Cutting expense claim review from hours to minutes using AWS with Visma

M2 by Visma has long been a trusted standard for managing travel and expense claims across Finland. The solution serves about 1,400 customer organizations and 1.1 million users, who file close to 2.7 million claims a year. Traditionally, validation has involved three roles. First, the traveler submits the claim. Second, a finance team member checks it for compliance. Third, a manager grants final approval. Each step requires careful attention to tax rules, receipt matching, value-added tax splits, and detailed descriptions of each cost. Manually reviewing complex multi-month claims could take up to 1 hour, and 9.58 percent of claims were returned for correction, extending the reimbursement time.

At roughly 8 minutes per review, manual checking consumed an estimated 360,000 hours a year—costing Visma’s customers over EUR €10 million. Some elements were hard to handle with traditional rules-based logic. Finnish tax regulations require detailed free-text descriptions for many types of expenses, and checking this contextual information has been beyond the reach of conventional automation. Using AI services on AWS, Visma could build a solution that interprets context, applies nuanced policy rules, and delivers the comprehensive automation that customers would benefit from most.

The software provider developed M2 Tarkka as a cloud-based microservice that brings intelligent automation to every stage of claim validation. At the heart of the solution is a library of approximately 50 rules, some being deterministic and others being powered by LLMs. These rules check everything from value-added tax splits to the contextual accuracy of free-text descriptions. Through a template-based editor, customers can adopt, tweak, or extend these rules, configuring each one as either a warning or a blocking error.

Validation tasks vary widely in complexity. Matching receipt totals to claim entries is straightforward, but interpreting a free-text justification for a client dinner under Finnish tax rules calls for far more nuanced reasoning. By adopting Amazon Bedrock, a service for building generative AI applications and agents, Visma gained access to foundation models that it could match to each validation task.

“We can even say, on every rule, ‘Instead of the default LLM, use this model,’” says Jani Vertanen, head of product, M2 Travel and Expense Solution at Visma. “We’re getting excellent results because Amazon Bedrock gives us the flexibility to choose the right model for each type of description evaluation.”

Visma also built an AI agent to help organizations upload their travel policies and users ask natural language questions about company-specific rules before submitting claims. The agent delivers relevant, accurate, and customized responses through Amazon Bedrock Knowledge Bases, a fully managed capability with in-built session context management and source attribution. Aiming to have 70–80 percent of customers self-serve, the team designed onboarding to be lightweight, typically taking 2–3 weeks.

Behind the scenes, M2 Tarkka runs as the company’s first microservice that is built entirely on AWS. The team used AWS Lambda, a service for running code without thinking about servers or clusters, to handle event-driven processing as claims move through the validation pipeline. Amazon Elastic Container Service (Amazon ECS), a fully managed container orchestration service, runs containerized workloads alongside AWS Fargate, a serverless, pay-as-you-go compute engine, removing the need to manage underlying infrastructure.

Since launching M2 Tarkka, the company has helped customers cut average claim lead times by 45 percent and reduce claim returns by 25 percent. And some organizations are seeing reductions as high as 50 percent. A complex multi-month travel claim that once took up to 1 hour to review can now pass validation in minutes. This frees finance teams from line-by-line checks and gives managers confidence that claims have already been measured against tax rules and company policies. Employees, in turn, receive their reimbursements far sooner. Visma expects M2 Tarkka to drive a 25 percent increase in total annual recurring revenue for M2 by Visma within 2 years, with the potential to reach 30 percent as adoption expands.

Looking ahead, the company is developing a solution to automate claim creation and approval, using agentic AI as a business and technology strategy rather than a bolted-on feature. “Some companies want people to spend as much time on their services as possible, but we’re the opposite,” says Vertanen. “Our customers still have our solution. The process is just so automated that they don’t even notice it.” For Visma, the future of expense management is one where claims are submitted, validated, and approved largely in the background, an invisible layer that returns hours to employees and keeps the focus on the work that matters.