Trovata Launches Treasury Agents on AgentCore: A Report in Under a Minute Instead of Days

Picture a finance professional who opens ten bank portals every morning, copies the balances into a spreadsheet, and consolidates them by hand. By lunchtime, they finally know how much cash the company has. This is exactly the pain Trovata addresses. It is a cash management platform with direct connections to more than 10,000 banks, and it is now handing the routine work over to autonomous agents on Amazon Bedrock and Amazon Bedrock AgentCore.

What the problem is and how it works

Treasury teams face three chronic problems. Data is scattered across bank portals. Reporting and reconciliation consume hours, and sometimes days, of skilled people's time. And legacy treasury management systems (TMS) were not designed for the cloud and AI.

In 2023, Trovata released a chat that answered questions about data. A user would ask, "What is our net cash position across all accounts as of yesterday?" and get an answer. Now the company has gone further: the agents don't wait for a question, they work continuously. They are launched on a schedule, by a trigger, or by a rule. One agent monitors account activity, another explains why actuals diverged from the forecast, and a third looks for suspicious transactions and odd balance jumps.

The key detail: the AI is built into the system itself and tied to real treasury objects, rather than sitting on top of it as a generic chatbot. The model is responsible for understanding the request and generating text. AgentCore handles the orchestration of multi-step scenarios.

How the agents' freedom was limited

In finance, the main question is not "what can the agent do" but "what can it not do." Trovata describes four safeguards.

First: the agent gets data only through the company's own vetted API. It has no direct database access, so it cannot run an arbitrary query and pull out anything extra.

Second: a role-based access control model (RBAC). The agent acts within the same permission boundaries as the person on whose behalf it works. If an accountant cannot see a subsidiary's accounts, neither can the agent.

Third: AgentCore provides an isolated execution environment. Fourth: agents' requests and actions are logged so that an audit can be passed later.

There is also a separate product move: banks can run such agents under their own brand. A bank gets AI features for its customers without building them from scratch.

Numbers and caveats

The company claims savings of more than $1 million a year, over 5 hours a day for an accounting team, a 77% reduction in annual TMS costs, and reports in under a minute. There is also a library of ten agent templates built from real deployments. According to Trovata's estimate, most teams have 5–15 processes ready for automation.

The caveat is obvious: this is an AWS post about a partner, and the figures come from Trovata's own case studies. There is no independent verification, and it is unclear how well the results will replicate elsewhere. In addition, the text has no data on how often the agents make mistakes or who is responsible for an error when money is involved.

What to take away

If you are building agents for finance, healthcare, or any other regulated field, the approach transfers in full. Don't give an agent direct database access; route everything through an API with validation. Inherit the user's permissions rather than issuing the agent separate, broad ones. Write an audit log from day one. Start with repeatable processes such as reconciliation and variance explanation: there it is easier to measure the gain and easier to catch an error.