and incident paths, and production is when they matter. Build four routines.

  • Logging: every agent action, input source, tool call and output is recorded so a reviewer can reconstruct a decision. Agree the retention period and who can read the logs.
  • Human oversight: define which outputs a person approves before they reach a client or a system of record, and record the approval.
  • Change control: model version changes, prompt changes and new data sources go through your change process.
  • Reassessment: review the vendor at a fixed interval and after any incident, subcontractor change or scope expansion.

A path from pilot to production should carry these controls with it. Our page on moving AI agents from pilot to production describes that route for banks.

Question list for the vendor questionnaire

Use these questions as the core of your questionnaire. Each requires a document or a named owner, not a yes.

  1. Which parts of your organization, and which sites, are in the scope of your ISO 27001 and ISO 27701 certificates?
  2. Which data categories does the agent process, and where is each stored and processed?
  3. Which model providers and subprocessors are involved, and where do they operate?
  4. Is our data used for training or evaluation by you or any provider?
  5. Who on your side can access our data, from which countries, and how is that access approved and logged?
  6. How do you test the agent for prompt injection, data leakage and incorrect outputs, and can we see results?
  7. What does the agent log, for how long, and can we export it?
  8. How and when will you notify us of an incident, including one at a subprocessor?
  9. What audit rights do we, our auditor and the regulator have, including over subcontractors?
  10. How do we exit: in what format do we get our data, and who deletes what, with what confirmation?

Our perspective

Banks that pass vendor reviews quickly treat security as part of the agent’s design. They decide at the start where data lives, which actions need a human, and what gets logged, and they write those decisions into the contract.

At S-PRO we build AI agents for banks on our own agentic platform, S-PRO Agents, and run them as a service. We have delivered 100+ AI projects, and our engineers work with bank security and vendor risk teams on exactly these questions. We hold ISO 27001 and ISO 27701 certifications, and we treat them as the start of a conversation about your scope, not the end of it.

Our advice to a CISO is to bring the questionnaire above to the first workshop. A vendor that can answer it with documents, early, will be easier to supervise for years.

The S-PRO team

How S-PRO supports banks through vendor due diligence

S-PRO is an international Swiss-American firm with offices in Zurich and Salt Lake City, and global delivery. Our AI agents for banks are built on S-PRO Agents and run as a service, with R&D centers in Zurich and Łódź. For wallet-level and platform work around the agent, we also deliver fintech software development.

In a due diligence process we provide our ISO 27001 and ISO 27701 certificates and answer your questionnaire with your security and vendor risk teams. Solutions are designed with EU AI Act, FINMA and DORA requirements in mind, and the final assessment stays with your bank.

Book a discovery call with S-PRO

FAQ

Does an ISO 27001 certificate replace a bank’s own vendor review?

No. The certificate shows that an auditor assessed the vendor’s security management system within a stated scope. Your review still has to check that the scope covers your service, and examine data flows, model providers, logging and contract terms for your specific use case.

Does FINMA Circular 2018/3 apply to every AI vendor?

It applies to outsourcing by Swiss banks and securities firms, and its fuller duties apply to significant outsourcing. Whether an AI agent counts as significant depends on the function it supports and its data. Document that assessment, who approved it and the reasons, before you start the vendor review.

When does DORA apply to a Swiss bank?

DORA applies to EU financial entities, so a Swiss bank is affected through EU subsidiaries, branches or services it provides in the EU. For those entities, Articles 28 to 30 on ICT third-party risk have applied since 17 January 2025. Ask legal counsel to map your group structure.

When do EU AI Act high-risk obligations apply?

Stand-alone high-risk (Annex III) obligations apply from 2 December 2027, and high-risk AI embedded in products (Annex I) from 2 August 2028, under the AI Act as amended by Regulation (EU) 2026/1744. Check Annex III against the use case and confirm dates on the Commission’s AI Act page.

What should happen to our data when the contract ends?

The contract should require return of your data in an open, agreed format, followed by deletion with written confirmation, including at subprocessors. It should also provide transition support so you can move the function to another provider or bring it back in-house without a gap in service.

Sources

Last updated 14 November 2026

Subscribe
S-PRO newsletter
Get technical insights, and industry news

    By subscribing, you agree to our Privacy Policy .
    This site is protected by reCAPTCHA and Google. Privacy Policy and Terms of Service apply.
    Subscription confirmed
    Thank you!
    Get ready! You will receive handpicked content right to your inbox.