Post-trade reconciliation automation for a trading platform

Trades landed in the three systems with mismatched identifiers, timing differences, and occasional partial fills that split one economic trade into several records. Matching them was a spreadsheet exercise that consumed the desk's entire morning: roughly four hours of concentrated, error-prone work before any real operations could start.

The stakes were asymmetric: a routine day produced a handful of legitimate breaks, but missing a single genuine discrepancy meant a mis-stated client position and a potential compliance incident. The desk was forced to treat every one of thousands of rows with the same suspicion, because the process had no way to concentrate attention where it mattered.

Volume was growing 30% year over year. The choice was to keep adding headcount to a process nobody trusted, or to rebuild it.

We built a reconciliation engine that ingests all three sources continuously and matches trades on a configurable rule cascade: exact identifier matches first, then fuzzy matches on instrument, quantity, price, and timestamp windows, with every rule's tolerance explicit and auditable rather than buried in a spreadsheet formula.

The engine auto-clears the overwhelming majority of the book. What remains are true breaks, and those go to an AI agent that investigates each one before a human sees it. The agent pulls context from all three systems, checks for the known break patterns (partial fills, late settlement, fee differences, corporate actions), classifies the likely cause, and drafts the correcting entry.

Every agent conclusion arrives in a morning review queue as a one-paragraph explanation with linked evidence. The operations analyst approves, amends, or escalates; nothing posts to the ledger without a human decision, and every decision is logged against the evidence the agent assembled.

The system runs on durable workflows, so a mid-run failure resumes instead of restarting: an operational requirement the client's compliance team specified after their previous automation attempt produced silent gaps during outages.