Sign up today for free. No credit card down.Get started

    Our approach to AI

    We Take a Different Approach to AI.

    A summary that is mostly right is still a good summary. A bill that is mostly right puts a wrong number in your books. So our AI reads the document, and then code checks what it read before anything is posted.

    An invoice for $4,812.50 is pasted into an AI chat. The chat reads the total as $4,182.50, gives no warning, and that wrong total is what gets posted, off by $630.00.INVOICETOTAL$4,812.50AI chatVendorBrindlecote SupplyInvoice #BS-20417DateSep 12, 2026Total$4,182.50BILLSBrindlecote Supply4,182.50Kellerby Freight312.00Posted as read.Off by $630.00

    AI on its own leaves holes

    Paste an invoice into ChatGPT or Claude and it will read most of it correctly. No model reads every field right every time, and the chat never tells you which one it missed. It also has no record of what is already in your books, so the same bill pasted twice gets entered twice. A wrong total or a duplicate bill posts looking like every other entry.

    What the AI read passes through dozens of checks written in code. It comes out the other side with every field verified.What the AI readDozensCHECKS IN CODEEvery field testedEvery document, every time

    Code checks every read

    After the AI reads a document, dozens of validation checks written in code test what it read, before anything is posted. Unlike a model, code gives the same answer every time.

    A document that passes every check posts to QuickBooks or Xero. A document that fails one is held for review and waits for you.PassedFailed a checkQUICKBOOKSXEROPOSTEDHELD FORREVIEWWaits for you

    Anything that fails waits for you

    A document that fails a check is held for review instead of being posted to QuickBooks or Xero. You see it, you decide, and only then does it reach the books.

    An illustration of the idea. An invoice is read by AI, which decides it is a bill from Brindlecote Supply for $4,812.50. That decision runs down a chain of yes-or-no checks. A missing field sends it back for the AI to read again. Three checks run side by side. A document already in the books is skipped. Any failed check drops it onto a line that ends in review. Only a document that passes every check posts to QuickBooks or Xero.INVOICETOTAL$4,812.50THE AI DECIDESTypeBillVendorBrindlecote SupplyTotal$4,812.50DueOct 12, 2026AccountSuppliesNONONONONONOYESRead it againRead it againYESSkip it.Already inyour booksSIDE BY SIDEDoes the mathadd up?Do the datesmake sense?Is the currencyright?Is anythingmissing?NOIs it the rightkind of document?YESDid all threepass?YESIs it the vendorwe think it is?YESIs it reallyyours to pay?YESHave we seenit before?NODoes every linehave an account?YESWill it postexactly as read?YESQUICKBOOKSXEROPOSTEDHELD FOR REVIEWPosted

    How We Measure It.

    WHAT WE SENT558documents, emailed to fiveclient companies of one firmClean PDF324Scan140Phone photo49Fax14Photocopy13Image quality of the 540 industry documents142with a deliberate trap18 kinds, built to give a plausible wrong number216with no text layer at allThe picture is all there is to readHOW IT SCORED98.3%Fieldsread right6,584 of 6,69595.9%Every keyfield right494 of 51598.6%Recordsclassified right551 of 559THE 21 DOCUMENTS WITH A KEY FIELD WRONGWHERE IT MISSED9Receipt subtotal read as an itemThe total was right on every one8One other key fieldA date, vendor, number or total3Bill date read from a terms dateA later date printed beside it1Set aside as a statementSo no record was madeWHAT HAPPENED NEXT8 posted correctlyThe QuickBooks record matched the document11 held for reviewWaiting for a person before anything posts1 posted wrongWent into QuickBooks with a field wrong1 never reached a recordNothing went into QuickBooks● 25 of 25 vendor statementskept out of the books

    We emailed 558 documents to five client companies of one bookkeeping firm account, then scored what DocStreamAI did field by field against an answer key.

    We made that pile hard and messy on purpose, because real inboxes are. 216 of the documents have no text layer at all: 140 scans, 49 phone photos, 14 faxes and 13 photocopies, where the picture is all there is to read. 142 carry a deliberate trap, from 18 kinds we built to produce a confident, plausible, wrong number. A test of tidy PDFs would be easier to pass and would tell you far less.

    We scored it strictly too. DocStreamAI read 98.3% of fields correctly (6,584 of 6,695). The key-field score is the hardest measure: a document only counts if every key field on it is right, so a single wrong field fails the whole document. On that measure, 95.9% of documents were perfect (494 of 515). It classified 98.6% of records correctly (551 of 559), and all 25 vendor statements were kept out of the books.

    A miss does not have to reach your books. Of the 21 documents with a key field wrong, 8 posted correctly, 11 held for review, 1 posted wrong and 1 never reached a record. Scored on October 2, 2026. Read the full accuracy report and every miss, one by one.

    Focus on growing your business,
    not data entry.

    Start your free trial today. No credit card required. Set up your account in under five minutes.