From Custom Code to Natural Language: How AppZen Accelerates Financial Automation with Claude on AWS

Financial control gets increasingly challenging as an enterprise grows. In today’s market, it no longer suffices to wait through weeks-long deployment cycles for custom applications with custom code. AppZen saw that an agentic approach is the way to provide financial controllership at scale in minutes, with reduced cost and far fewer engineering hours.

AppZen’s AI-powered tools provide intelligent automation opportunities for expense auditing, accounts payable, and corporate card management. Using Anthropic's Claude models on Amazon Bedrock, AppZen has transformed months-long custom development cycles into self-service, natural language workflows that can be deployed and tested in minutes. AppZen passes this improved speed at scale to its customers, helping to improve compliance and expand financial controllership within increasingly sophisticated business finance environments.

As an enterprise grows, their financial operations become more complex, and financial controllership requires increasingly precise controls. But custom solutions are expensive and require long deployment cycles. AppZen is changing that equation with AI.

“When we were building custom models, we had so many different constraints,” said Asad Malik, AppZen Director of Customer Success.

Although effective, this approach had inherent limitations around speed and scale. Agentic AI presented a transformative automation approach to development, testing, and deployment that could adapt as quickly as today’s business landscape by offering solutions in minutes rather than weeks.

AppZen looked to Amazon Bedrock and Anthropic Claude to create a platform that would allow customers as well as AppZen’s own teams to define finance operations workflows using natural language while maintaining enterprise-grade security and governance.

Having already tried self-hosting open source models on top of Kubernetes, AppZen knew that a managed solution would allow them to plug and play. No longer would their development team have to worry about GPU management, provisioning servers, or autoscaling.

The company uses the open source framework LangChain to build agents in their AI Agent studio platform, and then brings the agents into Amazon Bedrock. Agents created through the AI Agent Studio are solving the most complex compliance and automation needs for hundreds of companies globally, processing millions of transactions.

Prior to implementing agentic workflows, AppZen already had a long history of utilizing AWS tools, meaning Amazon Bedrock was already closely integrated with their ecosystem.

“It has all the built-in regional capacity concern, data protection region concern and access control that we already use for the rest of our product. And then obviously hosting Anthropic models on top of that gives us access to the best models on the market,” said AppZen Director of Engineering TJ White.

Amazon Bedrock serves as AppZen’s foundation for model management, governance, security, and regional inference endpoints, and Anthropic Claude models powered the financial agents that needed to understand structured and unstructured financial documents.

Rather than rely on a single model for every workload, AppZen selects Anthropic Claude models strategically along with its own ZenLM models with consideration to performance, latency, and cost. Claude Haiku powers high-volume agent workflows where speed, efficiency, and multimodal data processing are critical. Claude Sonnet supports customer facing interactions, such as chatbots, which require more complex reasoning.

AppZen also uses Amazon Bedrock Guardrails to protect against its agent creation workflows from prompt injection attempts and harmful prompts. When customers talk with AppZen’s chatbot to create an agent inside AI Agent Studio, the conversation and end agent instructions are run through Bedrock Guardrails, helping to flag concerns early on. For example, Guardrails flagged a customer issue in which a member of the audit team had created an agent to auto-approve all of the audit team’s expense reports. These types of interventions help build customer confidence in the security and governance of these agentic solutions.

The agentic workflow also helps AppZen’s customers be empowered to deploy, test, and improve solutions quickly and with far less friction than before. Inside their AI Agent Studio, customers can even simulate agents before deployment.

“When a customer sees a problem with an agent or some agent that they've deployed, changing that configuration is just as easy as it was creating it because it's also just natural language. So instead of having to go through the whole build cycle again, which could be days or weeks or months, they just change the natural language in the prompts and click save and deploy it,” White said.

AppZen has continual collaboration with AWS as they implement and iterate on agentic tooling, and plans to expand their usage within Bedrock, as well as AgentCore, for more cost tracking and observability.

By replacing custom development cycles with configurable agentic workflows, AppZen can deliver sophisticated financial controls at scale, delivering more quickly for existing customers and expanding their ability to reach a broader customer base overall.

One of the company’s production agents automates China VAT receipt verification for Amazon, a process that previously required manual review from finance specialists for tens of thousands of expense reports each year. The new agent performs the same policy validation automatically, achieving approximately 98 percent automation while freeing employees to focus on higher-value work.

“The speed is incredible. To be able to actually launch an agent, launch a platform, launch a pipeline immediately to solve a business problem right now instead of needing to adjudicate this over weeks. We are able to create the SOP, simulate it on the fly, and then deploy it same day. And then if tomorrow we want to create another one, we create another version. No problem,” said Malik.

AppZen has adopted the same AI capabilities to transform customer success operations. Configuration analysis, rejection trend analysis, and business reviews can be generated more efficiently, with a projected hour reduction of 1,500 per month to 212.

The company continues working toward becoming a foundation for modern financial operations utilizing agentic AI to solve complex problems faster.

AppZen helps enterprises automate complex back-office financial processes. Combining agentic AI and large language models, AppZen offers automation for expense auditing, accounts payable, corporate card management, and invoice processing. By enabling organizations to replace manually coded workflows with agentic tools, customers are able to solve new financial control challenges effectively and efficiently. The organization focuses on reducing manual effort of mundane yet necessary tasks to give finance teams the opportunity to focus on high-value work.