---
title: "Compliance Automation for Banks: AI Agents and Production Gates - S-PRO"
url: https://s-pro.io/blog/compliance-automation-for-banks-ai-agents-and-gates
lang: en
description: "Which AML, KYC and reporting tasks AI agents can take on in a bank, what stays a human decision, and the production gates to pass before go-live."
updated: 2026-11-16T06:24:05+00:00
source: S-PRO (https://s-pro.io)
---

[Back to insights](https://s-pro.io/blog)

Article

# Compliance Automation for Banks: AI Agents and Production Gates

Author

[Igor Izraylevych](https://s-pro.io/author/igor-izraylevych)

CEO and Co-founder of S-PRO

Published

November 15, 2026

Time to read

13 min

Table of content

- Key Takeaways
- Table of Contents
- What to automate and what stays human
- What regulators expect today
- Controls that must exist before the agent touches data
- Validation and monitoring by use case
- Risks agents create and how to handle failures
- Production gates checklist
- Our view: start with one task and build the evidence file
- From pilot to production with S-PRO
- FAQ
- Sources

Share

AI agents can take over the preparation work in bank compliance: collecting evidence, screening, drafting alert summaries and assembling files. The decision stays with a named human, and no agent goes live until it passes a fixed set of gates. This article shows what to automate, what to keep human and which gates to require before production.

It is written for Heads of Compliance, MLROs, CROs and CTOs at banks in Switzerland, Germany and the US who want to speed up KYC, adverse media, sanctions screening and alert triage without replacing their existing AML systems.

S-PRO

Move Compliance AI Toward Production

S-PRO builds AI agents for AML, KYC and compliance, with human-in-the-loop controls and regulatory requirements in mind.

[Visit](https://s-pro.io/)

## Key Takeaways

- Automate preparation and triage (gathering, matching, summarizing, drafting). Keep decisions on onboarding, escalation and suspicious activity reporting with accountable people.
- FINMA leaves the risk classification of AI applications to each institution, so your gate document must define it, with materiality and risk exposure as criteria.
- Production readiness rests on evidence: data quality checks, third-party due diligence, explainability, drift monitoring and independent review.
- Regulatory dates differ by jurisdiction. In the US, the 2018 joint statement welcomes pilots while AML duties remain, and the 2026 AML/CFT program rule is a proposal, not final.
- Treat the pilot as a gated process: each gate has an owner, a test and a written result.

## Table of Contents

- What to automate and what stays human
- What regulators expect today
- Validation and monitoring by use case
- Production gates checklist
- From pilot to production with S-PRO
- FAQ
- Sources

## What to automate and what stays human

Automate the work that precedes a decision, and keep the decision itself with a person. An agent that prepares a case in minutes still hands it to an analyst who owns the outcome.

The split below is our working view, and it is a starting point for your own risk assessment, not a classification of any use case. Check Annex III against the use case before you fix the split.

| Compliance task | Agent can do | Stays human |
| --- | --- | --- |
| KYC onboarding | Extract data from documents, check completeness, pre-fill the file, flag inconsistencies | Acceptance or rejection of the client, approval of high-risk relationships |
| Sanctions screening | Match names against lists, rank candidate hits, assemble context | Confirming or discharging a true match, any freeze or block decision |
| Adverse media | Search sources, deduplicate, summarize, cite the source of each finding | Judging relevance and weight for the client’s risk rating |
| Alert triage in transaction monitoring | Enrich alerts with client context, group related alerts, draft a case summary | Closing alerts, escalating, deciding on a suspicious activity report |
| Regulatory reporting support | Collect data, draft narratives, reconcile figures to source systems | Sign-off and submission |
| Regulatory change monitoring | Detect and summarize new texts, map them to policies | Interpreting obligations, changing policy |

Three rules keep the split stable.

1. **The agent never closes a case on its own.** It proposes, and a person disposes.
2. **Every agent output cites its inputs.** A reviewer can see which document, list entry or transaction produced each statement.
3. **Your existing AML system stays the system of record.** The agent reads from it and writes proposals back to it, which keeps your rules, scenarios and audit history intact.

Our pages on [AML and transaction monitoring agents](https://s-pro.io/ai-agents-aml-transaction-monitoring-bank) and [KYC onboarding agents](https://s-pro.io/kyc-onboarding-ai-agents-for-banks) describe how we apply this split in practice.

## What regulators expect today

No single rulebook governs AI agents in compliance. Switzerland relies on supervisory guidance and existing law, the EU has a dated AI Act timetable, and the US has supervisory statements plus a pending proposal.

### Switzerland: FINMA Guidance 08/2024 and the FADP

Switzerland has no overarching AI-specific legislation. The Federal Council signed the Council of Europe’s AI Convention in March 2025, but it is not yet part of Swiss law, and a consultation draft of the necessary legal measures is expected by the end of 2026, according to the [Legal 500 country guide for 2026](https://www.legal500.com/guides/chapter/switzerland-artificial-intelligence/?export-pdf=). The Federal Act on Data Protection (FADP) applies whenever personal data is processed. It covers transparency, data minimization, data protection impact assessments, automated individual decisions and profiling, all of which touch a KYC or screening agent.

FINMA Guidance 08/2024 sets supervisory expectations for AI. As summarized in an [academic white paper from OST](https://www.ost.ch/fileadmin/dateiliste/3_forschung_dienstleistung/institute/ifl/pdfs/white-paper_ai-scalable-framework-for-swiss-fi_v1.0-final.pdf), FINMA does not designate specific AI applications as high-risk. Each supervised institution makes that determination. The guidance calls for a broad definition of AI that covers traditional and generative systems, and for criteria that categorize applications by materiality and risk exposure, including the potential for regulatory non-compliance or reputational harm.

The observations that matter most for a compliance agent are data quality, third-party due diligence, explainability, drift monitoring and independent review. We use them as the backbone of the gates below.

### EU: AI Act and DORA

If your bank serves EU clients or operates in the EU, the AI Act applies to the relevant systems. Regulation (EU) 2026/1744 (the AI Omnibus) has been in force since 27 July 2026. Under the timetable on the [European Commission’s AI Act page](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), stand-alone high-risk (Annex III) obligations apply from 2 December 2027, and high-risk AI embedded in products (Annex I) applies from 2 August 2028. The date 2 August 2026 is the general application date of the Act, with exceptions, not the date for high-risk obligations.

The Digital Operational Resilience Act, [Regulation (EU) 2022/2554](https://eur-lex.europa.eu/eli/reg/2022/2554/oj), sets requirements for ICT risk and third-party ICT providers in EU financial entities. An agent running on an external model or hosted service belongs in that third-party register.

### US: joint statement and the proposed AML/CFT program rule

The US agencies’ [joint statement of 3 December 2018](https://www.fincen.gov/sites/default/files/2018-12/Joint%20Statement%20on%20Innovation%20Statement%20%28Final%2011-30-18%29_508.pdf) encourages banks to pilot innovative approaches to AML compliance. It also makes clear that Bank Secrecy Act and AML obligations remain in full. FinCEN’s April 2026 proposed AML/CFT program rule names generative AI as an innovative approach. It is a proposal, not a final rule, so plan for it without treating it as settled.

Nothing in this article is legal advice. Take your classification and notification questions to your legal and compliance function and, where relevant, to your regulator.

## Controls that must exist before the agent touches data

Every control below needs an owner and a written record. The goal is that an auditor can reconstruct what the agent saw, did and recommended on any given day.

- **Data access.** Give the agent read access only to the fields the task needs. Apply role-based access and log every query. This supports FADP data minimization.
- **Swiss data hosting.** Decide where client data and model inference run, and document it. Private banks often require Swiss hosting for client-identifying data. Confirm it for every component, including the model provider and logging store.
- **Model testing.** Test on your own historical cases, including known true positives, edge cases and multilingual names. Record the result per version.
- **Human oversight.** Define who reviews which output, the review time limit and what happens when the reviewer disagrees with the agent.
- **Audit trail.** Store inputs, outputs, model version, prompt version, reviewer and decision for each case. Keep them for the period your record-keeping rules require.
- **Integration with existing AML systems.** Connect through documented interfaces and keep the monitoring rules, case manager and sanctions engine as they are.
- **Third-party due diligence.** Assess model providers and hosting partners as outsourcing and ICT third-party relationships, with exit and substitution plans.

## Validation and monitoring by use case

Validate each agent against the task it performs, with a metric a compliance officer can read. A single accuracy figure across KYC, screening and reporting hides the cases that matter.

- **KYC and onboarding.** Measure extraction accuracy per document type, completeness flags that were wrong in either direction, and the share of files a reviewer had to rework.
- **Sanctions screening.** Measure recall on a labeled set of true matches first, because a missed match is the costly error. Then measure how many false candidates the agent correctly ranks lower.
- **Adverse media.** Check source coverage and whether every finding links to a retrievable source. Test for hallucinated sources and for confusion between people with the same name.
- **Alert triage.** Compare the agent’s priority ranking with analyst outcomes on past alerts. Sample the low-priority bucket regularly to catch missed risk.
- **Regulatory reporting support.** Reconcile every figure the agent drafts to the source system and flag any narrative statement without a reference.

After go-live, monitoring continues. Track input drift (new document formats, new client mix, new typologies), output drift (shifts in ranking or summary patterns) and reviewer override rates. A rising override rate is an early sign that the agent no longer fits the data.

## Risks agents create and how to handle failures

The main risks are wrong outputs, silent degradation, data exposure and over-reliance. Each needs a detection method, an escalation path and a record.

| Risk | Typical failure | Detection | Response |
| --- | --- | --- | --- |
| Operational | Agent unavailable, queue backlog | Service monitoring, queue age alerts | Fall back to the manual process, log the incident |
| Model | Wrong match, invented source, inconsistent summary | Sampling, source checks, override rate | Pause the affected function, review recent cases |
| Drift | Performance falls as data changes | Scheduled back-testing against labeled sets | Retest, retrain or roll back to the prior version |
| Cybersecurity | Prompt injection through a document or web page, data leakage | Input filtering, output scanning, access logs | Isolate the agent, notify the security team |
| Compliance | Decision taken without required review, missing audit record | Control testing, audit trail completeness checks | Re-review the affected cases, correct the process |

Two habits make failures manageable. First, define in advance what counts as a reportable incident and who decides. Second, keep a rollback path: the previous agent version, or the manual process, must stay ready to use.

If your team is mapping these risks for a first agent, talk to our engineers about your compliance use case. We can work through the task split and the controls with you.

## Production gates checklist

An agent moves from pilot to production only when each gate has a named owner, a test and a signed result. Use this list as the minimum, and add gates your risk function requires.

1. **Use case and classification gate.** Describe the task, the data and the decision it supports. Record the materiality and risk assessment under your own criteria, and check Annex III against the use case where the EU AI Act may apply. Owner: Head of Compliance with Risk.
2. **Data gate.** Confirm data sources, quality checks, lawful basis and FADP assessment, including a data protection impact assessment where required. Owner: data protection officer.
3. **Hosting and third-party gate.** Document where data and inference sit, complete due diligence on every provider and register the ICT dependencies. Owner: CTO with Procurement.
4. **Validation gate.** Pass tests on historical cases with thresholds set before the test, with results per use case. Owner: model validation or an independent function.
5. **Explainability gate.** Show that a reviewer can see the evidence behind each output and reproduce it. Owner: Head of Compliance.
6. **Human oversight gate.** Fix review roles, limits on agent autonomy and the escalation route. Confirm that no agent output closes a case alone. Owner: COO or Head of Operations.
7. **Security gate.** Test access controls, prompt injection handling and logging. Owner: CISO.
8. **Independent review gate.** Have a function independent of the build team review the file, as FINMA’s guidance observations on independent review suggest. Owner: internal audit or second line.
9. **Operational readiness gate.** Prepare runbooks, rollback, incident criteria and staff training. Owner: Head of Operations.
10. **Ongoing review gate.** Set drift monitoring, back-testing frequency, reviewer sampling and a date for periodic re-approval. Owner: Head of Compliance.

A pilot that passes gates 1 to 7 can run on a limited scope with reviewers checking every output. Gates 8 to 10 decide whether it scales.

## Our view: start with one task and build the evidence file

> We think the fastest route to a compliance agent in production is a narrow first task with a complete evidence file, not a broad pilot with a thin one.
>
>
>
> In our view, the first agent should read from your current AML and KYC systems, write proposals back to them and leave every decision with your staff. That keeps the system of record, the rules and the audit history where your supervisors expect them.
>
>
>
> We design our solutions with EU AI Act, FINMA and DORA requirements in mind, and we treat the gates above as part of the build, not as paperwork at the end. For custom compliance solutions, S-PRO builds it on RegRadar and its agentic platform S-PRO Agents.
>
>
>
> We believe a team that can show validation results, reviewer overrides and drift reports from week one finds approval conversations simpler, because the evidence already exists.
>
>
>
> *The S-PRO team*

## From pilot to production with S-PRO

S-PRO is an international Swiss-American firm with offices in Zurich and Salt Lake City, and global delivery. We build AI agents for banks, run them as a service on our own agentic platform S-PRO Agents. S-PRO is ISO 27001 and ISO 27701 certified. Our team has delivered 100+ AI projects, 40+ of them agentic.

Talk to S-PRO about your AI agent project

Get in touch

For compliance leaders, we start from your use case: [AML & transaction monitoring](https://s-pro.io/ai-agents-aml-transaction-monitoring-bank) for alert triage and screening support, or [KYC onboarding agents](https://s-pro.io/kyc-onboarding-ai-agents-for-banks) for document-heavy onboarding. Our engineers work with your compliance, risk and IT teams on the gates, so the evidence is produced as the agent is built.

## FAQ

### Can an AI agent decide whether to file a suspicious activity report?

No. The agent can assemble the case, summarize the transactions and draft a narrative, but the decision belongs to a trained person with accountability. Your policy should state this explicitly and your audit trail should record who decided and on what evidence.

### Does FINMA classify AI compliance agents as high-risk?

FINMA does not designate specific AI applications as high-risk. Under Guidance 08/2024, each institution defines its own criteria, focusing on materiality and risk exposure. Document your classification and the reasoning, and review it when the agent’s scope changes.

### How long does a first pilot take?

It depends on the use case, the data and your integration landscape. For scoping, we offer a design thinking pilot from USD 20k, 2–4 weeks. The output is a defined use case, a task split and a gate plan for moving toward production.

### Do we have to replace our existing AML system?

No. An agent can sit alongside your current transaction monitoring and screening systems, reading alerts and cases and returning proposals. Your system stays the system of record, which preserves rules, history and audit evidence while the agent reduces manual preparation work.

## Sources

- [Legal 500 Country Comparative Guides 2026: Artificial Intelligence, Switzerland](https://www.legal500.com/guides/chapter/switzerland-artificial-intelligence/?export-pdf=)
- [OST white paper: A Scalable Framework for Implementing AI in Swiss Financial Institutions (2025)](https://www.ost.ch/fileadmin/dateiliste/3_forschung_dienstleistung/institute/ifl/pdfs/white-paper_ai-scalable-framework-for-swiss-fi_v1.0-final.pdf)
- [European Commission: AI Act regulatory framework](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai)
- [EUR-Lex: Regulation (EU) 2022/2554 (DORA)](https://eur-lex.europa.eu/eli/reg/2022/2554/oj)
- [Joint Statement on Innovative Efforts to Combat Money Laundering, 3 December 2018](https://www.fincen.gov/sites/default/files/2018-12/Joint%20Statement%20on%20Innovation%20Statement%20%28Final%2011-30-18%29_508.pdf)

Last updated 15 November 2026

Subscribe

Get technical insights, and industry news

Thank you!

Get ready! You will receive handpicked content right to your inbox.

---

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 first**ISO 27001 and ISO 27701 certified.
- **Claude Partner Network**S-PRO is a Certified partner in the Claude Partner Network. Our team includes Claude-certified engineers and architects.
- **Proven in regulated finance**Banks, wealth managers and fintechs in Switzerland, the EU and the US.

Who you’ll talk to

**Michael Barskyi**Head of Banking & Fintech

**Dustin Schulze**Senior Business Analyst

Prefer email or phone?

[hi@s-pro.ioEmail](mailto:hi@s-pro.io)[+41 41 539 15 83Phone](tel:+41415391583)

Claude Certified Services Partner(Anthropic)

**4.9**46 reviews on Clutch

Contact us

### Send us your request

Share your ideas, findings and questions with us.

Link copied to clipboard

Get in touch
