Finance and insurance automation: inquiries, KYC, claims

Lookups and entry still span systems

When NTT DATA announced San-in Godo Bank’s RPA rollout, it described deposit inquiry work as heavy clerical work, because it repeats the same lookups in several systems across a vast number of accounts.

Financial institutions adopted RPA early. By 2019, Nippon Life had 200 robots covering 180 tasks, saving 70,000 hours of clerical work a year. Even so, the person in charge estimated that robots had been applied to only about 20% of the work. Reading handwritten documents, or deciding from what they say, cannot be automated by RPA that only repeats fixed steps.

Work	What stays manual	What already exists
Deposit inquiries from government agencies	Inquiries that arrive on paper	Roughly 80% or more of banks, shinkin banks, and credit cooperatives support online inquiry services (including those planning to)
Bar association inquiries and similar	Everything from reading the form to name matching, balance and transaction lookups, and writing the answer	—
Ongoing customer due diligence	Comparing returned forms with the customer record, and entering changes	Besides post, some institutions collect replies through apps
Insurance claims	Reading documents, looking up the policy, entering and checking the claim	Some large insurers have built their own AI OCR

Ongoing due diligence covers every customer

The FSA’s FAQ on its anti-money-laundering guidelines states that, in principle, every customer is subject to ongoing customer due diligence. The common way to confirm details is to send customers a letter and get their reply.

The Japanese Bankers Association says these requests draw many complaints and that reply rates are low. When a reply comes back, staff compare it with the customer record and enter what changed. When it doesn’t, they ask again. Both happen in volume.

Insurance claims are heavy on entry and checking

According to the Life Insurance Association of Japan, life insurers paid 8 million hospitalization benefit claims and 5.61 million surgery benefit claims in fiscal 2024.

Dai-ichi Life says that entering text from paper documents into its systems, and checking their content, was weighing down its operations. The forms in scope came in 700 types, at 17 million sheets a year.

Deciding on payment is a person’s job. The FSA’s supervisory guidelines for insurers expect claims that need advanced legal or medical judgment to be decided with input from the legal department or doctors, not by claims staff alone. What can be automated is the reading, lookup, entry, and checking that come before that decision.

Generative AI is already in use. The question is where, and with what data

In the Bank of Japan’s fiscal 2026 survey of 150 financial institutions, about 80% were using generative AI, and over 90% including those trialing it.

The FSA’s AI discussion paper notes institutions running open-source LLMs on on-premises servers, and others using proprietary generative AI services in closed environments over dedicated lines. Version 1.1 says that combining AI with RPA and similar tools to automate routine work may bring large cost savings and efficiency, while for uses that can significantly affect customers, such as loan approval documents, most institutions require a human in the loop.

Handing customer information to AI also raises these points:

• Japan’s Personal Information Protection Commission asks businesses entering personal data in prompts to a generative AI service to fully confirm, among other things, that the provider does not use that data for machine learning
• Sensitive information such as medical history may be handled, to ensure the proper operation of insurance and other financial business, only with the person’s consent and within what the work requires

What to check before automating

• Screens and computers: which computers operate core banking or policy screens, and with which permissions. Check your contracts with the system’s vendor too
• Whether existing services are enough: work that an online inquiry service or a shared industry system already handles should stay there
• What AI may see: separate what may go to outside AI models from what only models you host may handle
• Where people decide: staff identify the person, rate the risk, decide what to disclose, and decide on payment
• Records: whether you can show auditors who approved what was entered

What Kitewell does

Descarty’s Kitewell is integration and automation middleware that connects systems through APIs, files, databases, and screens.

• Operates screens: AI reads and operates web systems and desktop apps with no API. Operations that worked are remembered and repeated the same way on later runs
• Reads documents and drafts judgments: questions to AI models and judgments become steps in the workflow. You choose the model for each step, including in-house and local models
• People decide where it matters: approval can be set on any step. Requests reach Slack, Microsoft Teams, and email, and sending one back restarts the run from the step you choose
• Passwords stay out of AI’s hands: system passwords are referenced by name and typed in without showing the value to the AI
• Everything on record: each run keeps its logs, screenshots, and approval decisions
• Alerts when it stops: failures are sent to Slack, Microsoft Teams, email, and more

The computers that run workflows can be VMs or dedicated PCs in your own cloud, on your premises, or in a closed network. Licensing, team sync, and updates can all stay in your environment, and with in-house or local AI models nothing is sent outside. AI models that can read and operate a screen are limited, though, and small local models usually cannot operate one. The PoC settles which model to use in a closed network.

You can start from one piece of work, such as taking in an application through preparing the review and registering it in your system. A demo shows how it runs, a free connection test confirms Kitewell reaches your systems, and a one-month PoC measures the result against success criteria agreed up front, before you license it for production. See From evaluation to production for details.

Sources

• NTT DATA: San-in Godo Bank begins full operation of the RPA solution WinActor (March 13, 2018; in Japanese)
• Newswitch: Nippon Life and its “robot HR team” (March 8, 2019; in Japanese)
• Digital Agency: Expanding online deposit inquiries (February 14, 2025; in Japanese)
• FSA: FAQ on the Guidelines for AML/CFT (March 31, 2026; in Japanese)
• Japanese Bankers Association: Public outreach on key AML/CFT measures (November 26, 2021; in Japanese)
• Life Insurance Association of Japan: Life insurance trends, 2025 edition (in Japanese)
• ABBYY: Dai-ichi Life customer story (in Japanese)
• FSA: Comprehensive Guidelines for Supervision of Insurance Companies, II-4 (in Japanese)
• Bank of Japan: Use of generative AI and risk management at financial institutions, fiscal 2026 survey (August 2026; in Japanese)
• FSA: AI Discussion Paper, version 1.0, summary (March 4, 2025; in Japanese) and version 1.1 (March 2026; in Japanese)
• Personal Information Protection Commission: Alert on the use of generative AI services (June 2, 2023; in Japanese)
• Personal Information Protection Commission and FSA: Guidelines on personal information protection in the financial sector (March 2024; in Japanese)