Case Study: Agentic Reconciliation & Break Resolution for Fund Operations

Asset & Wealth Management

Agentic Reconciliation & Break Resolution for Fund Operations

The Challenge: Thousands of Breaks and a Shrinking Settlement Window

The operations team at a global asset manager reconciled cash and positions every day across more than a dozen custodians and prime brokers. The matching engine left about 4,000 breaks each morning for 30 analysts to clear by hand. Around 60% came from timing differences or reference data mismatches, and each of those still had to be traced to its source and explained. Shorter settlement cycles left less time to do it, older breaks were piling up, and the firm's experienced analysts had little time left for the exceptions that carried real risk.

Our Solution: Agents That Investigate Each Break and Propose a Fix

Lydatum built an agentic workflow on top of the existing matching engine that takes each break from detection to a proposed resolution:

• Classification Agent: Grouped related breaks and assigned each a likely cause, based on how the firm had resolved similar breaks in the past.
• Investigation Agent: Traced each break to the originating trade or cash movement on both sides and wrote a short explanation of the difference, with links to the supporting records.
• Resolution Agent: Prepared the corrective action, such as a draft adjustment or a query to the counterparty, and sent it to the right analyst for approval.
• Control & Sampling Agent: Applied the firm's tolerance rules so that low-value timing breaks cleared automatically, and picked a daily sample of those automated resolutions for independent human review.

Nothing above the agreed thresholds was posted without an analyst's approval, and each proposal carried the evidence the analyst needed to decide. We built an evaluation suite from two years of historical breaks and rerun it whenever a prompt or model changes, so operations and risk can check accuracy before an update is released.

The Impact: Analysts Focused on the Breaks That Matter

Six months after launch, agents were handling 82% of daily break volume and the team's attention had moved to complex and high-value exceptions.

Routine breaks are now cleared or queued for one-click approval before the team starts its day. The backlog of aged breaks was gone within the first quarter, and the firm met its shorter settlement deadlines without adding headcount. Sampled reviews have kept agent accuracy above the threshold set by operational risk.

Services Delivered: Agentic Workflow Automation, Multi-Agent Systems & Orchestration, Agent Evaluation & Observability, Data Pipelines & Integration

Technologies Used: LangGraph, Azure OpenAI, Model Context Protocol (MCP), Snowflake, dbt, Apache Kafka, Python