Where Agents Actually Work in Banking
Where They Earn Their Keep
I have built enough with AI agents on my own projects to tell the demo magic apart from what actually holds up. The hype says agents will soon run everything. What I have found is narrower and more useful. They are very good at a specific shape of problem and unreliable outside it. So the interesting question for banking is not whether to use agents, but where they genuinely earn their place, and how to box them in so they stay useful.
Start in the Back Office
The strongest fit, by a wide margin, is internal work that a person reviews before it counts for anything. Reconciling accounts at the end of the day. Combing through transactions to flag the handful that look off. Pulling together the pieces of a regulatory report so a person can check and finish it. Watching for the patterns a compliance team would want to know about.
These fit for the same reasons. The work is tedious and pattern-heavy, which is exactly what agents are good at. And a person sits between the agent and any consequence, so a mistake gets caught in review rather than out in the world. That combination, high tedium and a built-in checkpoint, is the sweet spot. It frees people from the grind without handing anything irreversible to a machine.
Bounded Client-Facing Work
Client-facing is where it gets more interesting and more dangerous. An agent can move from explaining how to reset a password to actually resetting it, from describing how to update an address to making the change. For routine, reversible, low-stakes requests, that is a genuinely better experience, the kind of thing a client wants handled at any hour without waiting.
The catch is the boundary. The moment a request involves real money, a judgment call, or anything not easily undone, the agent's job is to hand off to a person, not to push ahead. The version that works is narrow and a little boring, handling the small stuff cleanly and escalating the instant it leaves its lane. The version that ends in a headline is the one given room to improvise with things that matter.
The Real Work Is the Constraints
Here is the part the demos skip. Most of the effort in a real agent system goes not into the clever part but into the limits around it. What this agent is authorized to touch, and nothing beyond it. What hard lines it cannot cross. When it has to stop and escalate. And a complete, reviewable record of everything it did and why. Get those right and an agent is an asset. Skip them and you have built a fast way to make a mess.
This is also why the systems that hold up tend to use several narrow agents rather than one that does everything. A doer that drafts the work and a separate checker that reviews it catch more than a single agent grading its own homework. From building these, the failure I run into most is an agent confidently taking a wrong path and then compounding it step after step. Narrow scope and a checkpoint are what keep that contained.
Drawing the Lines
Agents are neither the everything-machine of the hype nor the toy the skeptics describe. They are a strong tool for a particular kind of problem, bounded, repetitive, reviewable, and well instrumented. In a bank, that means starting narrow and internal, proving the guardrails hold, and widening scope only as the trust is earned. The institutions that get real value from agents will not be the ones that deployed the most. They will be the ones that drew the lines carefully and held them.