Invoice processing & extraction

Invoice processing software that checks the bill, not just reads it

Every field is read with its own confidence score, and what it is unsure about goes to a person. Then the amount is checked against what that vendor billed for that property before.

The free plan processes 25 invoices a month with unlimited users, takes no credit card and does not expire. You get your own intake address at the end of setup.

Extracted fields

Vendor · Read
Bright Green Landscaping
98%
Invoice number · Read
BG-4417
97%
Amount due · Read
$1,595.00
96%
Property · Inferred
Harborview Apartments
40%

Amount due, read from: AMOUNT DUE $1,595.00

Needs reviewA guessed value cannot auto-approve.

Second look: usual $1,450.00, this one +$145.00 (+10.0%)

An example. Yours shows your own vendors, properties and amounts.

Reading the invoice is the easy half. The rest is knowing when to ask.

Two things decide whether automated invoice processing pays for itself: whether it knows when it does not know, and whether anything looks at the amount it just read.

Invoice processing software has three jobs. Capture the document in a form the system can work with. Extract the fields without a person typing them. And decide which of those extractions a human needs to look at.

The third job decides the other two

At 95% field accuracy with no way to tell which 5% is wrong, you check all thousand invoices. If the system hands you the eighty it is unsure about, you check eighty.

Confidence per field, not per document

A document with a 96% total and a 51% PO number is not “84% confident”. It is sure of one thing and unsure of another, and one number throws that away.

Some conditions bypass the score

A quarantined document waits for a person, and an unreadable or refused attachment goes to review whatever the model returned. A line-item total that does not add up refuses the QuickBooks export.

Then the amount gets a second look

Every extracted bill is compared with that vendor's history at that property, your approval thresholds and the invoices you already hold. A finding never holds the bill or contacts the vendor.

Why the confidence signal matters more than the accuracy number

Consider the arithmetic. If a system processes a thousand invoices at 95% field accuracy and cannot tell you which 5% it got wrong, you have saved nothing but typing.

The useful measurement is not accuracy. It is whether the confidence signal is honest.

Collapsing a 96% total and a 51% PO number into “84%” throws away exactly the information you need. Where the document had readable text, each field also shows the line it was read from, so verifying a value is a glance rather than a hunt through the PDF.

The second look finds a fixed contract amount that moved, a price above the usual, a bill sitting just under a threshold, or a copy of one you already have. A top-tier duplicate is the single exception to “never holds the bill”: it cannot be exported until a person decides.

Capabilities

What runs on every document

A second look at the amount

Nine checks against your own numbers: a contract amount that moved, a price above the usual at this property, a bill just under a threshold, changed bank details on an invoice, and a possible duplicate against up to 200 of your recent invoices from that vendor.

A guessed value can never auto-approve

Every field records whether it was printed on the page, handwritten, inferred from context, or absent. An inferred value is capped below the score that lets a document pass without a person — enforced by a constant in the pipeline.

Reads what vendors actually send

PDFs, photos and images of paper invoices, Office documents, and invoices typed into the email body with no attachment. Attachments are parsed out of the MIME structure on the server.

Confidence per field

Vendor, number, dates, totals, tax and line items each carry their own score. A document whose vendor, number, date or total falls below 70% goes to review instead of on to an approver.

Extraction evidence

“Extraction evidence” shows where a value came from and, where the document had readable text, the line it was read from — so verifying a value is a glance rather than a hunt through the PDF.

Internal consistency checks

Line items are totaled and compared against the stated total. A variance is named on the document rather than discovered at close.

Review built for correction speed

The rendered document and the editable fields on one screen, with zoom, page navigation, a pinned action bar and previous/next document buttons that keep a long queue moving.

Multi-invoice emails

One email carrying several attachments produces one document per attachment, each extracted and metered separately. Several invoices inside a single PDF stay one document — we do not split pages.

How the second look sets its baselines

Baselines are kept per vendor AND property, and nothing has to be configured. It also catches several bills in a week that clear one threshold together.

An inferred value’s cap is a constant in the pipeline, not a policy someone can relax. So a value the model worked out rather than read cannot clear the gate on its own.

Why this one

What to actually check when comparing invoice processing tools

  • Does anything look at the amount after it is read, against that vendor's own history at that property — or does the tool stop at extraction?
  • Is confidence reported per field, or one score for the whole document?
  • Can a value the model inferred rather than read still be auto-approved?
  • Can you see the evidence a value was read from, without leaving the screen?
  • Are internally inconsistent extractions detected, or only individually implausible ones?
  • What routes to a human regardless of score — unmapped vendors, quarantined files, total mismatches?
  • Are your corrections tracked against the model's proposal, so you can see which vendors need a rule?
  • Does it handle an email carrying several invoices as several documents?
  • Can you test it on your own vendors' documents before paying, or only on a demo dataset?

Pricing

Test it on your own vendors first

The price is published, users are unlimited on every plan including the free one, you pay per document rather than per seat, and the free plan repeats: 25 processed invoices every month, forever, with no card.

Forward twenty-five genuinely awkward invoices from your worst-formatted vendors — that is a far better evaluation than any accuracy claim, including ours.

For a single document and no account at all, email one invoice to the Invoice Inbox Test and read the extracted fields in the reply.

Prices are in US dollars and exclude sales tax, which is added where it applies.

Paid plans renew automatically at the price shown — every month, or every 12 months on annual billing — until you cancel. Cancel any time in billing settings: the next renewal stops and you keep the period you have already paid for.

See all plans

Invoice processing — common questions

Looking for invoice OCR software?
inletAP reads invoices with an AI model, not an OCR engine. It reads the text a PDF already carries, and a vision model reads photos and image-only PDFs. Every field gets its own confidence score, so you check the fields it is unsure of.
Is this text recognition or AI invoice processing?
Both descriptions are partly right and neither is the interesting part. Reading characters off an image is solved; understanding that a particular number is the tax and another is an emergency call-out surcharge on a specific building is not. Two things separate a usable system. It knows when it does not know — inletAP scores confidence per field and routes uncertainty to a person rather than guessing, and a value it inferred rather than read is capped below the score that would let it through on its own. And it checks the amount it read against what that vendor billed for that property before.
What file formats can it process?
PDFs, photos and images of paper invoices, and common Office documents, plus invoices typed into the body of an email with no attachment at all. Attachments are parsed out of the MIME structure server-side, so nobody downloads and re-uploads anything. Documents arrive however they already do — a vendor-facing intake address, an upload, or a forward from the mailbox you use today.
What fields does it extract?
Vendor name, invoice number, invoice date, due date, total amount, tax amount, currency, and line items where the document has them, plus location clues that help identify the property. The property is then matched from those clues, your property aliases and the mailbox, the entity follows the property, the GL account is set by your rules, anything uncertain is corrected on review, and the QuickBooks class is resolved from the property at export.
How do I know whether an extracted value is right?
Each field carries its own confidence score and an “Extraction evidence” view showing where the value came from and, where the document had readable text, the line it was read from, with the original document displayed alongside the fields rather than in a separate tab. Line-item totals are also checked against the stated total and any variance is named explicitly.
What happens with a vendor whose layout it keeps getting wrong?
Your corrections are tracked against what extraction proposed, which makes the pattern visible. The fix is usually a default GL account on the vendor record, or one coding rule keyed on the sender — a one-off change rather than a recurring cost. Persistent low confidence on a mapped vendor keeps routing to review rather than silently degrading.
Can it detect duplicate invoices?
Yes — after extraction, and again during review. Every arriving invoice is scored against the invoices you already hold on nine rules and six guards, including the attachment fingerprint and a vendor / invoice-number / date / amount match, so the same invoice arriving twice by two different routes is linked. A top-tier match is held back from export until a person decides; everything else is a card on the document.

Keep reading

Start free and run your own invoices through it

25 invoices a month, unlimited users, no credit card. Send this month's vendor bills and watch each one get checked, coded and posted to your books.