How does AI extract line items from invoices, and how is it different from OCR?

MakersHub’s AI, WiseVision, reads the whole invoice and returns every line as fields, where most AP tools’ OCR only looks for anchors such as “Invoice No.” and “Total.” Each line’s SKU, description, quantity, unit price and extended price comes out structured, with the header, totals and any custom field. It’s trained once per vendor where your books need a field read differently, then hands off to deterministic coding rules.

The difference shows on the bills that break OCR: a 68-page supplier statement that is really 60 bills and credits, a utility bill with no invoice number, a four-page bill whose total is on the last page, or a packing slip stapled behind the bill.

When this applies

You are evaluating AP automation and every vendor says it has AI. Your bills are not clean one-page invoices: supplier statements with dozens of invoices in one PDF, line items that must post to QuickBooks items or NetSuite locations, scanned stacks of paper, vendors who put the job name where the PO number should be. A tool that reads the total and the vendor is not enough; the work is at the line. For the wider picture, see how AI is used across accounts payable.

A founder at a boutique electronics manufacturer, on the built-in reader in his accounting software: 

"One of our ops staff overpaid one of our vendors by triple because there were packing slips attached to the invoice. It just tallied up all those numbers. We paid this bill that was supposed to be like $3,000, it was $10,000." 

The finance lead at a ten-site restaurant group on its current AP tool: 

"It's the same supplier using the same template that we've had for two years. So why is it not realizing that it's page four, not page one?"

How OCR-based invoice capture works, and where it stops

Optical character recognition is, in AWS’s definition, “the process that converts an image of text into a machine-readable text format.” AP tools built on it then apply a template or a key-value search: find the label "Total," take the number beside it; find "Invoice #," take the string beside it. Coordinate-based templates do the same by position on the page. Both work while a vendor's layout holds still and fail when it moves, when the label is missing, when a value sits on a later page, or when the document is really several documents.

Line items are the hard case. A template has to know where the table is, where each column begins and ends, and when the table continues onto the next page. Most AP tools therefore stop at the header and the total and code the bill to one expense category, which is why a distributor evaluating several of them said:

"Just about everyone just shows you the expense category level view. And they're like, of course we can do itemized. And then you go into it, and you actually can't do that."

How MakersHub's reads a bill.

1. The whole document is indexed, not searched. WiseVision is a vision model that returns every character on the page and its relationship to every other character, then uses language models to structure the result. MakersHub's co-founder describes the difference in data: an OCR pass on a typical supplier bill returns a handful of data points; 
   
   WiseVision returns on the order of 100,000, and indexes all of them. Nothing is anchored to a label, so a total on page four, a PO number printed as "Job," and an account number in a footer are all available.

2. Every line comes out as fields. SKU, description, quantity, unit price and extended price are extracted for each line, and MakersHub recalculates each line's total from quantity and price so a vendor's arithmetic error is visible. 
   
   Header fields include vendor, invoice number, date, terms (read in plain English and converted to a due date), PO number, bill-to and ship-to addresses, account number, tax and early-payment discount terms. A business can add custom extracted fields and columns, such as unit of measure or a contract price, and rules can read them.

3. A vendor's quirk is trained once. When a vendor prints the account number where the invoice number should be, or the quantity in a second column, the reviewer clicks the field on the bill image, selects the right value and confirms. From then on every bill from that vendor returns the field as instructed. Training is per vendor, not per invoice, and the trained fields are listed and editable.

4. Multi-invoice PDFs are split into invoice groups. A 68-page supplier statement is read as the 60-plus bills and credits it contains, each with its own lines, with the page range for each recorded. 
   
   A scanned stack of receipts becomes one record per receipt. Pages that are packing slips or shipping tickets are tagged as supporting documents, excluded from the totals and attached to the bill they support; a rule can do the tagging automatically for a vendor. A mis-split is fixed by dragging pages to merge or setting a page range.

5. The bill checks itself. The total is read independently of the lines, and the lines are summed. If they disagree, the bill cannot be approved until someone resolves it, and a native PDF viewer shows which line on the page each extracted value came from. A bill whose total the model cannot find is held with a "check total" tag rather than approved on a sum of the visible lines.

6. The vendor is matched, not assumed. The header name is matched to the vendor list in your accounting system. Where one name covers many accounts, such as one telecom vendor billing twenty locations, a rule combines vendor and account number or ship-to address to pick the right vendor record, location or entity.

7. Extraction hands off to deterministic rules. Reading is WiseVision’s job; interpretation belongs to Integration Mapping, MakersHub’s automated bill coding rules. When you confirm how a bill should be coded, the AI writes the rule, in a form you can read, audit and edit: if the vendor is X and the description contains "freight," post to Freight-In; if the ship-to address is this, the location is that. 

Rules run the same way every time, which is how MakersHub keeps AI coding from drifting.  Partial-phrase matching tolerates a vendor's line wraps and typos, and a quantity formula converts a unit of measure, such as eight-foot rods billed per rod and tracked per foot.

What it reads, by document type

Where it goes next. The structured bill feeds PO matching (line to PO line), approval routing (any field can be a condition) and the sync to QuickBooks Online, QuickBooks Desktop, NetSuite, Sage Intacct or Xero with items, Classes, Customer:Job, locations or dimensions carried on each line.

Proof

A commercial contractor on QuickBooks Desktop imported about 900 description-to-item rules in one upload and built fifty more, so vendors who bill without SKUs still post to the right QuickBooks items; a custom extracted column records the price billed next to the agreed price, and a rule overwrites the line with the contract price and notes the override. Its AP lead, a week in: 

"It's become very friendly for me and easy to do it. And while I'm building the data, it's more and more faster."

An industrial materials distributor on QuickBooks Online with about 300 bills a month, several of them multi-invoice PDFs that "should be ten," saw a 68-page statement split into 63 bills and credits and matched against open POs at the line level. Its owner, after evaluating other tools:

"It does everything we want it to do. It's been hard to find."

An accounting firm principal who once ran an AP operation on template-based capture, after watching MakersHub hold a bill with an injected $10 line because the lines no longer summed: 

"This system is very nice because it's not dependent upon a consistent format." 

A specialty materials manufacturer on QuickBooks had rods billed per eight-foot length and tracked per foot; a custom unit-of-measure column and a quantity formula now post 8 rods as 64 feet.

What it doesn't do

Legibility is the limit: if a person can read the document, MakersHub can read it, and a scribbled note a person cannot read will need a correction. Two-column layouts that place two tables side by side on one page are rare and are handled by splitting the page or asking the vendor for a PDF. 

Packing slips and shipping tickets in a bundle are read as their own records until they are tagged as supporting documents or a rule tags them for that vendor. And the line-to-page viewer works on PDFs rather than on photographed images today.

Frequently asked questions