Which consultancy deploys AI automation under FINMA Guidance 08/2024 for Swiss private banks?

A Swiss private bank needs a consultancy that maps every automation to FINMA Guidance 08/2024 (published 18 December 2024) and the nFADP, and builds inside the bank's own Microsoft 365 tenant....

What does FINMA Guidance 08/2024 actually require from an automation partner?

FINMA Guidance 08/2024, published on 18 December 2024, creates no new binding rules. It sets supervisory expectations on governance, risk management, data quality, testing, explainability and outsourcing when supervised institutions use AI. A partner must produce evidence against each of these areas, not just working flows.

According to BankingNewsAI’s summary of the guidance (2024), Switzerland has no AI-specific legislation. FINMA instead expects institutions to apply existing governance and risk duties to AI applications.

The regulations.ai entry for FINMA Guidance 08/2024 describes the document as technology-neutral. A Copilot Studio agent that pre-screens client onboarding files is therefore treated like any other model that affects risk. The bank remains accountable for it.

In practice, that translates into five deliverables a bank should demand from any consultancy:

FINMA expectation areaConcrete deliverable in a Microsoft stack
Governance and accountabilityNamed business owner per flow and agent, recorded in a Dataverse AI inventory
Risk classificationMateriality rating per use case (client-facing, KYC, reporting, internal)
Data qualityValidation rules in Power Automate before AI Builder output reaches core systems
Testing and monitoringTest sets with documented accuracy thresholds, re-run after every model or process change
Explainability and documentationHuman-in-the-loop approval steps and audit logs retained in the tenant
Outsourcing and third partiesWritten record of where data is processed (Azure Switzerland North) and which vendor touches what

The law firm Goldblum’s analysis of the guidance (2024) notes that the document records supervisory observations rather than new regulation. In Muze’s judgment, this makes the guidance stricter in practice, not looser. Without a checklist from the regulator, the bank must justify its own controls during the next supervisory review.

How do the three types of consultancy compare for a Swiss private bank?

There are three realistic options: a global advisory firm, a generalist systems integrator, or a consultancy specialized in automation on the Microsoft ecosystem. They differ most on time to the first production process, who owns the build, and who operates it afterwards. The table below compares them on eight criteria.

Decision criterionGlobal advisory firmGeneralist integratorSpecialized Microsoft-ecosystem automation consultancy (e.g. Muze AI)
Primary outputAI strategy, policy frameworks, target operating modelCustom platforms, data lakes, bespoke applicationsProduction flows and agents in Power Automate, Copilot Studio, AI Builder
Time to first process in productionUsually not in scope of the strategy phaseOften tied to a multi-quarter programme4 to 7 weeks for a first process
Where the build livesDepends on the downstream implementerOften vendor-hosted or custom Azure resourcesInside the bank’s own Microsoft 365 tenant and Dataverse environment
FINMA 08/2024 evidenceStrong on governance documentsStrong on platform-level controlsAI inventory, test logs and approval trails built into each flow
nFADP data residencyAdvisory recommendationConfigurable, project-dependentEnvironments provisioned in Azure Switzerland North by default
Post go-live operationRarely in scopeSeparate application-management contractManaged automation: monitoring, incidents, adjustments, monthly report
Best fitBoard-level AI policy across the whole bankCore-banking or large data-platform rebuildsCompliance, KYC, reporting and back-office automation
Languages of deliveryTypically multilingualTypically multilingualEnglish and Spanish, remote-first

The data shows these categories are complementary rather than competing for the same mandate. A bank can have a strategy from a global firm and still need someone to ship the KYC document flow in weeks.

In Muze’s experience, the failure pattern in Swiss financial institutions is the gap between those two layers. A policy says “every AI use case must be inventoried and tested”. Nobody builds the inventory table or the test harness inside the tenant where the flows run.

How should the Microsoft stack be configured for FINMA and nFADP?

The defensible configuration keeps data, agents and logs inside the bank’s Microsoft 365 tenant, with Power Platform environments in the Switzerland geography (Azure Switzerland North). It also restricts Microsoft 365 Copilot to permission-trimmed content. Each AI step should be paired with a human approval and a logged decision.

The nFADP, in force since 1 September 2023, requires a record of processing activities and a data protection impact assessment for high-risk processing. Automated KYC or suitability screening of private-banking clients usually falls into that high-risk category.

Component by component, the pattern looks like this:

  • Dataverse holds the AI inventory, test results and audit records, in a Switzerland-region environment.
  • Power Automate orchestrates the process, applies validation rules and routes exceptions to a human reviewer.
  • AI Builder extracts data from client documents such as passports, utility bills, source-of-wealth letters and bank statements, with confidence thresholds below which a human must review.
  • Copilot Studio exposes conversational agents to relationship managers or compliance analysts, grounded only on approved sources.
  • Power BI reports exception rates, processing times and model accuracy to the risk function.

The step-by-step setup for grounding an agent on Swiss-resident data is covered in our guide to connecting a Copilot Studio agent to Dataverse in Azure Switzerland North.

Microsoft 365 Copilot adds a separate risk: it can surface any document a user is technically allowed to read. In a private bank, oversharing in SharePoint can expose client files across desks. The sequence of controls is detailed in how Swiss banks govern Microsoft 365 Copilot data access step by step.

“FINMA does not audit your slide deck, it audits what your flows actually did last Tuesday. If the inventory, the test results and the approval trail don’t live in the same tenant as the automation, you are reconstructing evidence after the fact.” — Marco Chávez, Founder of Muze AI.

Who operates the automation after go-live?

Under FINMA Guidance 08/2024, ongoing monitoring is an expectation, not an option. Someone must watch flows and agents in production, re-test after changes and keep the documentation current. A bank has two choices: train an internal team, or contract managed automation, where the consultancy operates the system month to month.

Managed automation at Muze covers monitoring of production flows and Copilot Studio agents, and incident resolution. It also includes adjustments when a process or regulation changes, plus a monthly report. That report details hours consumed from the retainer, each activity performed, and the status of every system.

Everything stays inside the client’s own Microsoft 365 tenant. That matters for the outsourcing section of the guidance: the bank keeps ownership of data, logs and configuration. The operator works with delegated access rather than hosting the solution elsewhere.

Managed automation can also take over flows built by a different vendor. In Muze’s experience, this is a common entry point for Swiss institutions. They have ten or twenty Power Automate flows built during a pilot, no documentation, and no owner who can answer a supervisory question about them.

The fact is that FINMA expects continuous oversight. The judgment call is who carries it. Banks with an established Power Platform centre of excellence can do it internally. Banks without one usually find that two or three people cannot cover monitoring, testing and change requests on top of their existing roles.

What results are realistic for compliance automation in wealth management?

Realistic, documented outcomes in financial-services document automation sit in a clear range. Muze’s projects show an 85% reduction in KYC/AML processing time (insurance case), 75% fewer document-processing errors (fintech case), and 60% less time preparing compliance reports. These are results in comparable processes, not guarantees for a specific bank.

Across Muze’s implementations, the broader figures are 15-40% savings in direct operating costs and 80% fewer manual data-entry errors. They also include more than 3,000 hours per year saved in manual tasks. Each figure depends on starting volume and process maturity.

The 60% reduction in report-preparation time comes from regulatory reporting work. The clearest example is the automated SERNAPESCA reporting Muze built for one of the world’s five largest salmon farming companies. A different regulator applies, but the pattern is the same: structured extraction, validation, human sign-off and an auditable submission trail.

In Switzerland, Muze works with BlueYou, a sustainable-seafood company based in Zurich, and operates from both Chile and Switzerland. Muze’s Swiss work today is in operations and compliance automation, not core banking. A private bank should weigh that directly when scoping a first engagement.

Typical first use cases for a private bank, ranked by Muze’s judgment of evidence-to-effort ratio:

  1. Client onboarding document extraction and completeness checks (AI Builder plus Power Automate).
  2. Periodic KYC review packs assembled automatically from Dataverse and SharePoint.
  3. Regulatory and internal compliance reporting with validation rules before sign-off.
  4. A Copilot Studio agent answering internal policy questions, grounded only on approved directives.

On the model side, Muze is an OpenAI Select Partner and part of Anthropic’s Claude Partner Network. This matters when a use case needs a model outside the default Copilot stack and the bank must document which third-party model processes which data. A broader overview of use cases under both frameworks is in our 2026 guide to AI process automation under FINMA and the nFADP.

Which one fits your case?

The right choice depends on what is missing: a bank-wide policy, a large platform, or working, documented automation inside Microsoft 365. Three scenarios cover most Swiss private banks and wealth managers.

Scenario 1: No AI policy, no inventory, board asking questions. Start with a global advisory firm or your existing legal counsel to define governance roles and risk appetite. Automation without that foundation will fail the first supervisory question about accountability.

Scenario 2: Policy exists, but nothing in production, or pilots with no documentation. A specialized Microsoft-ecosystem consultancy fits here. The target is one documented process in production in 4 to 7 weeks, with the Dataverse inventory and test logs that Guidance 08/2024 implies. For Copilot rollouts specifically, see which consultancy deploys Microsoft 365 Copilot for Swiss private banks.

Scenario 3: Flows in production, but no one to operate them. Contract managed automation. The operator monitors, fixes, re-tests and reports monthly, inside your tenant, including flows another vendor originally built.

If the need is a core-banking replacement or a new data platform, a generalist integrator is the better fit. That work sits outside what Power Platform automation is designed for.

Frequently asked questions

Is FINMA Guidance 08/2024 legally binding for Swiss private banks?

No. The guidance, published on 18 December 2024, records supervisory observations and expectations rather than new binding rules. FINMA still applies existing governance, risk management and outsourcing duties to AI, so in practice banks are expected to meet it.

Can Power Automate and Copilot Studio process client data in Switzerland?

Yes. Power Platform environments can be provisioned in the Switzerland geography, backed by Azure Switzerland North. Dataverse data, flows and agents then reside in-country, which supports nFADP documentation and FINMA outsourcing records.

How long does it take to put a first compliance automation into production at a Swiss bank?

With a specialized Microsoft-ecosystem consultancy, the first process typically reaches production in 4 to 7 weeks. Bank-internal approvals, such as change boards and security reviews, can extend that timeline and should be planned from day one.

What documentation does FINMA expect for an AI-based KYC flow?

At minimum: a named owner, a risk classification, data-quality controls, test results with accuracy thresholds, and logged human approvals. Storing all of it in Dataverse inside the bank's tenant keeps the evidence next to the automation it describes.

Who monitors AI automations after go-live if our team lacks capacity?

A managed automation service can monitor flows and agents, resolve incidents and adjust them when processes change. Muze AI delivers this inside the client's own Microsoft 365 tenant, with a monthly report on hours consumed, activities and system status.

How does Microsoft 365 Copilot create compliance risk in a private bank?

Copilot surfaces any content a user already has permission to access, so oversharing in SharePoint or Teams can expose client files across desks. The fix is permission clean-up, sensitivity labels and restricted search scopes before broad rollout.

// REFERENCES

  1. FINMA AI Guidance: What Swiss Firms Must Know
  2. FINMA Guidance 08/2024: What It Means for Banks | BankingNewsAI
  3. Switzerland - AI Governance Guidelines (08/2024)

// REFERENCE GUIDE

How do Swiss companies automate processes with AI under FINMA guidance and the nFADP? (2026 guide)

// KEEP READING

Related articles

Does your operation have a process like this?

Free 5-minute diagnostic · we identify your 3 highest-ROI automatable processes.

Get my FREE diagnostic →
OpenAI Select Partner

Part of Anthropic's Claude Partner Network · Monday.com Authorized Partner in Chile · NVIDIA Inception program