Text extraction from images, from one photo to your order inbox

Text extraction from images turns the letters in a photo or a scanned page into text you can edit. For one image, a free OCR tool (optical character recognition) does it in a few seconds. For the purchase orders and supplier quotes your team retypes into the ERP every week, OCR is only the first step. You need software that reads each document into named fields and checks them against your records before a person approves the entry.

Derik Lawlis Founder, ThriveAI · builds AI systems inside manufacturers

Which option fits you

Start from what you need to do with the text once you have it. The table sorts the four common cases.

The first row is OCR on its own. The other three are field extraction, which is a larger job. OCR tells you that the page says 1,200. Field extraction tells you that 1,200 is the quantity on line three of a customer's purchase order, and the checks tell you whether it matches the quote you sent.

What free OCR gives you, and where it stops

OCR returns plain text. The converter on this site uses Tesseract, an open-source OCR engine, and its page lists what gets lost on the way. The headings, bold type and table borders of the original do not come across. A table comes back as lines of text, one row at a time, without its columns. Small marks such as decimal points are easy to misread, so 64.10 can come back as 64:10. Printed text reads best, and cursive handwriting usually does not.

On one document you catch those errors by eye. On fifty purchase orders a week, a misread decimal point becomes a wrong price in your ERP, because nobody rereads every line. At that volume you need field extraction, with checks that compare each value against your own records.

Reading documents into your ERP, in four stages

A system that reads documents into your ERP runs in four stages. Plan the third and fourth before you pick a model, because they decide whether the data can be trusted once it lands.

1. Capture

Documents arrive as email attachments, scans, phone photos and files in shared folders. The first stage puts them all in one queue and records where each one came from. It is also the place to flag documents that carry personal information, such as an employee's time card, so they only go to processing that is allowed to see them. The rules behind that flag are in where your documents go.

2. Extract

This stage turns a document into fields: the PO number, the ship-to address, and each line's part number, quantity and price. There are three common ways to do it, and you can mix them.

• A cloud service with a prebuilt model. Amazon, Microsoft and Google each sell one that reads invoices, described under cloud OCR services. It fits documents of the types the prebuilt model covers.
• A language model that reads into your fields. You describe the fields you want, and the model reads the page and fills them in. It fits documents whose layout changes with every sender. Anthropic's Claude API accepts a PDF directly, in a request of up to 32 MB, and processes each page as both text and an image.
• Open-source parsing, then a model. A parser such as Docling turns the page layout into text and tables on your own server, and a model fills the fields from that. It fits when the volume is high or the documents cannot leave your building.

3. Validate

Validation decides whether a field goes straight through or goes to a person. These four checks cover most business documents.

• Format. The required fields are present, a date is a date and a quantity is a number.
• Your records. The part number exists, the customer is on file, and the price matches the quote or the price last paid.
• Other documents. The supplier's invoice matches your purchase order and the receiving record.
• Duplicates. The same PO number has not arrived before.

A score for each field helps decide what a person sees. Amazon Textract's expense analysis, for example, returns a confidence score with each piece of data it detects. A field below the threshold you set goes to a reviewer.

4. Route

Routing writes the approved data into the system of record, which is usually the ERP or the accounting system. Where that system has an API, the entry is written directly once a person approves it. Where it has none, the system prepares the entry in a file your ERP can import, and a person imports it.

Test accuracy on your own documents

Published accuracy figures come from vendors and from outside tests on other companies' documents. To learn how a service reads your worst supplier's scan, or a purchase order with a change written in pen, run your own documents through it.

1. Pull 50 documents your team has already keyed into the ERP, and include the worst scans.
2. Run them through the option you are considering.
3. Compare each field with what your team keyed, and count the errors by field.
4. Look hardest at the fields that cost money when they are wrong: part number, quantity, unit price and date.

A field that comes back wrong on a few of the fifty needs a check or a reviewer in front of it. The error count by field tells you where to put them.

Cloud services, language models and open source

Cloud OCR services

Amazon Textract. Its Detect Document Text API reads text and handwriting from a document, according to the Textract pricing page. Its AnalyzeExpense API extracts fields from invoices and receipts, such as the invoice number, PO number, total and line items. AWS lists a Textract endpoint in Canada (Central). The same pricing page charges Detect Document Text and AnalyzeExpense per page, at rates that vary by region and monthly volume.

Azure Document Intelligence. Microsoft's prebuilt invoice model extracts key fields and line items from sales invoices, utility bills and purchase orders. Microsoft says it reads phone photos, scans and digital PDFs.

Google Document AI. Google lists an Invoice Parser and an Expense Parser among its processors. Its list of locations includes Montréal, where Enterprise Document OCR and Form Parser are available.

Before you commit to one, check the field list of its prebuilt model against the fields your ERP needs, and check which of its features are offered in the region you plan to use.

Language models

A language model that can read images takes a page image or a PDF and returns the fields you asked for, in a structured format your software can load. You define the fields once, and each new supplier layout is read against the same list.

Cost is billed by tokens. Anthropic's PDF documentation says a page typically uses 1,500 to 3,000 text tokens, depending on how dense it is, plus the tokens for the image of the page. Token prices are on Anthropic's pricing page. Where the model runs differs by provider, and where your documents go covers it.

Open-source software

Tesseract is an OCR engine released under the Apache 2.0 license, and it recognizes more than 100 languages. It is the engine behind the free converter on this site. Its documentation notes that better results often need a better image first.

Docling was started by the AI for knowledge team at IBM Research Zurich and is released under the MIT license. It parses PDFs, Office files and images into text and tables, includes OCR for scanned PDFs and images, and can run locally, including in air-gapped environments.

Both run on a server you control and carry no per-page fee. You take on installing, running and updating them.

Where your documents go

Most documents a manufacturer or distributor reads carry business information: part numbers, prices, quantities and the buyer's work contact details. Some carry personal information, such as an employee's time card or a ship-to address that is somebody's home. The rules below are about that second group.

Federal. The Office of the Privacy Commissioner of Canada's guidelines for processing personal data across borders say PIPEDA does not prohibit sending personal information to an organization in another jurisdiction for processing. The organization that sends it stays accountable for it. The guidelines also say to tell customers, in clear language, that their information may be processed in another jurisdiction and may be accessed there by courts, law enforcement and national security authorities.

Quebec. Quebec's private-sector privacy act, as amended by Law 25, sets these rules:

• Section 17 requires a privacy impact assessment before personal information is communicated outside Quebec, and a written agreement that takes the assessment into account.
• Section 3.3 requires a privacy impact assessment for any project to acquire, develop or overhaul an information system that involves personal information.
• Section 1 says Divisions II and III of the act, which include section 17, do not apply to information about a person's duties within an enterprise, such as their name, title, work address, work email and work phone number.
• Section 90.12 sets the maximum administrative monetary penalty for an enterprise at $10 million or, if greater, 2 percent of worldwide turnover for the preceding fiscal year. The Commission d'accès à l'information lists these penalties among the main changes brought by Law 25.

Claude on Amazon Bedrock. In the Canada (Central) Region, current Claude models are served only through cross-Region inference, with US or Global routing. AWS says a request from Canada (Central) is routed to a destination Region in the US or Global profile, where the inference processing may happen. All data at rest, including logs and knowledge bases, stays in the Canada (Central) Region. The Claude Sonnet 5 model card marks in-Region inference in ca-central-1 as not supported.

Claude on Anthropic's API. The inference location setting accepts only "us" and "global".

Cloud OCR services. Textract has an endpoint in Canada (Central), and Google lists Montréal as a Document AI location, both linked under cloud OCR services.

In a ThriveAI build. Your documents stay at rest on your own server in Canada, one dedicated machine for your business. Reading runs on that same server with an open-weight model, or through a frontier model under a written zero-data-retention control, with the model tier named in the contract. The tier is named because Anthropic retains prompts and outputs for 30 days on its Covered Models, and zero retention is not available for them.

This section describes the rules as published. It is not legal advice, and your counsel draws the conclusion for your own documents.

Two documents to start with

Supplier quotes and price files, at a distributor

Suppliers send quotes and price updates as PDFs, each in its own layout, and someone retypes them into the ERP or a spreadsheet. For one table at a time, the free PDF to CSV converter pulls the rows into a spreadsheet in your browser. The build reads each PDF into part number, description, unit price, quantity break and effective date. It checks each line against the price last paid and the previous price file from the same supplier, and flags every change above a threshold you set. A buyer approves the changes before anything is written to the ERP.

Customer purchase orders, at a manufacturer

Customer purchase orders arrive by email, each on the customer's own form. The build reads the PO number, ship-to address, terms and each line's part number, revision, quantity, price and due date. It checks each line against the quote you sent and the part master in the ERP, and flags a changed price, an unknown revision or a quantity outside the quoted breaks. A person approves the order before it is entered, and nothing is sent to the customer on its own.

Four mistakes to avoid

1. Choosing the model before the fields. Write down the fields your ERP needs and the checks each one must pass. Then pick the service that reads them best on your own test set.
2. Leaving out the reviewer. Every option on this page misreads some documents. Decide who reviews flagged fields, and how quickly, before the first document goes through.
3. Leaving privacy to the end. Which documents carry personal information, and where each one may be processed, decide how the capture stage is built. Settle both before you build it.
4. Stopping at the extracted text. The retyping only stops when approved fields are written into your ERP. Plan that write, and the approval in front of it, as part of the first version.

Sources

• Tesseract OCR on GitHub.
• Docling on GitHub.
• Amazon Textract pricing.
• Amazon Textract: analyzing invoices and receipts.
• Amazon Textract endpoints and quotas.
• Microsoft: Document Intelligence invoice model.
• Google Document AI: processor list.
• Google Document AI: supported locations.
• Anthropic: PDF support.
• Anthropic: pricing.
• Anthropic: data residency.
• Anthropic: data retention practices for Covered Models.
• AWS: Amazon Bedrock cross-Region inference in Canada (November 24, 2025).
• Amazon Bedrock: Claude Sonnet 5 model card.
• Office of the Privacy Commissioner of Canada: guidelines for processing personal data across borders.
• LégisQuébec: Act respecting the protection of personal information in the private sector.
• Commission d'accès à l'information: main changes brought by Law 25.