Building AI Agents for Bank Treasury: Think Less Chatbot, More Operating Model

Most people first experience AI as a chatbot.

You ask a question. It gives an answer. Sometimes the answer is excellent. Sometimes it is generic. Sometimes it misses the point. You adjust the prompt, add more context, ask again, and eventually you get something useful.

That experience teaches one of the first lessons about generative AI: it is not a deterministic calculator. It is a model. Its output is stochastic.

If you ask the same question twice, you may get different answers. If you change one sentence in the prompt, you may get a better answer or a worse one. This can feel frustrating, especially when one attempt produces something sharp and the next produces something flat.

That does not mean the technology is random in the way a slot machine is random. But the analogy is useful up to a point. You are not pulling a lever and getting a guaranteed result. You are interacting with a probabilistic system. The craft is learning how to shape the odds.

For treasury teams, that distinction matters.

If you treat AI as a magic answer machine, you will be disappointed. If you treat it as a junior analyst with infinite patience, no judgment, limited memory, and a tendency to sound confident when wrong, you are closer to the right mental model.

And if you want to use it seriously, you have to move beyond the chatbot.

A chatbot session has limits. It has a context window, which is the amount of information the model can hold in view at one time. The longer the session gets, the more crowded that context becomes. At some point, performance can degrade.

The model may lose track of the original objective, overweight recent details, forget constraints, or start blending old instructions with new ones.

Anyone who has worked with these tools for more than a few hours has seen this pattern. The conversation starts clean. Then it gets long. Then the answers become less focused. You keep adding more information because you want to help the model, but sometimes that makes the problem worse.

This is one of the most important practical lessons in AI adoption:

Better AI use is not just better prompting. It is better context design.

You have to decide what information the model needs, what it does not need, what tools it should have, what task it is responsible for, and how its output will be checked.

That is where the agent concept becomes useful.

An AI agent is not just a chatbot. It is a model placed inside a role, with a task, a context, a set of tools, and some process for deciding what to do next. You can think of it as a small specialized worker. It may have access to documents, data, code, search tools, spreadsheets, internal policies, or prior work. But it still has limits. Give it too much to do and performance falls. Give it unclear goals and it will improvise. Give it weak controls and it may produce something plausible but wrong.

This is why organizational design matters.

You do not build one giant AI agent and tell it to "help the treasury." That is too broad. Treasury work is not one task. It is a collection of workflows: data sourcing, reconciliation, forecasting, modeling, assumption governance, documentation, challenge, reporting, controls, audit response, and management interpretation.

The more useful approach is to break the work apart.

For example, imagine building an AI-enabled modeling team for bank treasury. The goal might be to develop, document, and challenge a model for IRRBB, liquidity, or deposit behavior. That process already involves many different kinds of work. AI does not remove that structure. It makes the structure more important.

You might design a set of specialized agents:

• A data sourcing agent that identifies required data and checks availability.

• A data cleaning agent that profiles files, flags missing values, and prepares transformations.

• An exploratory analysis agent that studies patterns, seasonality, outliers, and segment behavior.

• A modeling agent that tests candidate approaches.

• A statistical review agent that challenges model fit, stability, sensitivity, and assumptions.

• A documentation agent that drafts methodology, limitations, and usage guidance.

• A validation agent that asks whether the model is fit for purpose.

• An orchestrator agent that manages handoffs, tracks open questions, and decides what needs to happen next.

  
  

  

  This starts to look less like "using ChatGPT" and more like designing an operating model.

  
  

  That is the connection treasury leaders should pay attention to. The question is not simply, "Can AI write a memo?" The better question is:

What parts of our treasury process can be decomposed into repeatable tasks, given the right context, tools, controls, and review loops?

This is familiar territory for treasury and ALM teams. We already think in systems. We already break complex problems into assumptions, data, models, scenarios, governance, reporting, and challenge. AI agent design is similar. You are designing roles, handoffs, constraints, escalation points, and review standards.

The human does not disappear. The human becomes the architect and reviewer.

That role is not trivial. In fact, it may be the highest-value part of the work.

Someone has to define the objective. Someone has to decide what "good" means. Someone has to know when an output is merely fluent versus actually correct. Someone has to decide what evidence is sufficient. Someone has to decide whether a model is fit for purpose.

For bank treasury, this is especially important because the cost of being wrong is not just a bad paragraph or a clumsy spreadsheet. The risks include weak assumptions, misleading forecasts, bad management information, poor controls, audit findings, and regulatory scrutiny.

So the right posture is neither hype nor dismissal.

AI can help treasury teams do work that is slow, technical, repetitive, and documentation-heavy. It can accelerate analysis. It can draft. It can test code.

It can search large knowledge bases. It can compare assumptions. It can help generate challenge questions. It can maintain institutional memory if you design the memory layer properly.

But it needs structure.

A good AI system for treasury should have

• Clear goals and success criteria.

• Narrow agent responsibilities.

• Curated context rather than unlimited context.

• Access to the right tools and knowledge base.

• Review loops.

• Audit trails.

• Human approval gates.

• Controls around sensitive data.

• A way to learn from prior work without blindly carrying forward mistakes.

This is where the opportunity is.

Treasury teams do not need to become AI research labs. But they do need to understand enough about how these systems behave to design around their strengths and weaknesses.

The first stage of AI use is asking better questions in a chatbot.

The next stage is building repeatable workflows.

The more advanced stage is designing teams of specialized agents that can work through a process, produce artifacts, check each other's work, escalate uncertainty, and improve over time through a knowledge base and memory.

That is when AI becomes more than a productivity tool. It becomes part of the operating infrastructure.

For treasury, I think the biggest gains will come from combining domain expertise with thoughtful system design. The people who know ALM, liquidity, FTP, deposits, balance sheet behavior, model governance, and regulatory expectations are the people best positioned to build useful AI workflows.

Not because AI replaces that expertise. Because AI needs that expertise to be useful.

Disclaimer: The views and opinions expressed in this article are my own and do not reflect the views of my employer or any organization I am affiliated with. This content was written with the assistance of AI tools and is provided for informational and educational purposes only. It is not intended as financial, legal, or regulatory advice. Readers should consult with qualified professionals before making any decisions based on this content.

About The Author

Chih Chen

is a bank treasury and risk leader with over 20 years of experience in Asset Liability Management and IRRBB. Currently SVP, Treasury at East West Bank, he focuses on translating quantitative insights into earnings stability and risk strategy.

He has previously held leadership roles at BNP Paribas (Bank of the West) and MSCI Inc., and is a published voice on deposit behavior and ALM.