One large multinational company’s FX and commodities team is using AI agents to help traders decide which exposures to hedge, which banks to approach and how to get better pricing. By drawing on a constantly updating database of bank quotes, the agents are helping negotiate tighter pricing while doing less manual work along the way.
- The member described the project at the H2 meeting of NeuGroup for Foreign Exchange , sponsored by RBC Capital Markets . The team began overhauling its operating model as part of a broader corporate push to adopt AI, with a planned treasury management system replacement creating an opportunity to reconsider how the desk operates.
- The same data that informs trading recommendations now feeds dashboards used to track hedges, eliminating over 10 hours of manual reporting work each month. An agent uses the data to reconcile month-end trade invoices from more than 10 banks has cut another half-day of work.
- “We invested a ton of time up front to create this, but we’ve made that back in the amount of time that we've saved from not having to constantly create and refresh dashboards,” the member said.
Better bank pricing starts with better records. Built using a third-party enterprise AI platform customized for the company's needs, the trading agent draws on historical trade data, market rates, outstanding exposures and bank pricing to recommend which hedges to execute and which counterparties to approach. It also checks counterparty credit limits and identifies five banks for traders to consider based on their past pricing.
- Traders can specify which currencies or commodities they want to hedge or ask the agent to identify where action is needed. If a team member wants to focus on copper or aluminum, for example, the agent identifies the relevant exposures, hedge maturities and amounts, then recommends which banks to approach. A person decides which recommendations to pursue.
- The data includes quotes from banks that didn't win the business. Rather than recording only executed trades, traders enter each bank's initial quote, the midmarket rate and the final all-in price. That information goes back into the agent's data sources, giving it a more complete record of how banks have priced similar transactions.
- When a bank asks why it is winning less business, traders can point to the pricing history, which the member says has helped tighten spreads.
- The system updates its recommendations as trades are completed. If a trader hedges part of an exposure today, the agent accounts for that transaction when calculating what remains to be hedged tomorrow.
The hardest work came before the agents. When the team began reviewing its processes before implementing agents, it identified dozens of recurring monthly and quarterly tasks that rely on disconnected spreadsheets and databases. Market data, exposure information, trade history and bank spreads often had to be collected or updated separately.
- The team consolidated those sources into data lakes. Making the information usable by agents required considerable work to standardize and tag fields so they could be retrieved consistently across workflows. The member said that preparation “was frankly the hardest part of this process."
- They learned that giving an agent more information or a longer sequence of tasks did not necessarily produce better results. Some early workflow designs took 15 to 20 minutes to run and returned inconsistent answers. The team responded by dividing the work into smaller stages, with separate agents gathering and analyzing information before another generates instructions for the trader.
- The trading workflow now involves three agents handling successive stages of the process, which the member said is faster and more accurate than one single agent. Prompts have also been narrowed to prevent searching unnecessarily through decades of historical market data.
Traders make trades. The trading agent drafts the request-for-quote language traders send banks over Instant Bloomberg , but a human reviews the proposed hedge, including its maturity and other trade details, before proceeding. - The member estimates the agent now completes its tasks correctly about 98% of the time, following adjustments to problems such as incorrect market calendars. He tested its limits by instructing it to trade with a bank that had reached its credit limit; the agent refused.
- The company already requires employees to pass an exam covering trading policy and the reasoning behind transactions before becoming senior traders. As agents take on more work, the member said those assessments may become quarterly or semiannual and include questions about how the technology operates. “I don't want my team to be blindly relying on it; I want them to use professional skepticism as they go through it,” he said.
