The Reconciliation Work That Shouldn't Take Humans Anymore

If your close is slow, reconciliations are probably why. Tie the sub-ledger to the GL, match the bank to the books, clear the intercompany — line by line, account by account, usually under time pressure with coffee going cold. It's the least glamorous part of the close and often the longest.

The good news is that most of it is pattern-matching, and pattern-matching is exactly what machines do well.

Auto-match first, then agents

Machine learning handles the matching. Instead of exact-match rules that break on a rounding difference or a timing lag, the model matches on fuzzy logic — it learns that these two amounts three days apart with slightly different references are the same transaction. Bank recs, sub-ledger-to-GL, intercompany — auto-reconciled percentages that sat around 50–60% on rules alone move into the 80s and beyond once the model's had a few cycles to learn your patterns.

The newer layer is agentic. For the items that don't auto-match, an agent can go investigate — pull the supporting document, check the bank memo, look at the prior-period treatment, and either propose a resolution or draft the adjusting entry for a human to approve. It's the difference between "here are 400 unmatched items, good luck" and "here are 400 items, 350 with a proposed explanation attached." That second version is a very different Tuesday.

Days-to-close is the number that matters. Auto-reconciliation doesn't shave hours — it can pull whole days out, because reconciliations sit on the critical path and everything downstream waits on them. Clients who were closing in eight or nine days routinely get to five or six once recs stop being the bottleneck. The exception queue becomes the work, instead of the matching.

The honest caveats

A high auto-match rate is only good if the matches are right. A model that matches aggressively to clear items can hide a genuine discrepancy — and a hidden discrepancy in a reconciliation is precisely the thing reconciliations exist to catch. So we tune for precision over recall early on, and we keep unmatched items visible rather than letting the system force a match.

Agentic resolution needs guardrails. An agent drafting an adjusting entry is helpful; an agent posting one unsupervised is a control weakness auditors will find. Keep a human approval gate on anything that hits the ledger, and log the agent's reasoning so it's auditable.

Data quality upstream sets your ceiling. If your bank feeds are incomplete or your intercompany data doesn't agree entity-to-entity, no amount of clever matching fixes that — it just surfaces it faster, which is genuinely useful but not the same as solved.

Start with the highest-volume, most mechanical recs — bank and clearing accounts — prove the auto-match rate and the day saved, then work up to the judgmental ones. The close gets faster and, almost as valuable, less dependent on three heroic people knowing where everything's buried.