The reconciliation mindset

Reconciliation is the least glamorous system I've ever built and the one that taught me the most about agentic design.

The job sounds trivial: match bank-statement lines against expected collections. The reality is that money arrives mangled — truncated references, merged transfers, sums that are two invoices and a rounding error in a trench coat. Our agentic flow auto-matches 99% of lines. The design insight is entirely about the other 1%.

The naive framing of automation is "replace the human." The reconciliation framing is "spend the human where the machine is uncertain." The system's real output isn't the 99% it matched — it's the ranked, contextualized exception queue it hands to a person, with its own reasoning attached. The human stops being a matcher and becomes a judge.

I now apply this shape to every agent system I build:

• Make the agent classify its own confidence, honestly. An agent that can't say "I don't know" will say something worse.
• Design the exception surface as carefully as the happy path. That's where the humans live now.
• Measure the escaped errors — wrong matches that sailed through — not just the match rate. A 99% auto-match rate with silent mismatches is a liability, not a feature.

Boring domains are underrated teachers. Nobody lets you be sloppy with other people's money.