--- title: "Claude for Financial Services: What Banks Get and Must Build - S-PRO" url: https://s-pro.io/blog/claude-for-financial-services lang: en description: "Claude for financial services gives banks agent templates, not a finished system. See what Anthropic ships, what you must build, and which path fits." updated: 2026-10-24T18:08:38+00:00 source: S-PRO (https://s-pro.io) --- [Back to insights](https://s-pro.io/blog) Article # Claude for Financial Services: What Banks Get and Must Build Author [Igor Izraylevych](https://s-pro.io/author/sproadmin) CEO and Co-founder of S-PRO Published October 24, 2026 Time to read 13 min Table of content - TL;DR - Table of Contents - What Anthropic actually ships for financial services - What a bank still has to build - Three ways to use Claude in a bank - Why connecting to client accounts is a different question - Governance: where Claude projects are won or lost - The regulatory frame for a Swiss bank - Use cases that suit Claude first - A practical sequence from pilot to production - Our perspective - How to choose - Where S-PRO fits - FAQ - Sources Share Claude for financial services gives a bank a strong starting kit, not a finished system: Anthropic supplies models, agent templates and governance primitives, and the bank still has to build the core integration, the governance, the evaluation and the human sign-off around them. This article compares the realistic ways to put Claude to work in a regulated institution, shows where each one fits, and lists the questions a CIO, COO or Head of Compliance should settle before a pilot starts. S-PRO s-pro.io Move Claude from pilot to production S-PRO implements Claude and builds AI agents for financial services, with solutions designed for FINMA, DORA, EU AI Act and data protection requirements. [Explore S-PRO’s Claude expertise](https://s-pro.io/claude-implementation-partner-financial-services) ## TL;DR - On 5 May 2026 Anthropic released [ten ready-to-run finance agent templates](https://www.anthropic.com/news/finance-agents), covering research and finance operations such as KYC screening, as plugins or Managed Agents. - Templates are a starting point. Connecting them to core systems, defining permissions, proving outputs and assigning sign-off remain the bank’s work. - The sanctioned path for a bank is to expose its own systems through MCP connectors under its own rules, rather than letting external agents act through client logins. - Three options are realistic: use Anthropic’s tools directly, build in-house, or work with a specialist implementation partner. The right one depends on your engineering depth and regulatory exposure. - Treat every agent as a privileged user and keep a human as the decision-maker on exceptions. ## Table of Contents - What a bank still has to build - Three ways to use Claude in a bank - The regulatory frame for a Swiss bank - Use cases that suit Claude first - Our perspective - How to choose - Where S-PRO fits - FAQ - Sources ## What Anthropic actually ships for financial services Anthropic’s [announcement of 5 May 2026](https://www.anthropic.com/news/finance-agents) describes ten ready-to-run agent templates. They split into research work, such as pitch builders and earnings reviewers, and finance operations work, such as KYC screeners and month-end closers. There are two ways to deploy them. You can install them as plugins in Claude Cowork or Claude Code, which suits individual analysts and small teams. Or you can run them as Managed Agents through cookbooks, which Anthropic lists as a public beta. Managed Agents add per-tool permissions and a centralized audit log, which are the controls a compliance function asks about first. That is a real contribution. A template shortens the path from an idea to a working demonstration, and the permission and logging features show that the vendor expects regulated buyers. What a template cannot know is how your bank is organized: which systems hold the client record, which policy defines a politically exposed person, or who may approve an exception. ## What a bank still has to build The gap between a demonstration and a production system is where most financial AI projects stall. The table below sets out what the vendor provides against what stays with the institution. | Layer | What Anthropic ships | What the bank must build | | --- | --- | --- | | Model and agent logic | Claude models, ten finance agent templates, plugins for Cowork and Code | Adaptation of each template to your products, policies and document types | | Core integration | Support for the Model Context Protocol and cookbook examples | Connectors to the core banking system, CRM, KYC and AML tools, document stores and market data, each under your security rules | | Governance | Per-tool permissions and a centralized audit log in Managed Agents | Role design, a minimal tool set per agent, approval of new tools, incident handling, records for the EU AI Act and your supervisor | | Evaluation | General model capability | Test sets built from your own cases, acceptance thresholds, regression checks after each model or prompt change, re-runnable decisions | | Human sign-off | The ability to keep a person in the loop | Defined decision rights: who reviews, who approves exceptions, what the agent may never do alone | None of the right-hand column is exotic. It is ordinary engineering and control work, but it takes most of the project time, and it decides whether your risk function lets the agent near a live process. ## Three ways to use Claude in a bank Most institutions end up choosing among three paths. They are not mutually exclusive, and many banks use one path for low-risk productivity and another for processes that touch clients. | Criterion | Anthropic tools used directly | Build in-house | Specialist implementation partner | | --- | --- | --- | --- | | Time to first demonstration | Shortest | Longest | Short | | Fit for individual analyst productivity | Strong | Possible | Possible | | Integration with core and legacy systems | Your team does it | Your team does it | Shared, with experience from prior projects | | Governance and audit documentation | Your team writes it | Your team writes it | Built into delivery | | Dependence on internal engineering capacity | High | Very high | Moderate | | Control over architecture and IP | Within vendor limits | Full | Defined in the contract | | Best suited to | Research desks, pilots, small teams | Banks with a large AI platform team | Banks moving from pilot to production in regulated workflows | Using the tools directly is the right start for an analyst team that wants faster memos and models. Building in-house suits the few institutions with a dedicated platform team and the appetite to maintain it. A partner makes sense when the use case is client-facing or compliance-relevant, the integration is complicated, and you need delivery plus documentation more than you need another internal project. ## Why connecting to client accounts is a different question A separate debate has opened around agents acting on a person’s behalf. As of October 2026, large US institutions have restricted external AI agents and connectors from client-facing account tasks, according to [Mindful Markets’ review of AI in trading](https://www.mindfulmarkets.ai/ai-in-trading-2026-from-access-anxiety-to-operating-model-reality-who-verifies-the-agents-and-who-sets-the-rules/). Reporting covers [JPMorgan’s position on AI agents](https://fortune.com/2026/10/15/jamie-dimon-ai-agents-meta-muse-jpmorgan) and [Schwab blocking agents that use client logins](https://thenextweb.com/news/schwab-blocks-ai-agents-meta-muse). Some platforms offer specific AI connectors, while others keep a closed in-app assistant to retain control. The lesson for a Swiss or European bank is practical. An agent that borrows a client’s credentials is outside your control. An agent that reaches your systems through connectors you built, with permissions you set, is inside it. That is the sanctioned path: MCP connectors operated under the bank’s own rules. The [Model Context Protocol](https://noah-news.com/open-standards-for-ai-governance-enable-safer-agentic-regtech-applications/) matters here because it sits between the model and sensitive systems. It lets an institution apply tool-level permissions, traceability and human-in-the-loop checks without hardwiring each integration. ## Governance: where Claude projects are won or lost ### Treat the agent as a privileged user Security-focused banks implement hard controls at the platform layer, not the prompt layer. An agent’s credentials should reach only a minimal, defined set of tool calls. Prompt-level instructions are soft controls and can be bypassed, as [practitioner analysis of the ten finance agents](https://buttondown.com/dodatathings/archive/the-diligence-wall-anthropic-shipped-10-finance/) points out. Design the agent’s access the way you would design a new employee’s, with least privilege and revocation. ### Make every output reconstructable An examiner who is not an engineer should be able to see which prompt, model version, tool calls and data sources produced a result, and in what order. The same source notes that model-risk reviewers want deterministic re-runs, and that standard LLM agents are non-deterministic. That is why teams pin model versions, settings and connector timestamps so that a decision can be re-tested a year later. The source frames this in terms of the US SR 11-7 framework, so check how your own supervisor’s model-risk expectations apply. ### Keep a person accountable Practitioners, including Forrester analysts quoted in [reporting on bank AI agents](https://longbridge.com/en/news/275919255.md), argue that AI should not be the system of record. Its value in trade accounting and KYC lies in extraction and comparison, while human analysts decide exceptions and sign off. Write this into the process: name the reviewer, the escalation path and the cases the agent must hand over. ## The regulatory frame for a Swiss bank A Swiss institution typically faces several regimes at once. Swiss data protection law applies to client data, FINMA supervises the institution, and the EU AI Act applies where you serve EU clients or place AI systems on the EU market. For DORA, the same applies to EU-facing operations. The EU AI Act is [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/LSU/?uri=CELEX%3A32024R1689), and its consolidated text is available on [EUR-Lex](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng). EIOPA notes that [AI-based creditworthiness assessments by banks are treated as high-risk](https://www.eiopa.europa.eu/publications/ai-act-and-its-impacts-european-financial-sector_en) use cases. A KYC document reader and a credit decision engine therefore do not carry the same obligations, so classify each agent before you build it. Germany’s BaFin has also published non-mandatory [guidance on ICT risks in the use of AI](https://www.bafin.de/SharedDocs/Downloads/EN/Anlage/dl_Anlage_orientierungshilfe_IKT_Risiken_bei_KI_en.pdf?__blob=publicationFile), dated 23 January 2026, which is a useful reference for the technology controls a supervisor may expect around AI systems. For FINMA specifics, rely on FINMA’s own publications and your compliance team’s interpretation. ## Use cases that suit Claude first The strongest early candidates share three traits: the work is document-heavy, the output can be checked by a person, and the agent does not move money or change a client record on its own. - **KYC and onboarding.** Extracting data from identity and corporate documents, comparing it with the application, and flagging gaps for an analyst. - **Adverse media and sanctions screening.** Reading and summarizing hits so that analysts triage faster. The screening system of record stays in place. - **Relationship manager copilots.** Preparing meeting notes, summarizing client history and drafting follow-ups for the RM to review. - **Document processing.** Pulling terms from contracts, prospectuses and statements into structured fields. - **Research and reporting.** Drafting memos and earnings summaries for analysts to verify. The weakest candidates are those that need autonomous action on client accounts. Even consumer-facing guidance, such as [a 2026 review of agentic trading apps](https://walnutinvest.com/resources/best-agentic-trading-apps), advises checking that a tool defaults to read-only and asks for explicit confirmation on every order. A bank should hold itself to at least that standard. ## A practical sequence from pilot to production 1. **Pick one process with a measurable baseline.** Record today’s handling time, error rate and escalation volume before the agent touches anything. 2. **Classify the use case.** Decide whether it falls under high-risk categories of the EU AI Act, and note FINMA and data protection implications. 3. **Define the tool set.** List the minimum systems the agent may call, and enforce it at the platform layer. 4. **Build the evaluation set.** Use real, anonymized cases with known correct outcomes, and set acceptance thresholds before testing. 5. **Wire in human review.** Specify who signs off, on what, and within what time. 6. **Log everything.** Capture prompt, model version, tool calls, data sources and order, so a result can be reconstructed. 7. **Run in shadow mode.** Compare agent output with analyst output without affecting decisions. 8. **Go live in a narrow scope, then widen.** Re-run the evaluation after every model or prompt change. ## Our perspective A common pattern is that the model is rarely the obstacle. The obstacle is that a template meets a core banking system that was never designed for an agent, a compliance team that rightly wants every output explained, and a process owner who has not decided who signs off. Our view is that an agent in a bank should be designed from the controls outward. Start with the permissions, the audit trail and the human decision points, then choose the model behavior that fits inside them. Hard limits belong in the platform, not in the prompt. An agent should assist an analyst who stays accountable, not replace the system of record. A sound approach is to start narrow. One KYC or document workflow, taken to production with documentation for the EU AI Act, FINMA and DORA, teaches an institution more than five pilots that never leave the sandbox. S-PRO is a Claude Partner Network partner, and our team includes Claude-certified engineers and architects. Anthropic’s tooling has clear limits, and the bank’s own engineering begins where they end. > *— Igor Izraylevych* ## How to choose Use these criteria when you weigh the three paths. - **Client exposure.** The closer the agent is to clients or to a regulatory decision, the more you need documented controls and delivery experience. - **Integration depth.** If the agent must read from a legacy core or several AML and KYC tools, budget most of the effort for connectors. - **Internal capacity.** Be honest about how many engineers and model-risk reviewers you can assign for the next 12 months. - **Evidence needs.** If your supervisor or internal audit will ask for reconstruction of decisions, plan the logging and re-run design on day one. - **Cost.** We do not publish prices, and Anthropic’s licensing and usage costs depend on the plan and volume, so request current figures from Anthropic and from any partner you consider. ## Where S-PRO fits S-PRO is a Zurich-based software company for banks, private banks, wealth managers, asset managers and fintechs. We take AI agents from pilot to production with audit trails, human-in-the-loop controls and documentation aligned with the EU AI Act, FINMA and DORA. Typical Claude use cases for financial institutions include agents for KYC and onboarding, document processing, adverse media and sanctions screening, and RM copilots. You can read more on our [page for Claude implementation for financial services](https://s-pro.io/claude-implementation-partner-financial-services). ## FAQ ### What is Claude for financial services? It is Anthropic’s offering of Claude models, agent templates and connectors for finance work. As of May 2026 it includes [ten ready-to-run agent templates](https://www.anthropic.com/news/finance-agents) for research and finance operations. It is a toolkit, and a bank still needs integration, governance and evaluation to use it in production. ### Can Claude be a financial advisor? We would not deploy it as one. It can prepare material, summarize client history and draft notes for a relationship manager, but the advice and the accountability stay with a licensed person. In regulated workflows, humans should remain the final decision-makers on exceptions and sign-offs. ### How much does Claude for financial services cost? Costs depend on the plan, usage volume and the integration work around it. We do not publish S-PRO prices, so ask Anthropic for its current licensing terms and request a scoped proposal from any implementation partner. Compare the total cost of the build and controls, not only the model usage. ### Can a Swiss bank use Claude under FINMA rules? It depends on the use case, the data involved and your outsourcing and risk framework, so confirm it with your compliance function and FINMA’s published requirements. Classify each agent first. If you serve EU clients, also check whether the EU AI Act applies to that system. ### Should agents log in with a client’s credentials? No. Large US institutions have restricted agents that act through client logins, and the safer route is connectors built and governed by the bank. Give each agent a minimal tool set and enforce it at the platform layer. ## Sources - [Agents for financial services, Anthropic](https://www.anthropic.com/news/finance-agents) - [AI in Trading 2026, Mindful Markets](https://www.mindfulmarkets.ai/ai-in-trading-2026-from-access-anxiety-to-operating-model-reality-who-verifies-the-agents-and-who-sets-the-rules/) - [Fortune, JPMorgan and AI agents](https://fortune.com/2026/10/15/jamie-dimon-ai-agents-meta-muse-jpmorgan) - [The Next Web, Schwab blocks AI agents](https://thenextweb.com/news/schwab-blocks-ai-agents-meta-muse) - [Open standards for AI governance and agentic RegTech](https://noah-news.com/open-standards-for-ai-governance-enable-safer-agentic-regtech-applications/) - [DoDataThings, The Diligence Wall](https://buttondown.com/dodatathings/archive/the-diligence-wall-anthropic-shipped-10-finance/) - [Bank AI agents in trade accounting and onboarding](https://longbridge.com/en/news/275919255.md) - [Regulation (EU) 2024/1689, EUR-Lex summary](https://eur-lex.europa.eu/legal-content/EN/LSU/?uri=CELEX%3A32024R1689) - [Regulation (EU) 2024/1689, consolidated text, EUR-Lex](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng) - [EIOPA, AI Act and the European financial sector](https://www.eiopa.europa.eu/publications/ai-act-and-its-impacts-european-financial-sector_en) - [BaFin, Guidance on ICT risks in the use of AI](https://www.bafin.de/SharedDocs/Downloads/EN/Anlage/dl_Anlage_orientierungshilfe_IKT_Risiken_bei_KI_en.pdf?__blob=publicationFile) - [Best agentic trading apps in 2026, Walnut](https://walnutinvest.com/resources/best-agentic-trading-apps) Subscribe Get technical insights, and industry news Thank you! Get ready! You will receive handpicked content right to your inbox. --- Link copied to clipboard [Get in touch](https://s-pro.io/get-in-touch)