Treasury teams are using AI to turn files into agents, decks into chatbots and two-day processes into five-minute workflows.
Treasury teams are finding traction with AI by taking on specific problems, testing practical tools and building expertise by getting their hands dirty. At recent meetings, members of NeuGroup for Foreign Exchange, NeuGroup for Mega-Cap Assistant Treasurers and NeuGroup for Large-Cap Assistant Treasurers, described how small, practical projects can centralize knowledge, visualize data and automate parts of FX hedging and other treasury work typically requiring heavy manual effort. - “If you wait for a perfect data lake, you'll wait a long time,” said one member who recommended that peers “start small” as their teams build skills and practical use cases while larger data projects continue to take shape.
- “Transformative stuff should be worked on the side,” he added, because “AI moves fast, and it will be easier to implement in the future. And they'll be more skilled once we get there.”
Making existing tools smarter. Some of the clearest examples involved using AI to get more out of existing reports, files and systems. One member described creating “mini data lakes” inside treasury by gathering the reports and information his team uses most often, then connecting those datasets to agents and dashboards that can answer questions and extract relevant details.
- He pointed to insurance as one early use case, saying the team built a Q&A agent around policy documents so users could pull key exclusions and underlying data without digging through files manually.
- The same member also uses an agent that draws on systems including Kyriba, CWAN (formerly Clearwater Analytics), Chatham Direct and Bloomberg Anywhere to monitor investments and flag issues tied to policy limits, credit ratings and counterparty risk.
Taking hours off tasks. Other examples include using Microsoft Copilot to speed up routine treasury work that normally eats up hours in Excel, PowerPoint and Outlook. One AT described AI as a way to shorten the path from raw information to something usable, but not as a replacement for treasury judgment. - Another member uses Google Gemini to take investment data and quickly produce region-level breakdowns and charts, compressing work that could take analysts hours to assemble manually. He uploads a file, asks for amounts by region and it “literally gives you a nice pie chart in two minutes.” That helps the team work from a single source of data instead of having multiple analysts build separate files from different versions.
- Many members find Anthropic’s Claude especially useful for drafting, rewriting and sharpening business materials; it emerged as the strongest all-around tool for some of them. But one member said using it at scale for the kinds of internal applications his company had in mind could become expensive quickly. He pointed to Context as a less expensive, more specialized option for presentation work, especially for creating and improving and executive decks.
- The same member said some senior leaders now feed presentations into a chatbot and ask it questions instead of reading the deck front to back, a sign that AI is starting to change not just analysis but how treasury information is consumed.
- He showed how broad that push has become in his treasury organization, citing eight agents that have increased efficiency by eliminating 175 hours of work. The team has 34 ideas in progress, including FX analysis, trade preparation and reporting, a treasury dashboard, and automation for KYC reporting, bank administration and insurance exposure gathering.
Automating complex FX workflows. The FX meeting featured a more technical example of how one member uses tools of varying complexity, ranging from LLMs to the Python coding language, to build automations for FX processes. For example, the team used AI as a “junior developer” to build a custom, logbook-style process for tracking long-dated revenue contracts, hedge accounting and hedge periods—an area existing systems could not easily automate. This pulls together the data behind a trade recommendation, including large contracts to be hedged and pooled trades, and generates separate files for accounting with the hedge accounting information needed downstream.
- After review, the file can be exported for approval and pushed into FXall, cutting a process that once took about two days down to roughly five minutes.
- In describing the toolkit, he said ChatGPT helps with planning out the project, Cursor can take a prompt and generate an end-to-end solution, Python serves as the underlying language, and VS Code and Jupyter Notebook make it easier to inspect the logic, run steps in sequence and show the work in a format others can follow.
- He added that the larger goal is to keep expanding AI-driven coding automations in FX now that he has direct reports with coding experience who can help build and maintain them. The point, he said, is to spend less time compiling files so treasury can spend more time analyzing exposures and making decisions.
- Asked how to get started, he echoed his peer’s suggestion to start small: “Don't try to boil the ocean; begin with what you can easily validate." Then refine, test and improve from there.
