AI Agents in Reconciliation: Compressing the Cycle from Days to Hours

Background: Manual Reconciliation Hits the Scale Ceiling

A large-scale e-commerce company, operating multiple platforms and stores, processes documents at the 100k level per month. Its finance team reconciled the classic way — manual exports plus Excel cross-checks: statements downloaded from each platform portal at the start of the month, several accountants spending days checking platform by platform, discrepancy causes dug out of chat history, and shared costs (logistics, warehousing, promotion) allocated by gut feel. The long cycle was painful enough; worse, the process was untraceable — any question about a variance meant starting the investigation from scratch.

The Solution: A Fact Layer as Foundation, Agents Doing the Work

The company deployed the Qeasy intelligent reconciliation platform in three layers.

Fact-data layer. Statements from every platform import in one click and persist via async tasks. Each platform × statement type gets its own fact table — 24 in total — with full raw-row snapshots, amounts at Decimal(20,4) and unit prices at Decimal(20,6). Every reconciliation result traces back through a minimal bridge table to the original statement row; traceability is the precondition for everything that follows.

Reconciliation engine. Reconciliation splits into twin plans running in parallel. The revenue plan auto-matches platform statements against supply-chain orders line by line using dual codes (platform order number plus internal business code), persisting variance amounts and reasons per row and writing back item details on successful matches. The expense plan follows a three-tier architecture — plan, aggregation, allocation detail — with sandboxed allocation scripts computing shared expenses and writing them back to revenue lines in real time. The engine runs on BullMQ async queues driven by a seven-state state machine, so high-volume processing never blocks business operations and every step is queryable and re-runnable.

AI agent matrix. Three domain agents — a Bill Parsing Engineer, a Reconciliation Script Engineer and an Expense Allocation Engineer — plus a general assistant as the single entry point. Finance staff state needs in natural language ("set up parsing rules for this new platform's statements"), and the agent follows a five-step workflow: requirement dialogue, automatic reconnaissance, script writing, sandbox testing, then execution. Agents autonomously reconnoiter, write, test and save; a human confirms once, in business language, before execution.

Safety: No Second Write Path

The design baseline: every business side effect of an agent goes through the existing script + JobTask + BullMQ pipeline — agents never mutate business data directly. Tests never persist, script versions are rollbackable, all tool calls land in audit tables, and the sandbox isolates user scripts behind isolated-vm plus dedicated child processes.

Results: From Days to Hours

After go-live, the monthly reconciliation that used to be measured in days now completes in hours end to end — import, parsing, matching and allocation all run asynchronously, and accountants shifted from row-by-row checking to reviewing variances and confirming executions. Onboarding a new platform's statement format went from a scheduled development task to a same-day conversation with the parsing agent. Every variance traces back to its raw statement row and script version, turning month-end audits from file-digging into link-clicking. Shared-expense allocation moved from gut feel to scripted, re-runnable computation.

Lessons

AI Agents in finance require a trustworthy fact layer first — precision, snapshots and bridge-table lineage are all non-negotiable. An agent's value is not replacing the system but letting business users drive it without scheduling; converging all writes onto the existing pipeline is what earns finance and audit trust. And a light-confirmation model — autonomous reconnaissance and testing, one confirmation before execution — delivers efficiency and safety together.