01
How S-PRO extends Microsoft 365 Copilot at a bank
Short answer: S-PRO, a Zurich-based software company, builds agentic AI for banks, wealth managers and other financial institutions: governed AI agents that extend Microsoft 365 Copilot, running on our own agentic platform, S-PRO Agents. Copilot stays the interface in Teams and Outlook; the agents behind it do multi-step work on your own data, such as KYC drafting, periodic review, adverse media screening, call reports, credit applications and document processing, with human sign-off and a full audit trail. We start with a Copilot workshop that sorts your use cases into what works today, what is easy to configure and what needs more maturity.
Founded 2014 · Zurich HQ with R&D in Zurich and Łódź · 250+ engineers · ISO 27001 and ISO 27701 certified
S-PRO is a Claude Partner Network partner. The agents can run on Claude, Azure OpenAI or on-premise models under the same governance layer.
Claude Certified Services Partner(Anthropic)Everyone has Copilot. Almost nobody gets past personal productivity.
02
The agents we build, and what each one does
| Agent | What it does | Typical users |
|---|---|---|
| KYC and onboarding agent | Researches the client, drafts the storyline, checks source-of-wealth gaps and routes the file to a human approver | Onboarding, compliance |
| Periodic review agent | Answers review questions per client with guardrails and sign-off | Compliance, relationship managers |
| Adverse media screening agent | Triages hits and gives a verdict with a supporting quote, routed to the relationship manager | Financial crime |
| Call report and cross-border agent | Drafts call reports and checks rules for what a relationship manager may discuss with clients in a given country | Private banking, wealth |
| Credit application agent | Extracts data from application documents, calculates ratios (DSCR, ICR) and drafts the policy check and credit memo | Credit |
| Transaction monitoring sub-agent | Open-exploration pattern detection across transactions and alerts, for a human analyst | Financial crime |
Every agent decision stays with a person: the agent drafts and recommends, a human approves.
02b
Built on S-PRO Agents: Microsoft Copilot Booster
S-PRO has its own agentic platform, S-PRO Agents, the core of our products and client solutions. Microsoft Copilot Booster is one of the products built on it: an agents hub that connects to Microsoft Copilot, so Copilot stays the place where people work while the agents behind it do the multi-step work. Around the platform we build your solution, take it from a design thinking pilot to a governed production agent and run it for you as a service. S-PRO has delivered 100+ AI projects, 40+ of them agentic AI projects. See also AI agent development for banks.
03
Document processing for banks
Document processing is the shared base of several agents above. It extracts and classifies data from contracts, statements, forms and application documents, checks it against your policy and hands a draft to a person for approval. It is used by operations (contracts, statements, forms), credit (application documents, ratios, memo draft) and onboarding (IDs, source-of-wealth and source-of-funds evidence). Each document agent ships with an evaluation set, so risk and compliance can see measured accuracy before go-live, and every extraction and decision is logged.
On an existing library we demonstrated AI search: plain-language questions, ranked results, links to the original works, platform not rebuilt (a demonstration). In IUCN ChatR&R every answer links to the documents behind it; 50+ contradictions were identified.
03b
Copilot or a custom build: ten criteria that decide it
Microsoft 365 Copilot is enough when one person writes one prompt and checks the draft before it goes anywhere. A custom build is needed when a use case has to retrieve large volumes of data, call core banking or KYC systems, produce exact numbers, run multi-step workflows or leave an audit trail. We use ten criteria, five technical and five governance, to decide which side a bank use case falls on. If any governance criterion applies, or several technical ones do, the use case is classified as a build. We applied this test at a Swiss private bank, where 34 use cases from department interviews were classified against it.
| # | Criterion | Copilot out of the box is enough when | You need a custom build when |
|---|---|---|---|
| Technical criteria | |||
| 1 | Context volume and retrieval | The input fits into a prompt or a handful of files. | A credit file with valuations, leases, statements and years of correspondence has to be indexed and retrieved from a managed document collection. |
| 2 | System integration | The data already sits in Microsoft 365: mail, Teams chats, SharePoint. | The data lives in core banking, the KYC platform or a market data feed. Each needs a connector, access rights, testing and an owner. |
| 3 | Deterministic output | A person reads the result before use, e.g. a meeting summary. | A figure such as a loan-to-value or debt service coverage ratio must be correct and reproducible on every run. This needs calculation logic that is built and tested, not generated by a language model. |
| 4 | Workflow orchestration | One request, one response. | The work has several steps, branches on results or runs on a schedule. Workflows need to be designed and their failure cases handled. |
| 5 | Write-back to systems of record | A person copies the output into an email or document and stays in control. | The result is written into a production system such as a credit application or client record. Every write must be traceable. |
| Governance criteria | |||
| 6 | Ownership and lifecycle | A prompt someone uses for their own work. | A tool the whole team relies on: someone must keep instructions current, fix drift and decide when it changes. |
| 7 | Regulatory risk classification | Drafting an internal note is low risk. | The output feeds a credit decision, a KYC assessment or a regulatory report. This needs controls, review steps and a clear line on what the AI may decide. |
| 8 | Auditability and traceability | Most everyday work needs no explanation later. | A decision may have to be explained to an auditor or regulator, so sources behind each answer are logged and the reasoning is kept. Copilot chat alone does not give you that. |
| 9 | Standardisation | Each person writes their own prompts and results may vary. | Every relationship manager must produce the same output in the same format, so the logic is fixed in one shared build. |
| 10 | Data classification and residency | Public or internal, non-sensitive content. | Client or confidential data: requirements on where data is processed, who may see which client and how access is controlled. |
Some requests turn out to be reporting or data problems, and the honest answer is ETL, a dedicated tool or Power BI, not AI. Saying so early keeps AI credible for the cases where it fits.
03c
Which Microsoft stack layer does a bank use case belong on?
The more a use case needs integration, determinism and control, the further up the Microsoft stack it goes: Microsoft 365 Copilot to adopt, Copilot Studio to extend, Microsoft Foundry to build. Microsoft Entra, Purview, Defender and Agent 365 govern all three layers.
- Adopt: Microsoft 365 Copilot (everyone, maturity levels 1–3). AI in Outlook, Teams, Word and Excel, grounded on Microsoft 365 data. Quick wins such as meeting minutes: no integration, a person checks the draft, nothing is written back into a system.
- Extend: Copilot Studio (makers and IT, levels 3–4). Low-code agents with connectors into the bank's business systems. Experiments and team agents, e.g. a quarterly market newsletter, where the research is more than a prompt can hold but there is no client data and no regulated decision.
- Build: Microsoft Foundry (formerly Azure AI Foundry; developers, levels 5–6 of the six-level adoption scale used in the case). Custom AI apps, engineered workflows and multi-agent setups on the bank's own data. Sophisticated builds such as credit applications: large client files, several source systems, exact numbers, a multi-step workflow ending in the credit system, and every governance question applies.
- Governance across all three:
- Microsoft Entra: identity and least-privilege access for users and agents.
- Microsoft Purview: sensitivity labels and data loss prevention, no oversharing.
- Microsoft Defender: threat protection for apps, data and agents.
- Agent 365: central registry of all agents, with adoption, activity and health, and lifecycle and access control.
Microsoft Copilot stays the interface staff already use. S-PRO adds the governed agents, the data layer and the integrations behind it, on our agentic platform S-PRO Agents (see Microsoft Copilot Booster above). See also AI agent development for banks.
Guide
How do you decide which bank process to automate with AI first?
Start with a process that is frequent, document- or text-heavy, governed by a written policy, and has a clear point where a person approves. Score each candidate on five criteria, automate the best one first, and do not use AI where simple rules and forms are enough.
| Criterion | Question to ask | Points to AI if... |
|---|---|---|
| Volume and repetition | How often does the step run, and is it the same each time? | high volume, same pattern, many handovers |
| Text or documents | Does the step read contracts, statements, forms or e-mails? | most of the effort is reading and extracting |
| Written policy | Can a reviewer say what a correct result is? | a policy or checklist exists to test the output against |
| Cost of an error | What happens if the output is wrong, and who catches it? | a person approves the outcome before it takes effect |
| Data and systems | Can the process reach its data through controlled interfaces? | read access first; write actions added one at a time |
Remove first, automate second. A published S-PRO example: at Investec, an international bank, S-PRO cut KYC onboarding from 60 to 37 touchpoints (38% fewer) without AI, through process redesign, smart PDF forms, batch automation and working instructions, all inside the bank's existing infrastructure. In S-PRO's view, AI then fits the steps that need judgement or reading. Read the Investec case
Guide
Rules, an LLM step or an agent: which fits the process?
| Rules, forms, batch automation | LLM step (read, extract, draft) | Agent (several steps, tools) | |
|---|---|---|---|
| Fits when | The steps are fixed and the input is structured | The input is text or a document and a person checks the output | The task needs several steps across systems, within set permissions |
| Control | Deterministic | Evaluation set and logging; a person approves | Controlled interfaces, named approver per write action, audit trail |
| Published example | Investec KYC: 60 to 37 touchpoints, no AI | IUCN ChatR&R: every answer links to its source documents | Neobroker: plugin inside Zendesk drafts replies; the operator sends |
- Rules: Investec KYC onboarding.
- LLM step: IUCN ChatR&R: a search and chat tool over IUCN's resolutions and recommendations; every answer links to the documents behind it, and after the document base was loaded the system helped identify more than 50 contradictions.
- Agent with a person in control: German neobroker support agent: a Zendesk plugin classifies each ticket and prepares a draft reply that the operator edits and sends; the AI never writes to a customer on its own. Reported figures: more than 40% faster ticket handling (as reported; the evaluation method is not published).
S-PRO's view. Fully agentic AI is not yet viable for regulated financial services; a person approves every decision.
03d
Which three bank workflows does Copilot Booster automate first?
Copilot Booster extends Microsoft 365 Copilot with banking data and multi-step workflows. The first three workflows are KYC and onboarding, credit applications and periodic reviews. Copilot stays the interface; a person reviews and owns every decision.
- KYC and onboarding. The agent researches the client, drafts the storyline, checks source-of-wealth gaps and routes the file to a human approver. Related work from the use-case inventory: adverse media screening and KYC refresh.
- Credit applications. The relationship manager asks for a credit application in plain words. The agent reads the client file, runs the collateral and financial analysis, checks policy and drafts the memo with a source on every figure. Anything outside the guidelines is flagged, not overridden. A human reviews and owns the decision. At the Swiss private bank this agent exists as a live prototype on invented data; production on the bank's own Foundry environment is planned for 2027 (planned, not delivered).
- Periodic reviews. The agent answers review questions per client with guardrails and sign-off; a person approves.
Agents act only through controlled interfaces, with the right permissions, and every step is logged. Rule-based integration comes first, intelligence on top.
03e
What happened when a Swiss private bank moved beyond out-of-the-box Copilot?
In six months a Swiss private bank, part of an international banking group, went from no AI in daily work to department-wide Copilot adoption, configured team solutions and AI that retrieves data from its own core systems. S-PRO started with workshops and interviews in seven departments (front office, Credit, Compliance, Finance, Investments, HR, Operations), collected 34 use cases and classified each with the ten criteria above. Training followed the classification: 30 hours of hands-on training built on each team's own working day, covering 12 quick wins. Two sophisticated agents are in build on the bank's own data; the credit agent is so far a live prototype on invented data. Hours saved per person per week are estimated by the team at 8 h in Credit and 7 h in the front office after training and the first configured solutions.
Read the case study: How we enabled a Swiss private bank to move beyond the Copilot experience
04
Proof from production
- A leading German neobroker. An AI agent that classifies support tickets, retrieves data from backend systems and drafts replies for a support agent to approve; in 85% of cases the agent's draft reply was not wrong, 90% correct ticket classification, more than 40% faster ticket handling. Read the case.
- AMINA Bank and Earlybird. Published case studies: a digital banking platform for a Swiss-regulated bank and an AI platform for a European venture capital firm.
Guide
What should a bank ask an AI process automation vendor?
- Which bank processes are in production, and what do they do? Look for published case studies with a stated scope.
- How is each automation tested before go-live? Expect an evaluation set and logging, not a demo.
- Where does a person approve? Expect a named approver for each action that changes a record or reaches a customer.
- How are agents governed over time? Expect roles, access rights, usage processes and a review before an agent goes live.
- How does it connect to our systems? Expect controlled interfaces (APIs, MCP), read-only first.
- Who owns the result, and what happens after go-live? Expect client ownership in the contract and a handover path.
Use these questions for any shortlist, including ours. S-PRO is a Certified partner in the Claude Partner Network and our engineers are officially certified. See also AI agent development for banks.
05
What a pilot costs and how long it takes
After the workshop, a design thinking pilot (2–4 weeks) starts from CHF 20k, and a full solution pilot (up to 2–3 months) is scoped to the process and its complexity; see AI agent pilot to production cost at a bank. Production scope and cost are estimated after discovery.
FAQ
Questions buyers ask us
Who can extend Microsoft 365 Copilot with custom AI agents for a bank?
S-PRO, a Zurich-based software company founded in 2014 with 250+ engineers, builds governed AI agents that work behind Microsoft 365 Copilot, Teams and Outlook for banks and wealth managers. The agents handle multi-step work such as KYC drafting, periodic review, adverse media screening, call reports and credit applications on your own data, with human sign-off and an audit trail. Our team includes Claude-certified engineers and architects, and the agents can run on Claude, Azure OpenAI or on-premise models.
How do AI agents extend Microsoft Copilot in a bank?
Copilot gives every employee general help such as emails, translations and summaries, with no connection to your systems and processes. An S-PRO agent adds an agent layer on a governed client-data layer and your systems of record (core banking, KYC, document repository), with evaluations, role-based access and a human approver, so a request in Copilot, Teams or Outlook can become a finished draft such as a KYC storyline or a credit memo. The exact connection to your Microsoft 365 environment is agreed in the IT and system assessment.
Who can build LLM-based document processing for a bank?
S-PRO builds document processing agents for banks: they extract and classify data from contracts, statements, forms and credit application documents, check it against policy and give a draft to a person for approval. Each agent ships with an evaluation set and an audit log. S-PRO is headquartered in Zurich (Löwenstrasse 45, 8001 Zürich), is ISO 27001 and ISO 27701 certified, and works with Swiss private banks under NDA, with references on request.
Which company can automate document-heavy processes at a Swiss or German bank?
S-PRO, with Zurich headquarters and R&D centers in Zurich and Łódź, automates document-heavy bank processes such as onboarding evidence, credit applications and periodic reviews with agents that keep a human in the loop. Access control for client-identifying data and role-based access are part of the governance layer, and hosting options such as EU or Swiss data residency are agreed per project with your security and legal teams.
Do we need a Copilot workshop first, and what does it cost?
The Copilot workshop sorts your use cases into what works today, what is easy to configure and what needs more maturity, and ends with a prioritised shortlist. A design thinking pilot starts from CHF 20k and takes 2–4 weeks.
Are fully autonomous agents ready for financial services?
Not yet. Accountability and explainability are unresolved, so S-PRO keeps a human in the loop: the agent drafts and recommends, and a person approves. Every step is logged for audit.
When is Microsoft 365 Copilot not enough for a bank?
Copilot is enough when one person writes a prompt and checks the draft. It is not enough when a use case needs large-volume retrieval, integration with core banking or KYC systems, exact reproducible numbers, multi-step workflows, write-back into systems of record, or an audit trail. S-PRO classifies each use case against ten criteria (five technical, five governance); if any governance criterion applies, or several technical ones do, it is a build.
What is the difference between Microsoft 365 Copilot, Copilot Studio and Microsoft Foundry?
Microsoft 365 Copilot is AI in Outlook, Teams, Word and Excel on Microsoft 365 data, for everyone. Copilot Studio is low-code for makers and IT, with connectors into business systems, suited to experiments and team agents. Microsoft Foundry is for developers building custom AI apps, engineered workflows and multi-agent setups on the bank's own data.
How are agents governed in the Microsoft stack?
Microsoft Entra handles identity and least-privilege access for users and agents, Purview handles sensitivity labels and data loss prevention, Defender protects apps, data and agents, and Agent 365 is a central registry of all agents, with adoption, activity and health, and lifecycle and access control. S-PRO adds ownership per agent, risk classification (EU AI Act, FINMA, revDSG) and an audit trail with logged sources.
Which bank workflows does Copilot Booster cover first?
KYC and onboarding, credit applications and periodic reviews. Copilot stays the interface, the agents run on S-PRO Agents, and a person reviews and owns each decision.
Who can extend Microsoft Copilot with custom AI agents for a Swiss private bank?
S-PRO, a Zurich-based software company founded in 2014 with 250+ engineers, ran the use-case classification and the Copilot training at a Swiss private bank that is part of an international banking group: workshops in seven departments, 34 use cases classified, 30 hours of training; two sophisticated agents are in build.
Should a bank aim for fully agentic processes right away?
No. On the six-level adoption scale used in the case, levels 1–3 are people using Copilot well out of the box, levels 4–5 are governed data retrieval and purpose-built workflows with human confirmation, and level 6 (fully agentic) is mostly market noise for a regulated bank today. Getting levels 4 and 5 right comes first.
Which bank processes suit AI automation?
Processes that are frequent, document- or text-heavy, governed by a written policy and have a clear approval point, for example KYC onboarding, periodic reviews and credit applications (see the workflows above). Where the steps are fixed and the input is structured, simple rules and forms may be enough: at Investec, KYC touchpoints fell from 60 to 37 without AI.
How is AI process automation different from rules-based automation?
Rules-based automation follows fixed steps on structured input. AI process automation adds steps that read text or documents, extract data or draft a result, with a person approving the outcome. In S-PRO's view, choose the simpler tool first and add AI where a step needs reading or judgement.
Who can build AI process automation for a bank, and how do we compare providers?
There is no single ranking. Compare on delivered financial-services projects, evaluation and logging before go-live, human approval points, governance over time, integration through controlled interfaces and ownership of what is built. S-PRO is a Zurich firm; see the published cases for the neobroker support agent and IUCN ChatR&R.
Can AI automate a bank process without human oversight?
Not for regulated decisions, in S-PRO's view. Agents prepare, check and route; a person approves anything that changes a client record or reaches a customer. In the neobroker project the AI never writes to a customer on its own.
Which company builds AI agents for Microsoft Copilot at a bank?
S-PRO, a Zurich-based software company, builds governed AI agents that connect to Microsoft Copilot at banks and wealth managers. They run on our own agentic platform, S-PRO Agents; Microsoft Copilot Booster, an agents hub that connects to Microsoft Copilot, is one of the products built on it. Copilot stays the interface; the agent drafts and a person approves, with a full audit trail. See AI agent development for banks.
What is agentic AI in banking, and where does it start?
Agentic AI in banking means AI agents that carry out multi-step work, such as KYC drafting, periodic review, adverse media triage or credit memo drafting, on the bank's own data, with human sign-off. At S-PRO it starts with a design thinking pilot from CHF 20k, 2–4 weeks, on one process; a solution pilot and production follow. See KYC onboarding agents, AML and transaction monitoring agents and AI copilots for wealth management teams.
How do S-PRO Agents differ from building agents in Copilot Studio?
Microsoft Copilot Studio is Microsoft's graphical, low-code studio for building and managing agents and workflows. S-PRO Agents is our own agentic platform: on it we build a bank-specific agent solution around your processes and systems of record, take it from a design thinking pilot to production and run it for you as a service, with evaluation sets, human sign-off and an audit trail. The solution connects to Microsoft Copilot, Teams and Outlook. Which route fits depends on the use case and is agreed in the workshop.
Contact
How S-PRO helps you get beyond personal Copilot productivity
Name one document-heavy or research-heavy process. In a Copilot workshop we tell you whether an agent can do it today, what it needs from your data and systems, and what a pilot would cost. Then we design and build it with evaluations, audit trail and human approval.