AI Agents in Finance

An AI agent doesn't just answer a question — it plans a sequence of steps, pulls data on its own, and works toward a goal. Here's how that shows up in research, monitoring, and decision support.

From chatbot to agent

A standard AI chatbot takes a prompt and returns a response — one exchange, one turn. An "agentic" system is built differently: it's given a goal, and it plans a sequence of steps to get there, deciding along the way which tools or data sources to call, evaluating what comes back, and adjusting the plan before producing a final result. The distinction matters because it changes what the system can be trusted to do unsupervised, and what still needs a checkpoint.

What agents get set up to do

In finance, agent-style systems are typically built around narrow, well-defined tasks: a research agent that pulls a company's recent filings, cross-references them against a watchlist, and drafts a summary; a monitoring agent that checks a set of data sources on a schedule and raises an alert when something crosses a threshold; a support agent that walks through a checklist of due-diligence questions and flags gaps. The common thread is a multi-step task that would otherwise require someone to manually chain several lookups together.

The tool-use loop

Under the hood, most agents run a loop: decide what to do next, call a tool or data source, look at the result, and decide again. This is what lets an agent "go get" information rather than only working with what's already in the prompt. It's also where things can go wrong — if an early step returns bad data, or the agent misinterprets a result, the error can carry forward into every subsequent step.

Autonomy versus oversight

The more steps an agent is allowed to take without a person checking in, the more useful it is for saving time — and the more it matters where the checkpoints sit. Well-designed systems build in a human review point before anything consequential happens, especially before any step involving money movement, client communication, or a public-facing action. Autonomy and oversight aren't opposites here; the practical question is where along the chain a person signs off.

Risks specific to agents

• Compounding errors. A wrong assumption early in a multi-step task can propagate and get harder to spot by the final output.
• Tool misuse. An agent given access to more systems than it needs has more ways to make an unintended change.
• Over-trust. A confident, well-written final summary can mask a shaky chain of intermediate steps that nobody checked.