Receipt reconciliation: from shoebox to close package.

Problem

Solo operators (plumbers, freelancers, Etsy sellers, one-truck landscapers) run the same broken pipeline. Receipts pile up in a glovebox. Invoices sit in an email folder. Every January the whole mess lands on an accountant who has to reconstruct twelve months of history.

The operator does not want accounting software. What they actually do is receive documents. What the accountant needs is those documents, categorized, totaled and exportable.

The real failure mode is silent guesses. A wrong total in a tax export is worse than no product.

Before

Kyle mapped what actually happens before designing anything:

• Capture is ad hoc: glovebox, email folder, phone camera roll, no single inbox.
• Categories are guesses: no vendor memory.
• January is archaeology: the accountant reconstructs, and deductions get missed on faded receipts.
• No exception path: duplicates, ambiguous vendors and low-confidence reads have nowhere to land.

How it works

Ingestion. Invoices are forwarded to a per-org address, or receipts are photographed through a mobile capture flow. Attachments go to object storage and a content hash deduplicates within the org.

Extraction. A worker calls the extraction model with a structured-output schema. Fields below threshold create review items. Nothing below threshold ever reaches an export.

Vendor rules. A category correction (“this Shell station is Fuel, not Meals”) becomes a vendor rule, so next month’s identical receipt skips review.

Monthly close. A job attempts to close the prior month and refuses if unresolved review items exist, emailing the operator a deep link instead. When clean, it produces a PDF cover (category totals, flagged list, missing-receipt gaps), QBO- and Xero-importable CSVs and a ZIP of source images. The accountant share link is hashed, revocable and access-logged.

Exceptions are the product

Six exception types route to review instead of silently writing bad data:

• Amount mismatch: total does not fit the expected range or prior pattern.
• Date window: receipt date falls outside the close period.
• Vendor ambiguity: normalized name matches more than one vendor.
• Duplicate: same vendor, amount and date window as an existing entry.
• Recurring vendor, no receipt: history shows monthly spend but no document arrived.
• Low confidence: extraction below the floor, so the operator is prompted to retake.

The close gate is deliberate: the package is clean or it does not exist.

What I learned

Confidence is the product. Every field shows its score; anything below threshold lands in review.

The close package is the artifact, not a dashboard.

Cut bank-feed matching from v1. A half-working matcher would destroy the trust the confidence flags build.

Block the close, don’t warn and proceed. The gate has to be structural.