Multi-Agent AI for Foreign Exchange Operations in South Korea
Multi-Agent AI for Foreign Exchange Operations in South Korea
- 60%Faster FX processing
- 40%Higher operational efficiency
- 5AI agents in one workflow
Project Overview
Savvycom built a five-agent AI assistant inside Mattermost for a South Korea-based foreign exchange company in three months. It cut FX processing time by 60% and improved operational efficiency by 40%. Quote ID locking, role-based access and automatic logging sit inside the workflow, so the faster path is also the controlled path.
Transaction volume was growing, and the team needed faster FX operations without loosening control. Internal communication already ran in Mattermost. The brief was to bring FX operations into that channel, not to replace the tools the team used every day.
About The Client
A South Korea-based financial services company specializing in foreign exchange trading and currency exchange. It handles high-frequency BUY/SELL transactions and real-time rate inquiries. Facing growing transaction volumes, the team needed faster and more controlled FX operations without replacing their existing internal collaboration tools.
Challenges
Of the four problems the client faced, three were control problems rather than speed problems.
- Manual FX operations: rate checking and trade confirmation were slow and labor-intensive as volume grew.
- Complex trading rules: settlement handling, quote validity and BUY/SELL logic needed strict, consistent enforcement.
- Inconsistent rate risk: periodic and on-demand rates circulated in the same channel and could cause confusion.
- Approval and access control: secure execution required clear role-based permissions.
A faster assistant that let the wrong person change settlement rules, or traded at a rate the user never confirmed, would add risk. The design had to make the fast path and the controlled path the same path.
Solutions
Savvycom built a multi-agent system with LangGraph and GPT-4o, embedded in Mattermost. Users check rates, manage settlement rules and execute trades in natural language. Each agent owns one part of the workflow, with its own rules and permission boundary.
- User Management Agent: role-based member and permission control.
- FX Rate Agent: on-demand rates for exchange and periodic rates for reference.
- Settlement Agent: channel-level BUY/SELL settlement configuration, admin-only.
- Exchange Agent: chat-based BUY/SELL execution with quote ID locking, confirm or reject, and automatic re-quote on expiry.
- Transaction Agent: search and review of past FX transactions in the chat channel, for audit and tracking.
Splitting the work this way means a change to settlement rules touches one agent and one permission boundary, not the whole assistant.
Quote ID locking
The Exchange Agent locks a quote ID when a user starts a BUY or SELL. The user then confirms or rejects. If the quote expires first, the assistant requests a fresh quote instead of executing at a stale rate.
Reference rates versus executable rates
Periodic rates inform a decision. Only a locked, on-demand quote can be traded. Keeping the two apart removed the confusion the client had seen when different rates circulated in one channel.
Role-based permissions
A regular user can request rates and trade within the rules but cannot change the rules. Settlement configuration is admin-only.
Automatic audit trail
Every transaction is logged by the workflow, not by the user. Authorized users can search past trades without leaving Mattermost, so an audit question is answered where the trade was made.
Results
The assistant went live after a three-month project. FX processing time fell by 60% and operational efficiency improved by 40% against the client’s previous manual process. These figures reflect this client’s volume and workflow, so other teams should read them as a reference point, not a forecast. Automatic logging, quote locking and role-based access reduced trading risk. Adoption needed no new tool, because the assistant lives in the existing Mattermost channel.
Technical Overview
AI: LLM (GPT-4o), embeddings, multi-agent orchestration with LangGraph
Backend and data: Python, PostgreSQL, Redis, REST APIs
Infrastructure: Docker, hybrid cloud on AWS
Integration: Mattermost
Frequently asked questions
What is quote ID locking in FX trading?
A quote ID ties a trade to one specific quoted rate. If the quote expires before the user confirms, the trade does not execute at the old rate. In this project the assistant requests a new quote automatically, so users always confirm the rate they actually trade at, even during busy periods.
How does a multi-agent AI system enforce role-based access in financial workflows?
Permission checks sit with a dedicated agent, and sensitive actions are restricted to admins. Here the User Management Agent handles roles, and the Settlement Agent is admin-only. Ordinary users can request rates and execute trades within the rules, but they cannot change the settlement configuration that governs those rules.
How long does it take to build an AI assistant for FX operations?
This project took three months for a five-agent assistant integrated with Mattermost. The main scope drivers were the number of agents, the systems to integrate, and the trading and approval rules to enforce. An assistant covering a single workflow stage can take less time; a multi-system deployment takes longer.
How does an AI trading assistant keep an audit trail?
The workflow logs every transaction automatically, without relying on the user. A dedicated Transaction Agent lets authorized users search and review past transactions inside the Mattermost channel. The record is created by the execution path itself, so there is no separate compliance step that a busy desk could skip.
Can the AI assistant execute FX trades without human approval?
No. Execution runs through quote locking and a confirm-or-reject step by the user, and settlement configuration stays under admin control. The assistant speeds up each workflow step, but people keep the actual decisions. The assistant cannot change the settlement rules that define what it is allowed to do.
Have a similar challenge?
Tell us about your project. We will share how we approached comparable cases and what results to expect.
Talk to our Experts
