Raghav Anand
← All postsNov 2026 · Raghav Anand

Why three-way reconciliation is the long pole

Matching your ledger to a processor is tractable. Adding the sponsor bank is where weekends, delayed ACH and returns pile up.

At Dave I worked on reconciliation for cash advances. We delivered two-way reconciliation, our ledger against our payment processors, with reporting for Finance. We also put together a framework for the next step: bringing in the sponsor bank. That third leg was the long pole, and it's worth explaining why.

Two-way is mostly a join

Your ledger and your processor usually share identifiers and see the same transaction within a short window. Matching them is mostly a join with some tolerance for timing and fees. It has rough edges, but it's a known problem.

The bank runs on a different clock

The sponsor bank sees money move in batches, on banking days, with its own references. Three things make the third leg hard:

  • Timing. Long weekends and holidays push settlement out, so a transaction that's done in your ledger may not appear at the bank for days.
  • Returns. An ACH debit can come back days after it looked successful, for reasons like insufficient funds. The original match was correct when you made it and wrong later.
  • Shape. The bank often reports aggregated or differently keyed records, so one ledger line may map to part of a batch.

Add duplicates, late postings and fee mismatches on top, and a manual process turns into a recurring fire drill for Finance and Ops.

You can't automate what you can't score

When I started prototyping automation for this, the first gap was data. Real financial data can't leave a company, and synthetic data without known breaks can't tell you whether your matcher works. So I built a generator that produces ledger, processor and bank records for advances and repayments, injects realistic breaks, and writes an answer key for each one.

With an answer key, any approach gets a score: exact rules, fuzzy matching, or an AI step that explains the leftover exceptions for an ops analyst. It's the same lesson as with claims. The exact matches belong to rules. AI is useful for the exceptions, and only if you can measure it.

The generator is on GitHub. All data is synthetic.