−40% reconciliation cost with an AI matching pipeline

A US payments scale-up was growing faster than its finance team could close the books. Three payment providers, three export formats, and thousands of daily transactions that someone had to match by hand against the ledger.

The problem

Reconciliation ran on spreadsheets. An analyst exported settlement files, eyeballed them against internal records, and resolved mismatches one at a time. At low volume it worked. At the company's new volume it meant a multi-day close, recurring errors, and a hiring plan that scaled headcount linearly with payments — exactly the cost curve the business was trying to avoid.

What we built

We replaced the manual pass with a matching pipeline that ingests each provider's settlement file, normalizes it, and matches line items against the ledger. Deterministic rules handle the clean majority; an LLM-backed matcher proposes resolutions for the ambiguous remainder, each with a confidence score and a reason an analyst can audit. Anything below threshold is queued for human review rather than silently guessed.

The pipeline is observable end to end: every match, every exception, and every model decision is logged, so finance trusts the numbers and can explain them.

The result

The team now reconciles in hours, not days, and the close happens earlier in the month.

• Reconciliation cost fell 40%
• 98.5% of transactions auto-match without human touch
• Live in 11 weeks from kickoff

The matcher gets the volume; the analysts get the exceptions worth their time.

Figures are anonymized at the client's request.