Less routine work in claims, applications and customer service at banks and insurers

AI is already in daily use in this industry. Whether it pays off is less clear. In the 2025 study by Cofinpro with VÖB-Service, 78% of the 317 experts from German banks surveyed reported productive AI use, up from 52% in 2023, yet 70% of AI initiatives run in isolation. In the IAB establishment panel 2025, published in 2026, 50% of establishments in finance and insurance use generative AI, second only to information and communication at 59%.

In Visa's 2026 survey of European banking decision-makers, cost reduction is the top goal for 77% of German respondents, and Visa notes that many banks have started AI projects while the economic benefit often remains limited. When it works, the change is concrete: claims and applications arrive in the case system already sorted and filled in, staff find the right guideline or product condition with the source shown instead of searching folders, and customers get a precise reply sooner. People still make every decision that affects a customer.

We concentrate on work where an assistant or agent prepares and a person decides. Knowledge management, document processing and customer service each have their own page under Use cases, linked at the end of this page.

Not everything needs to be built. Voice bots for inbound calls are a specialist product: Württembergische reports one third shorter waiting times with an AI agent that classifies and routes calls, and an unnamed German health insurer reports a 71.4% automation rate for sick-leave calls, both reported by the vendor Parloa in 2025. Ask your core banking, policy or claims system vendor what it already offers. We build where your documents, rules and case systems have to come together, and we test what you buy.

• Knowledge assistants for claims handlers, credit staff and advisers: answers from conditions of insurance, credit guidelines, product sheets and internal circulars, with the source paragraph shown and each user's permissions respected. In the 2025 Cofinpro and VÖB-Service study, 59% named intelligent information retrieval as an AI use.
• Intake of claims and applications: the agent reads incoming mail, forms and attachments, flags duplicates, extracts policy or customer number, dates and amounts, and opens a pre-filled case for a handler. In Sollers Consulting's 2025 survey of 35 insurers in several countries, about 64% use AI in claims and 69% for data extraction.
• Document checks that prepare credit and underwriting work: completeness of the file, figures from payslips, annual accounts or valuation reports compared with what the customer stated, and deviations listed for the person who decides. In Cofinpro's 2025 survey of lending heads, document analysis came first among the use cases with most potential, at 44%. The check only prepares. It never scores or decides.
• Customer service drafts: replies to emails and portal messages drafted from approved texts and the customer's case, checked and sent by a person. In Bitkom's 2026 consumer survey, 43% would let AI answer questions about existing cover, but only 15% think AI recommends better than a human.
• Testing of AI systems you already run or buy: acceptance tests and adversarial tests on your own cases, so you can show supervisors and internal audit how the system behaves.

BaFin asks financial firms to test their AI systems. Its guidance on ICT risks when using AI systems in financial firms, published on 18 December 2025, is not binding, applies to CRR institutions and Solvency II insurers and sits under DORA. It covers the full lifecycle of an AI system, ICT third-party risk, human oversight for AI self-service assistants, adversarial testing, stress tests under shifted data and involving the system's manufacturer in tests. The Bundesnetzagentur confirmed in 2026 that BaFin remains the competent authority for financial firms under the AI Act.

The AI Act lists credit scoring and risk assessment and pricing in life and health insurance among high-risk uses. A deployer of such a system must, among other duties, have trained people oversee it, use it as the provider instructs, monitor it and keep its logs. After the Digital Omnibus these obligations apply from 2 December 2027. We do not build systems that score creditworthiness or set life or health insurance prices, and we do not automate such decisions. Our systems prepare the file for the person who decides.

DORA, the EU regulation on digital operational resilience, applies when we work for a bank or insurer. Its rules on ICT third-party risk and your outsourcing rules apply to our work, and we agree in the contract how they are met, together with your outsourcing management and information security teams.

Banking secrecy and data protection decide where data may be processed, which is why we build in your own cloud account or on your own servers. The legal assessment stays with you, advised by your data protection officer, compliance and lawyers. This page is not legal advice.

• Credit decisions, cover decisions, settlement amounts and prices stay with your staff.
• A person checks and sends every message to a customer. Where customers talk to an assistant directly, they are told it is AI, as Article 50 of the AI Act requires from 2 August 2026.
• When a customer objects to an AI-supported pre-check, a person reviews the case, and the log shows which documents, fields and rules the preparation used.

The system runs where your data may be processed: in your own cloud account in an EU region, or on your own servers with an open-weight model, which is a model whose weights you can run on your own hardware. All model calls go through one LLM gateway, a single controlled access point with keys, budgets and logs, so you can see which data went to which model.

We connect the systems the work already happens in: inbound mail in Outlook, your document management system such as d.velop or ELO, SharePoint and file shares for guidelines and circulars, the CRM, for example Salesforce, and the claims, policy or credit workflow system through the interfaces it offers. Permissions come from those systems, so a handler only sees what they may open. If a central IT service provider runs your core systems, we work through the interfaces it provides.

Agents act through approved tools with their own permissions, and every step that writes to a case or sends a message waits for a person's approval. Logs stay in your own store, with retention agreed with your data protection officer and works council.

Before go-live we run an acceptance test: a fixed set of your own past cases with known correct results, against a pass mark agreed with your business, IT and risk owners. For intake we measure accuracy per field, for knowledge assistants whether the answer and the cited source are correct, and for drafts how often a handler can send them with no or minor edits.

Following the topics in BaFin's guidance, the test set includes adversarial inputs, such as instructions hidden in an attachment, and shifted data, such as a new form layout or a new product generation. Where you buy a model or product, we use its manufacturer's documentation and, where possible, involve the manufacturer in the tests. After go-live the same tests run after every change to the model, prompts or data, and we compare time per case with the baseline. Test results and logs are kept so internal audit can trace them.

Start with one process that has many similar cases, documents you already hold and a clear person who approves, such as claims mail intake or a knowledge assistant on credit guidelines. In a free 30-minute first call with an engineer, which needs no access to your systems or data, we work out which process fits, where it should run and what the acceptance test should measure. Each step after that has a fixed price, agreed in writing after the first call.

If your teams already use AI tools, the free AI usage review gives you one page on spend and quality, built from exports you choose to send.