Second Read: receipt and invoice OCR that tells the bookkeeper which pages to check

For bookkeepers, receipt and invoice OCR with a review queue is a crowded, commoditized category. Dext, AutoEntry, DocuClipper and ParseField all sell to the same practices. Hubdoc and QuickBooks receipt capture are free inside the ledger software. Showing low-confidence fields for review is already standard: ParseField, Mindee, Taggun and Extend all do it. So "tells you which pages to check" will not stand out as a pitch on its own. The open space is narrower. The first opening is local processing. The only offline desktop tool I found, Bookkeeping-OCR, makes the user confirm every result and has no automated triage. A local tool that never uploads client documents could appeal to privacy-sensitive or regulated practices, and to bookkeepers whose clients object to cloud processing. The second opening is trust. Vendor confidence scores are self-reported and nobody independently checks whether they are calibrated. If Second Read can show that its "fine" pile really is fine, for example a measured miss rate on real messy thermal receipts, it could sell as a second check on top of whatever extraction tool a practice already uses, rather than as a Dext replacement. The risk is that freelance bookkeepers are price-sensitive and already pay for, or get free, tools that are "good enough" and post straight into the ledger. A one-way CSV export is weaker than Dext's direct publishing. The OCR Gate idea is much less contested. I found no dedicated GitHub Action that fails a build when the rate of untrusted extractions goes up. all2md's quality gate and UiPath's evaluation feature come closest. Teams currently build this themselves from golden datasets. The catch is that the market is small and developer-tooling buyers expect it free or open source. Cleanlab TLM or Azure Content Understanding could add a CI wrapper easily. It probably works better as an open-source wedge that markets the judge model than as a standalone paid product.

• Dext (formerly Receipt Bank)

  This is the incumbent receipt and invoice extraction tool for bookkeepers, with Xero and QuickBooks publishing built in. Dext says it is 99.9% accurate, but that figure comes from the vendor. It runs in the cloud and is sold through accounting practices, which is the same buyer Second Read targets.

• Hubdoc (Xero)

  Hubdoc comes free with Xero subscriptions. Practitioners report that its accuracy trails Dext and AutoEntry, and one source says it is in maintenance mode. It sets the price floor at zero for Xero users.

• AutoEntry

  A cloud extraction tool for accountants. One 2026 reviewer measured about 93% first-pass accuracy on supplier, date, VAT and total, which leaves a small but real share of receipts that need checking.

• QuickBooks Online Receipt Capture

  Free built-in OCR that suggests matches against the bank feed, and nothing posts until a person reviews it. Accuracy is reported to drop on thermal receipts, busy mobile photos and multi-page documents.

• ParseField

  Markets per-field confidence scores so a reviewer checks only the flagged items, plus balance-variance checks before export to CSV, Excel or JSON. This is almost exactly the 'which pages to check' pitch, but it runs in the cloud.

• DocuClipper

  Cloud receipt, invoice and bank-statement OCR aimed at accountants, with QuickBooks and Xero export.

• Bookkeeping-OCR

  An offline desktop app for Mac and Windows that reads vendor, date and total from receipts and keeps data on the local machine. It costs $119 once. It is the closest match to the local desktop angle but has no automated trust flagging: the user confirms every result.

• Taggun

  A receipt OCR API with a confidence score per document. Its docs recommend a threshold of about 0.7 and sending lower scores to manual review.

• Mindee

  A receipt and invoice extraction API with confidence on each field and an accuracy-boost feature. Its docs recommend sending medium or lower confidence results to human validation.

• Cleanlab TLM

  Trust scores for LLM structured extraction, both for the whole output and for each field, with an invoice extraction example. This is essentially the 'judge model' component offered as an API to developers.

• Microsoft Content Understanding

  Azure's document extraction service with confidence, grounding and human-review routing. A big-company version of the extraction-plus-trust-flag idea, aimed at developers.

• all2md conversion-quality GitHub Action

  A GitHub Action that scores document conversion quality, including OCR fallback confidence, and fails the job below a threshold you set. It is the nearest existing thing to OCR Gate, though it is aimed at conversion fidelity, not extraction trust rates.

• UiPath Document Understanding evaluation

  A platform feature that reruns OCR to measure how it affects extraction accuracy when models change. It covers regression evaluation inside an enterprise platform, not as a CI gate.

Checked 2026-10-28