OCR in accounts payable turns a scanned or PDF invoice into machine-readable text, and the software worth buying adds a layer on top that maps that text into typed fields: vendor, invoice number, due date, tax, totals and line items. Plain OCR returns characters and coordinates, and nothing in that output tells you which number is the total. Priced per page, the engines AP teams actually buy run from $1.50 to $70 per 1,000 pages, and the most common configuration mistake costs seven times more than the correct one. Upload an invoice below and see typed fields and line items instead of raw text.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
OCR is a solved problem. Reading an invoice is not. The gap between the two is where AP teams lose the hours they expected to save, and it shows up in four specific places.
A raw OCR call returns every word on the page with a bounding box and a confidence score. It does not label the vendor, and it cannot tell the invoice total from the subtotal, the freight line or the prior balance. Someone or something still has to decide, and if that decision is a set of regular expressions and coordinates, you have just bought a template library to maintain.
Published third-party testing on scanned invoices found Google Document AI at 82% on header fields and 40% on line items, and a GPT-4o pipeline at 98% on header fields and 57% on line items. Header accuracy is the number vendors quote. Line item accuracy is the number that decides whether a three-way match runs without a human.
AWS Textract charges $10 per 1,000 pages for Analyze Expense, its purpose-built invoice API, and $70 per 1,000 for Forms plus Tables. Teams reach for Forms plus Tables because an invoice looks like forms and tables. At 20,000 pages a month that choice costs an extra $14,400 a year and returns less usable invoice structure.
QuickBooks Online extracts vendor, date and total from an uploaded bill and stops there. It does not read line items, QuickBooks Desktop has no OCR at all, and spreadsheet invoice import caps at 100 invoices or 1,000 rows per file. Those are product limits, so no amount of process change gets past them.
The useful test is not whether a tool can read the page. It is whether the output can be posted without opening the original PDF. That takes six things.
Vendor, invoice number, invoice date, due date, PO number, subtotal, tax, freight and total come back as named values with types, so they can be written straight into an AP record instead of parsed a second time.
Each row of the invoice table returns as a structured line: description, quantity, unit price, extended amount, and the tax or GL hints when the invoice carries them. This is what three-way matching needs and what plain OCR cannot supply.
AP arrives in batches: a folder from a scanner, a month of vendor PDFs, an export from a mailbox. Send the whole set through in one pass rather than clicking upload 300 times.
Extraction is only useful if it lands somewhere. Excel and CSV for review and reconciliation, JSON for an integration, and QBO for QuickBooks import, all from the same extraction.
Per-word confidence tells you nothing actionable. Per-field confidence lets you auto-post the clean invoices and route only the uncertain ones to a person, which is the entire mechanism behind touchless processing rates.
Per-page rates that do not change when a vendor changes their template, and no per-layout setup fee. Our published plans are $49 and $149 a month, or $288 and $888 billed annually, which works out to $9.60 and $7.40 per 1,000 pages.
Four steps that take a pile of vendor PDFs to posted AP entries, with the decisions that matter called out at each one.
Scanned paper, emailed PDFs, vendor portal downloads and photographed invoices all have to reach the same place before anything can be automated. Decide now whether that queue is a folder, a mailbox or an API endpoint.
Tip: Photographed invoices are worth accepting. Field crews and drivers will send them regardless of policy.
Run the batch through an engine that returns typed fields, not text. If you are building on a cloud API, choose the invoice-specific model rather than a generic forms-and-tables call, both for accuracy and because it is a fraction of the price.
Tip: Test on your ten worst vendor layouts, not on a clean sample invoice. The worst ten set your real touchless rate.
Set a threshold and let clean invoices pass. The point of field-level confidence is that a person opens 5% of invoices instead of 100% of them, and the ones they open are the ones that were genuinely ambiguous.
Tip: Track which vendors trip the threshold. It is usually a handful of templates, and fixing those moves the rate more than tuning the model.
Push the result out as Excel or CSV for review, JSON for a direct integration, or QBO to bring bills into QuickBooks. Keep the source PDF linked to the record so an auditor can get back to the original in one click.
Tip: Retention rules apply to the extracted data as well as the PDF. Keep both for the same period.
Every rate on this page was read from the vendor rate card or billing API on 26 August 2026. Accuracy figures come from a published third-party test, and the labor benchmarks from APQC. Nothing here is estimated except where the arithmetic is shown.
You need a per-invoice number your CFO will accept and a defensible view of what the software costs against what the process costs. Both are on this page with the sources named.
You process invoices for several entities or clients and cannot maintain a template per vendor. What you need is line item extraction that survives layout changes and exports that your ledger accepts.
You are choosing between Textract, Document Intelligence, an LLM and a hosted extraction API. The rate cards and the configuration traps are laid out below so the build estimate is not a guess.
Above roughly 20,000 pages a month the commitment tiers and the touchless rate matter far more than the headline per-page price. The tier tables below are the ones the vendors do not publish.
Every rate below was read from the vendor's own rate card or billing API in August 2026, not from a summary article. Figures are per 1,000 single-page US Letter invoices in a US region, and the LLM rows include output tokens, which is where most of that bill actually lands.
| Engine | What it returns | Per 1,000 pages | Verified |
|---|---|---|---|
| Alibaba qwen-vl-ocr (Global scope) | Raw text, no typed fields | $0.16 | Aug 2026 |
| DeepSeek vision endpoint (off-peak) | Raw text, heavily downsampled | $0.70 | Aug 2026 |
| AWS Textract Detect Document Text | Words and bounding boxes | $1.50 | Aug 2026 |
| Azure Document Intelligence Read | Words and bounding boxes | $1.50 | Aug 2026 |
| OpenAI GPT-5 nano | Whatever your prompt asks for | $0.47 | Aug 2026 |
| Mistral OCR 4.1 | Structured text and layout | $4.00 | Aug 2026 |
| AWS Textract Layout | Reading order and blocks | $4.00 | Aug 2026 |
| AWS Textract Analyze Expense | Typed invoice fields and line items | $10.00 | Aug 2026 |
| Azure Document Intelligence prebuilt-invoice | Typed invoice fields and line items | $10.00 | Aug 2026 |
| Azure Document Intelligence custom extraction | Your own trained schema | $30.00 | Aug 2026 |
| AWS Textract Forms plus Tables | Key-value pairs and table cells | $70.00 | Aug 2026 |
| InvoiceExtractor Plus, annual | Typed fields, line items, Excel, CSV, JSON, QBO | $7.40 | Our published pricing |
Two rows in that table do the same job for AP. Two other rows cost 7 times more and do it worse. That gap is the single most expensive mistake in accounts payable OCR, and it is worth understanding before you sign anything.
Textract is not one price. It is fifteen, and AWS bills you for the combination of features you request on each call. Teams building AP capture very often reach for Forms and Tables, because an invoice obviously has key-value pairs and an obvious table. AWS publishes a purpose-built invoice API instead, and it is one seventh the price.
| Textract feature | First 1M pages | Above 1M |
|---|---|---|
| Detect Document Text | $1.50 | $0.60 |
| Layout | $4.00 | $3.00 |
| Signatures | $3.50 | $1.40 |
| Analyze Expense (invoices and receipts) | $10.00 | $8.00 |
| Tables | $15.00 | $10.00 |
| Queries | $15.00 | $10.00 |
| Queries plus Tables | $20.00 | $15.00 |
| Custom Queries | $25.00 | $15.00 |
| Analyze ID | $25.00 (first 100K) | $10.00 |
| Custom Queries plus Tables | $30.00 | $20.00 |
| Forms | $50.00 | $40.00 |
| Forms plus Queries | $55.00 | $45.00 |
| Forms plus Custom Queries | $65.00 | $50.00 |
| Forms plus Tables | $70.00 | $55.00 |
| Analyze Lending | $70.00 | $55.00 |
| Forms plus Custom Queries plus Tables | $80.00 | $60.00 |
Read this before you pick a Textract feature set. An AP department running 20,000 invoice pages a month pays $200 on Analyze Expense and $1,400 on Forms plus Tables. That is $14,400 a year of pure configuration error, and Analyze Expense returns more usable structure for invoices because it emits normalized field types rather than raw key-value guesses. Textract's rate card was last republished on 29 October 2024 and we re-read the AWS Price List API on 26 August 2026 to confirm nothing has moved since.
Azure's public pricing page has rendered a dash instead of a price for most of 2026, so these come from the Retail Prices API for the eastus region, read on 26 August 2026. The meter an AP team wants is prebuilt, which covers the invoice, receipt and ID models.
| S0 pay-as-you-go meter | Per 1,000 pages |
|---|---|
| Read (plain OCR) | $1.50, then $0.60 above 1M |
| Document classifier | $3.00 |
| Add-on features | $6.00 |
| Prebuilt invoice, receipt, ID | $10.00 |
| Layout | $10.00 |
| Query fields | $10.00 |
| Custom extraction | $30.00 |
| Custom generative | $30.00 |
| Standalone query pages | $200.00 |
| Training | $3.00 per hour |
Azure gives no batch discount. Every batch meter is priced identically to its synchronous equivalent, which is unusual: Mistral, OpenAI and Anthropic all take 50% off batch work. What Azure does offer, and never shows on the pricing page, is a commitment tier. For a mid-sized AP department these are worth real money.
| Prebuilt commitment tier | Azure-hosted | Effective per 1,000 | Container | Effective per 1,000 |
|---|---|---|---|---|
| 20,000 pages per month | $190/mo | $9.50 | $152/mo | $7.60 |
| 100,000 pages per month | $900/mo | $9.00 | $720/mo | $7.20 |
| 500,000 pages per month | $4,000/mo | $8.00 | $3,200/mo | $6.40 |
| 1,000,000 pages per month | $7,500/mo | $7.50 | $6,000/mo | $6.00 |
A 20,000 page commitment already beats pay-as-you-go by 24%, and 20,000 pages a month is an ordinary mid-market AP department, not an enterprise. Overage is billed at the same effective rate as the tier you bought, so exceeding a commitment is not punished. If you train a custom model instead of using the prebuilt invoice model, the commitment tiers start at $540 a month for 20,000 pages ($27 per 1,000) and fall to $18,000 a month for a million pages ($18 per 1,000), and the connected-container versions of those are $459 and $15,300. Read has its own tiers starting at $375 a month for 500,000 pages, which works out to $0.75 per 1,000.
Vendors quote 99% and mean character accuracy on clean text. AP cares about field accuracy, and specifically about line items, because that is what feeds a three-way match. A published third-party test by BusinesswareTech in January 2025, run on scanned documents rather than clean digital PDFs, separated the two and the results are not flattering.
| Engine | Header field accuracy | Line item accuracy | Gap |
|---|---|---|---|
| GPT-4o with third-party OCR | 98.0% | 57.0% | 41 points |
| Azure Document Intelligence | 93.0% | 87.0% | 6 points |
| Google Document AI | 82.0% | 40.0% | 42 points |
| AWS Textract | 78.0% | 82.0% | 4 points against |
The pattern matters more than any single number. The engine that reads header fields best is second worst at line items. If your AP process stops at vendor, invoice number, date and total, almost anything works. If you match invoice lines to purchase order lines, header accuracy tells you nothing useful and you have to test line items on your own documents before you commit.
Software price is the small number. APQC benchmark data puts the fully loaded cost to process a single invoice at roughly $2 in top-quartile organizations, about $6 at the median, and above $10 in the bottom quartile. At the median, a department handling 5,000 invoices a month is spending around $30,000 a month on the process itself. Against that, a $10 per 1,000 pages extraction bill is $50.
| Quartile | Cost per invoice | 5,000 invoices per month |
|---|---|---|
| Top quartile | about $2 | about $10,000 |
| Median | about $6 | about $30,000 |
| Bottom quartile | $10 and above | $50,000 and above |
That is why arguing about $4 versus $10 per 1,000 pages is usually the wrong argument. The number that moves is how many invoices a person has to touch, and that depends almost entirely on whether the extraction gives you line items you can trust without opening the PDF.
| Question | Plain OCR | AI invoice extraction |
|---|---|---|
| What you get back | Characters, words, coordinates | Named fields with types and values |
| Handles a new vendor layout | Yes, but you map it yourself | Yes, no template needed |
| Templates or zones to maintain | Usually one per layout | None |
| Line item tables | Cells without meaning | Quantity, description, unit price, amount |
| Knows which total is the total | No | Yes |
| Confidence per field | Per word only | Per field, so you can route exceptions |
| Typical rate per 1,000 pages | $1.50 | $7 to $10 |
Plain OCR is not obsolete. It is the right tool when you need searchable archives, redaction, or full text indexing. It is the wrong tool when the output has to become an AP entry, because everything between the characters and the entry is work someone still has to do.
Most AP teams already have some capture in the tools they own, and it is worth knowing its ceiling before you buy anything else. QuickBooks Online reads header fields from an uploaded bill: vendor, date, total, sometimes the invoice number. It does not read line items. QuickBooks Desktop has no OCR at all. Spreadsheet invoice import in QuickBooks caps at 100 invoices or 1,000 rows per file. Those are hard limits, not settings, and they are the reason most growing AP departments end up with a separate extraction step that feeds the accounting system rather than relying on the one inside it.
OCR in accounts payable is optical character recognition applied to vendor invoices, converting a scan or PDF into machine-readable text so the data can be captured without typing. On its own OCR returns words and their positions. AP software adds a layer that maps those words into named fields such as vendor, invoice number, due date, tax and line items.
The invoice is captured from a scanner, mailbox or upload, OCR converts the image into text with coordinates, and an extraction model identifies which values belong to which invoice field. Confidence scores route uncertain invoices to a reviewer while clean ones pass through. The structured result is then exported to Excel, CSV, JSON or straight into the accounting system.
For header fields, yes: published third-party testing on scanned invoices put Azure Document Intelligence at 93% and a GPT-4o pipeline at 98%. Line items are a different story, with the same test showing 87% and 57% respectively, and Google Document AI at 40%. Test line item accuracy on your own worst vendor layouts before committing to any engine.
Plain OCR runs about $1.50 per 1,000 pages on AWS Textract and Azure Document Intelligence. Typed invoice extraction, which is what AP actually needs, is $10 per 1,000 pages on both AWS Analyze Expense and the Azure prebuilt invoice model. Our own annual plans work out to $9.60 and $7.40 per 1,000 pages.
For most US AP departments the deciding factor is line item accuracy and export format, not the OCR engine underneath. Azure Document Intelligence leads published line item accuracy at 87%, AWS Analyze Expense is the cheapest way to get typed invoice fields on AWS, and a hosted extraction tool wins when you do not want to build and maintain the mapping and review layer yourself.
OCR converts pixels into characters. AI extraction decides what those characters mean, returning named fields with types and per-field confidence rather than a page of text. In practice OCR needs a template or a set of rules per vendor layout, and AI extraction reads a layout it has never seen before, which is why it handles new vendors without setup.
QuickBooks Online can read an uploaded bill and fill in header fields such as vendor, date and total, but it does not extract line items. QuickBooks Desktop has no OCR at all. Spreadsheet invoice import caps at 100 invoices or 1,000 rows per file, so AP teams needing line item detail usually extract separately and import the result.
Modern engines handle clean scans well and handwriting unevenly. Handwritten amounts and signatures are the weakest case, and scan quality matters more than most buyers expect: a 150 DPI scan and a 300 DPI scan of the same invoice can produce noticeably different field accuracy. Route handwritten documents to review by default rather than trusting confidence scores alone.
No, it changes what they touch. APQC benchmarks put the fully loaded cost to process an invoice at about $2 in top-quartile organizations and about $6 at the median, and most of that is human handling. Extraction removes the keying and the lookup, leaving exceptions, vendor disputes and approvals, which are the parts that need judgment.
Invoice OCR software compared, and what to test before you buy.
Getting quantity, unit price and amount off every invoice row.
Send a folder of invoices through in one pass.
The complete Textract rate card, including the Analyze Expense trap.
Live Azure Document Intelligence rates and commitment tiers.
Getting extracted invoices into QuickBooks without retyping.
A typed invoice API with line items and no prompt engineering.
Where template OCR stops and model-based extraction starts.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
per month
$288 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 2,500 pages |
| Pro AI Best accuracy | 500 pages |
Scale invoice extraction across your whole team with automation.
USD
per month
per month
$888 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 10,000 pages |
| Pro AI Best accuracy | 2,000 pages |
Enterprise‑grade invoice extraction, security, and controls.
USD
per month
per month
$ charged today
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |