Receipts & reconciliation
Matching at scale: expense receipts to card feeds, statements to ledgers — tiered auto-match thresholds, the unmatched lane, and reconciliation as the pattern behind it all.
The third document family is really a matching family: two streams that should agree — receipts vs. card transactions, bank statement vs. cash ledger, vendor statement vs. AP subledger — and the work of pairing them up and explaining the residue. Humans are slow at it and it numbs them; AI is fast at it and never gets bored; the design question is purely where to set the confidence tiers.
The three-tier matching pattern
- Auto-match (high confidence) — exact or near-exact on amount + date-window + counterparty: paired automatically, logged, no human touch. In mature deployments this is 80-90% of volume. The tier's threshold is a control decision, documented in Module 5's matrix, not a default accepted from a vendor.
- Suggested match (medium confidence) — amount off by a tip or FX rounding, date off by a settlement day, merchant name garbled ('AMZN MKTP' vs the Amazon receipt): AI proposes, human confirms with one click. The human is confirming a specific hypothesis, which is 10× faster than searching cold.
- Unmatched lane (everything else) — the residue is the point of reconciliation: the duplicate charge, the personal expense on the corporate card, the bank fee nobody booked, the deposit in transit. AI accelerates the pairing precisely so human attention concentrates on the residue — reversing the traditional ratio where 95% of reconciliation time went to the easy 95%.
Receipts specifically: policy checking rides along
Expense receipts add a second AI job beyond matching: policy screening. Amount within category limit? Alcohol on the itemized bill? Date on a weekend without a travel record? Attendees listed for the client dinner? The AI flags; the approver decides — the Module 1 caution about approver independence stands (AI screens for the approver, it doesn't become the approver, and the approval log records a human name). Screening at 100% quietly ends the era where expense policy was enforced by the 5% audit and the honor system — announce it kindly before rollout, and watch behavior adjust in a month.
Auto-matching accelerates reconciliation; it does not perform it. The reconciliation as a control = a named preparer asserting 'these agree and here's the explained residue' + a named reviewer checking it, both logged with dates. Keep those signatures human even at 95% auto-match — the day something systemic breaks (a duplicated bank feed, a double-posted batch), the auto-match tier will pair the wreckage beautifully, and only the human staring at the residue line will feel that something's wrong.