Robinhood Transforms Financial Crimes Investigations Using Amazon Bedrock

Founded in 2013, Robinhood provides trading and investment services to millions of users. To help keep its platform secure, the company’s FinCrimes team actively monitors for suspicious activity that might indicate money laundering and other types of crimes. This monitoring plays a vital role in protecting users, upholding trust, and maintaining compliance with regulatory requirements and expectations.

While parts of the financial crimes investigation process were automated, much of the work remained manual. When an unusual transaction was detected, an alert would be generated, prompting an analyst to review it and conduct in-depth research to determine whether further action was needed. “Our analysts review mountains of data for every alert—a necessary step to protect customers and the integrity of our platform,” says Karthik Kenchaiah, senior engineering manager of FinCrimes engineering at Robinhood.

As traffic on Robinhood’s platform increases, so does the number of alerts, making it crucial to scale its investigative workflows quickly and efficiently. Seeing an opportunity to enhance these workflows, the company turned to generative AI. The goal was to build a responsible, scalable solution that accommodates increasing traffic patterns while remaining compliant with industry regulations.

Robinhood built its FinCrimes Agent, a multilayered system that combines large language model (LLM)-enhanced analysis and scalable AWS orchestration. The solution is optimized for high-volume alerts, with investigator oversight built into every stage for verification of accuracy. Using information from Robinhood databases and publicly available sources, the FinCrimes Agent generates concise summaries from structured and unstructured sources, highlighting key information from investigative records, long-form documents, and attachments. This streamlines annotation and reporting, saving analysts time.

When Robinhood detects potential criminal activity, an alert triggers the FinCrimes Agent to begin its workflows. The solution orchestrates multiple smaller agents that perform tasks individually. These agents communicate asynchronously through a task queue managed by Amazon Relational Database Service (Amazon RDS), a simple-to-manage relational database service. This orchestration helps to determine the sequence of agents, the tools they use, and how the results are validated.

The FinCrimes Agent runs self-contained workflows, each powered by specialized modules for summarization, classification, validation, and synthesizing external data. These modules use LLMs that run on Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models. The solution uses multiple LLMs through Anthropic’s Claude in Amazon Bedrock (both Claude Haiku and Claude Sonnet) and DeepSeek in Amazon Bedrock. Different models are chosen for different tasks to optimize both performance and cost efficiency.

Security was also a key consideration. By running models within its own virtual private cloud on Amazon Bedrock, Robinhood ensures that sensitive data stays within its control. “Using Amazon Bedrock, we feel comfortable knowing that the data isn’t leaving our control,” says Kenchaiah. “We have a lot of confidence in the security and reliability of the solution.”

To meet regulatory standards for explainability, each agent is paired with a validation agent that ensures outputs are factually accurate and coherent, and a hallucination check is applied to every response. “Until the verification agent is satisfied with the outputs, it doesn’t allow the workflow to proceed further,” says Nemil Timbadia, staff engineer at Robinhood. “This is a layer of safety that we’ve built so that our agents can reliably depend on the information.” Additionally, the solution includes a benchmarking framework that compares AI-generated outputs to resolved cases, which facilitates continuous monitoring and auditability. It also creates audit logs, so the company can provide verifiable and explainable outputs for governance teams—crucial for regulatory compliance.

With the FinCrimes Agent, Robinhood has achieved roughly a 20 percent cumulative efficiency gain in its investigative workflows. By delivering consistent, high-quality outputs, the solution has significantly reduced the time spent collecting and summarizing data. “The FinCrimes Agent has put all relevant information at our investigators’ fingertips, helping them to quickly identify illicit activity with increased focus and accuracy,” says Keith Kelley, head of financial crimes at Robinhood.

Currently, Robinhood is using the FinCrimes Agent to synthesize and summarize information for a subset of FinCrimes investigations, but the company plans to expand this to include additional use cases. As the solution evolves, the company remains focused on ensuring it aligns with regulatory expectations, helping set the stage for broader industry adoption of generative AI in investigations processes.

“Our goal wasn’t just to keep up with the industry; it was to lead. That meant building a solution that not only accelerated investigations but also upheld the highest standards for accuracy, explainability, and safety,” says Melissa Babin, senior director of investigations at Robinhood. “Every summary includes a human in the loop, ensuring that our investigators remain fully accountable for the final decision. We’re proud to have set a new benchmark in how generative AI can responsibly support financial crimes investigations.”

“We feel very confident about the solution that we’ve built,” says Kenchaiah. “Innovating in a space that involves compliance is hard. Fast-forward a few years, and I think that what we’re working on today with AWS will be the gold standard across the industry.”