How to extract data from PDFs and scans with AI

By the end of this guide you will have a repeatable way to extract data from PDFs and scans with AI: documents go in, and named fields such as supplier, date, total or tax come out as a spreadsheet row or JSON, with a person checking anything uncertain. The job is done by a document extraction tool, which reads the page layout, pulls out the fields you define and passes the result to your accounting, ERP or spreadsheet.

The real decision is how much of the plumbing you want to build. Some tools are finished platforms with a review screen, some are APIs for developers, and some are general automation or document-chat tools. This guide compares seven from our Automation and agents category so you can match the tool to your volume and your skills.

Key takeaways

• Define the fields you need first, test a tool on a small sample of your real documents, then add a human review step for low-confidence values before anything flows into your books.
• The main decision is platform versus API: Rossum offers a review screen for finance teams, while Mindee and Reducto return structured data to code you write.
• Costs differ sharply: Rossum starts at $18,000 per year with a one-year minimum, while Reducto and Multimodal charge per page processed.
• Extraction is a draft. Check totals, dates and tax fields against the source document, especially for financial or legal records.

What you need before you start

• A sample of real documents. Gather 10 to 20 PDFs or scans, including the messy ones: rotated pages, poor scans, different layouts from different suppliers.
• A list of target fields. Write down exactly what you want out, for example supplier, invoice date, total and tax, and where the data goes next.
• An account or an API key. Platforms need a sign-up or a sales conversation. API tools need someone comfortable sending a file and reading JSON.
• Time. A first test on a sample takes under an hour with a web tool. A full pipeline into your accounting system takes longer.
• Difficulty. Moderate. Easy with a hosted tool and a sample, harder if you wire an API into your own systems.
• Permission. If the documents contain personal data such as identity documents, confirm you are allowed to process them and where the vendor stores them.

Which kind of AI tool fits this job

Pick by who will run the process and how many pages you handle.

A document automation platform with human review. Rossum captures data from invoices, sales orders and similar business documents that arrive by email, API or manual upload. Anything uncertain goes to a validation screen where a person confirms or corrects values. Higher plans add master data matching and duplicate detection, plus connections to systems such as SAP, Coupa, Workday and Oracle. It suits accounts payable teams and carries a one-year minimum contract.

An extraction API for developers. Mindee turns invoices, receipts, identity documents, contracts and forms into structured JSON, optionally with a confidence score for each value. Reducto offers Parse, Extract, Split and Classify APIs, with schema-based extraction that includes citations and bounding boxes. Both return structured data for your developers to pass into your own software.

An industry platform. Multimodal is aimed at banks and credit unions, with an Unstructured AI extraction layer behind a serverless API and playbooks for lending and compliance work. It is a poor fit if you are not a financial institution.

General automation or document tools. Stepper is a workflow builder whose example workflows include invoice extraction from PDFs, useful if extraction is one step in a longer chain. FileDrop adds OCR for scans and PDFs, from its Business plan, inside a Google Workspace-centred suite. Elephas is different again: it lets you ask questions of your own documents on an Apple Silicon Mac, with cited answers, rather than producing a repeatable data feed.

Step 1: Define the fields and the output you need

You finish this step with a one-page spec that any tool can be tested against.

1. List every field you want, such as supplier, date, total and tax.
2. Decide the output format: a spreadsheet, JSON for an app, or a direct entry in your accounting or ERP system.
3. Note which fields are critical (amounts, dates, account numbers) and need stricter checks.
4. Estimate monthly page volume, because most pricing is tied to pages.

Check it worked by handing the spec to a colleague and confirming they would know what a correct result looks like.

Step 2: Prepare a test set of documents

A small, honest sample tells you more than any feature list.

1. Pick a mix of clean digital PDFs and poor scans.
2. Include your most awkward layouts, such as multi-page documents.
3. Remove or mask anything you are not permitted to upload to a trial account.
4. Keep a spreadsheet of the correct answers for a few documents so you can score the output.

Check it worked by confirming that your answer sheet covers every field in the spec for at least a handful of documents.

Step 3: Choose a tool that matches your volume and skills

You end with one or two tools to trial.

1. If a finance team needs a review screen and system connections, look at Rossum. Check the stated limits first: billing is per page processed, including duplicates, and extraction defaults to the first 32 pages of a document.
2. If a developer will integrate, trial Mindee or Reducto. Mindee has a 14-day free trial covering 200 pages, and Reducto's Standard tier includes $150 in free usage.
3. If you are a bank or credit union, ask Multimodal for a quote, since pricing beyond its Unstructured AI per-page rate is quoted on request.
4. If extraction is only one step of a larger process, look at Stepper's workflow editor. Its Free Plan includes 200 steps per month.
5. Use the comparison table further down to line up pricing and limits.

Check it worked by confirming that the trial or free allowance covers your whole test set.

Step 4: Run the sample and review field by field

Run your test set through the tool, then compare each result with your answer sheet.

1. Upload or send the documents. Rossum accepts email, API or manual upload, while API tools such as Mindee take a file in a request.
2. Define or select the model or schema for your fields. Mindee lets you pick or build a model, and Reducto's Extract follows a schema you define.
3. Compare each extracted field with your answer sheet, noting which document types fail.
4. If the tool gives confidence scores, see whether low scores line up with the wrong values. Mindee lists per-field confidence scores as optional, and a page with a confidence score uses 1.5 credits instead of one.

Check it worked by calculating how many fields were correct and which kinds of document caused the misses.

Step 5: Add a human review step for uncertain values

No extraction is safe to trust blindly, so route doubtful results to a person.

1. Decide a rule, for example that low-confidence fields or totals above a set amount are always reviewed.
2. In Rossum, uncertain documents go to a validation screen. In an API-based setup, you build that screen or spreadsheet check yourself.
3. Let the reviewer correct values and note recurring errors so you can adjust the schema.
4. Keep duplicate detection in mind. Rossum lists it on higher plans, and billing counts duplicates.

Check it worked by sending a deliberately wrong document through and confirming it reaches a reviewer.

Step 6: Connect the output to where the data lives

Extracted data only saves time if it lands in the right place without retyping.

1. Choose the destination: ERP, accounting software, a Google Sheet or a database.
2. Use the platform's own connections where they exist. Rossum names SAP, Coupa, Workday, Oracle and QuickBooks.
3. For custom destinations, send the JSON through an API or webhook, or through an automation tool such as Stepper.
4. Run a small batch end to end before switching over the full volume.

Check it worked by finding a test document's values in the destination system and matching them to the source.

Step 7: Monitor cost and accuracy over time

New suppliers and layouts will appear, so keep an eye on both.

1. Track pages processed against your plan. Mindee credits expire at the end of each subscription period and do not carry over.
2. Re-check a small random sample every month against the source documents.
3. Update your schema or model when a new document type appears.

Check it worked by comparing this month's error rate and spend with your first test.

Tools for this job compared

Tool	Good fit for	Pricing	Stated limits
Rossum	Accounts payable teams capturing invoices and orders	Paid, from $18,000/yr (checked 24 Oct 2026)	One-year minimum contract; billed per page including duplicates; first 32 pages by default
Mindee	Developers adding OCR through a web or API setup	Free trial, paid from €44/mo (checked 24 Oct 2026)	Credits expire each period; 20% overage surcharge by default; data deleted 30 days after cancelling
Reducto	API, web or self-hosted parsing with schema-based extraction	Free trial, paid from $10, usage-based (checked 24 Oct 2026)	Dense extractions over 100 fields per page may add a surcharge; Deep Extract excluded from batch discount
Multimodal	Banks and credit unions using an API	Contact sales, from $10, usage-based (checked 24 Oct 2026)	Aimed at financial institutions; pilots take about 4-6 weeks; other pricing on request
Elephas	Cited document Q&A on Apple Silicon Macs	Free trial, paid from $15/mo (checked 24 Oct 2026)	macOS only; files capped at 200 MB; no free plan beyond the 7-day trial
FileDrop	Google Workspace users needing OCR and upload forms	Freemium, paid from $7.42/mo (checked 24 Oct 2026)	OCR is available from the Business plan; Free plan allows 50 uploads a month
Stepper	Workflows where extraction is one step among many	Freemium, paid from $19/mo (checked 24 Oct 2026)	Free Plan: 200 steps per month; workflows pause when credits run out; Team Plan coming soon

Prices and plan limits change, so confirm the current terms on each vendor's site before you commit.

Common mistakes to avoid

• Testing only on clean files. Vendors demonstrate tidy documents, and your real mix includes skewed scans and odd layouts. Fix it by building your test set from your worst examples as well as your best.
• Trusting totals without a check. A single misread digit in an amount can pass straight into your books. Fix it by reviewing critical fields, or all low-confidence ones, before export.
• Ignoring how pages are billed. Rossum bills per page processed, including duplicates, and Mindee counts a confidence-scored page at 1.5 credits. Fix it by modelling your real volume against each plan before you sign.
• Locking in before testing. Some contracts run for a minimum term. Fix it by using free trials and sample allowances first, and by asking about the contract length early.
• Uploading sensitive documents carelessly. Identity documents and financial records carry privacy obligations. Fix it by confirming consent and data handling, and by asking a vendor about private deployment where it is offered.

How we put this together

This is an evidence-based guide built from each tool's listing in our directory, which is compiled from the vendor's own site and pricing page. It is not a review based on using the tools. We don't rank tools, and nobody paid to be included. Some outbound links may be affiliate links; see our disclosure.

Frequently asked questions

Can I extract data from PDFs with AI for free?

Partly. Mindee has a 14-day free trial covering 200 pages, Reducto's Standard tier includes $150 in free usage, and Stepper's Free Plan includes 200 steps per month. These suit testing rather than steady production volumes, so check each allowance against your page count.

How accurate is AI data extraction from PDFs?

It depends on scan quality, layout variety and the fields you ask for, and none of the listings in our directory publish a guaranteed accuracy figure. Test on your own documents and keep a person in the loop for amounts, dates and identity details, especially for financial or legal records.

Does AI work on scanned documents and not just digital PDFs?

Several tools list OCR for scans. FileDrop offers OCR for scans and PDFs from its Business plan, Elephas lists OCR for scanned PDFs and images, and Reducto lists multilingual OCR and automatic page rotation. Run a poor scan through your shortlisted tools before deciding.

Which tool is good for extracting data from invoices?

Rossum and Mindee both list invoices as a core use. Rossum is a platform with a review screen and system connections, while Mindee returns fields as JSON for developers. Beyond that, the choice comes down to whether you need a finished workflow or an API.

Can I keep my documents private?

It varies by tool. Elephas indexes documents on your own computer and hides personal details before a question goes to a cloud model. Reducto lists self-hosted and VPC deployment for Enterprise customers. Read each vendor's data terms before uploading confidential documents.

Compare more options in our Automation and agents category.