Agent Reconciliation: Receipts, Evidence Logs, and ‘Explain This Charge’ Workflows

The best agent payment systems don’t win by letting the agent buy things.

They win by making every purchase explainable.

That’s reconciliation: turning “a line item on a statement” into an evidence-backed story a human can trust.

The reconciliation problem agents create

If you share a primary card with an agent:

• ▸transactions mix with human spend
• ▸receipts are inconsistent
• ▸disputes become ambiguous (“did we mean to buy this?”)

The fix is not “better dashboards.”

The fix is evidence logs that start before the purchase happens.

The evidence record (minimum viable)

Every completed purchase should produce a durable record:

{
  "agentId": "agent_travel",
  "intentId": "int_abc123",
  "purpose": "Book flight to SFO (economy)",
  "policyDecision": "approved",
  "expectedMerchant": "United",
  "expectedAmount": 45000,
  "cardId": "card_xyz",
  "transactionId": "txn_123",
  "postedAmount": 44912,
  "descriptor": "UNITED AIRLINES",
  "createdAt": "2026-01-29T18:22:10Z",
  "verifiedAt": "2026-01-29T18:29:44Z",
  "receiptUrl": "https://merchant.example/receipt/..."
}

If you can’t produce this, your system is not audit-ready.

Receipt capture: what to aim for

Receipts are messy. That’s reality.

Aim for:

• ▸receipt URL (best)
• ▸invoice number + PDF link (good)
• ▸confirmation email thread id (fallback)
• ▸screenshot of order confirmation (last resort)

For many ecommerce flows, the agent can extract a receipt URL at checkout time if the browser executor exposes it.

Explain-this-charge: the 1-page output

Build a single summary view that answers:

1. ▸What was this for? (intent purpose)
2. ▸Who/what approved it? (policy decision + approver)
3. ▸Did it match the intent? (merchant/amount/time)
4. ▸Where is the receipt? (link)

If you can produce that for every charge, finance teams stop fearing autonomy.

Anomaly flags that catch real issues

Add simple flags:

• ▸descriptor does not match expected merchant
• ▸amount exceeds expected amount + tolerance
• ▸multiple authorizations in a short window (retry loop)
• ▸purchase made outside intent window

These four catch most incidents.

Bottom line

Reconciliation is where agent payments become a real product, not a demo.

If you’re building agent spend:

• ▸require intent
• ▸generate an evidence record
• ▸capture receipts
• ▸output an explain-this-charge summary

That’s how autonomous spend becomes operationally sane.