Buy document and form data for AI training
Last updated How we research
In short
Document and form data is licensed collections of real business documents, such as invoices, contracts, tax forms, reports and scanned records, used to train document understanding, extraction and drafting models. Mercor's published listings value tax returns, workpapers and bookkeeping records at $25,000 to $865,000 per accounting firm, and closed legal files at $50,000 to $1.2 million per law firm.
Key takeaways
- AI teams use document and form data for (1) document understanding and OCR models, (2) field extraction from forms and invoices and (3) contract review and drafting models.
- Usually priced per document or page for labeled sets, per archive for company records.
- Common delivery formats: Original PDFs and scans in TIFF or PNG, native DOCX and XLSX files and extracted text and layout in JSON.
- Check before you license document and form data: (1) US: IRC Section 7216 limits on tax return information, (2) US: HIPAA de-identification for health records and (3) US: GLBA notice and opt-out rules for financial institutions' customer records.
- Every seller we shortlist passes 8 vetting checks before you see a sample.
Document and form data: buyer's spec
- Used for
- Document understanding and OCR models
- Field extraction from forms and invoices
- Contract review and drafting models
- Agent training on document-heavy workflows
- Delivery formats
- Original PDFs and scans in TIFF or PNG
- Native DOCX and XLSX files
- Extracted text and layout in JSON
- Field-level labels and key-value pairs
- Usually priced
- Per document or page for labeled sets, per archive for company records
- Check before you sign
- US: IRC Section 7216 limits on tax return information
- US: HIPAA de-identification for health records
- US: GLBA notice and opt-out rules for financial institutions' customer records
- Personal and financial data removed from text and images
- Client and counterparty confidentiality
- Scans legible and complete
- Labels checked against source documents
- Document counts and types match the manifest
Why AI teams work with us
- 1
Every seller of document and form data we shortlist has passed eight vetting checks, from authority to sign through to a re-identification test.
- 2
Each shortlist comes with a vetting file per seller: rights evidence, provenance record, de-identification method, data manifest and proposed license terms.
- 3
Samples are drawn at random across the archive and labeled with the sampling method, so you can judge the whole package rather than its best rows.
- 4
You contract directly with the seller, and we help agree acceptance tests, staged payments and remedies before any data is delivered.
What is document and form data?
Document and form data is the paperwork of real organizations in its original layout: invoices, purchase orders, contracts, tax forms, insurance claims, inspection reports and scanned records. Layout matters as much as text, because models must learn where information sits on a page.
Sources include accounting and law firms, insurers, logistics companies handling shipping documents, manufacturers with quality records and organizations with large scanned archives.
What do AI teams use document data for?
AI teams use document data to train models that read, extract from and produce business documents. The same archive can feed several uses.
| Use | What the data needs |
|---|---|
| OCR and layout | Scans and PDFs at original quality, across varied templates |
| Field extraction | Labeled key-value pairs checked against the source |
| Contract review | Executed agreements with clause-level annotations |
| Drafting | Final documents paired with drafts and instructions |
| Workflow agents | Documents linked to the decisions they triggered |
How much is document data worth?
Document data has no list price, but published payouts for document-heavy company archives give a guide. Mercor's listings, compiled by ddiy.co in October 2026, include these.
- Accounting and tax firms: $25,000 to $865,000 for tax returns, workpapers and bookkeeping records
- Law firms: $50,000 to $1.2 million for complete files from closed matters and completed deals
- Consulting firms: $150,000 to $300,000 for client deliverables, models and the drafts behind them
- Manufacturers: $100,000 to $1.5 million for production, maintenance and inspection records
Mercor's listings are payouts to the firms, not prices to buyers. Sellers in these sectors can read our pages for accounting and tax firms, law firms and insurance companies.
What legal rules apply to document data?
Document data is governed by the rules for whatever the documents contain, and those rules differ by record type.
- Tax records: in the US, IRC Section 7216 prohibits tax return preparers from knowingly or recklessly disclosing or using tax return information
- Health records: HIPAA permits de-identification by expert determination or by removing 18 identifiers under 45 CFR 164.514
- Financial records: under GLBA, at 15 USC 6802, a financial institution may not disclose nonpublic personal information to a nonaffiliated third party unless the consumer has had notice and a chance to opt out
- Payment records: PCI DSS states that sensitive authentication data, including card verification codes, must not be stored after authorization
- Personal data in the UK and EU: UK GDPR and EU GDPR apply unless the data is anonymized
We are a broker, not a law firm, so take independent advice on the rules that apply to the records you plan to license.
How should document data be delivered?
Document data should be delivered in original form, as PDFs or scans at the resolution they were captured, with native files where they exist. Extracted text and layout should come alongside in JSON, keyed to page and position, so you can check the OCR against the image.
Labels for field extraction should name the field, its value and its location on the page, with the labeling guide attached. Ask for counts by document type and template, because a set dominated by one template teaches a model little about variety.
Handwritten forms need their own check. Ask whether handwriting was transcribed by people or by a model, and compare a sample of transcriptions with the scans.
What should you check before buying document data?
Before buying document data, check redaction on the images as well as the extracted text. A name removed from the OCR output but still visible in the scan has still been disclosed.
Check legibility on a random sample, confirm document types and counts against the data manifest and verify labels against source documents. Our guide to de-identification for AI training data covers redaction methods. Book a free consultation to source vetted document archives.
Frequently asked questions
What documents are most useful for document AI training?
How are scanned documents redacted for AI training?
Can accounting firms license client documents?
Should document data include drafts as well as final versions?
Sources
- How to sell company data to AI companies (2026 guide), ddiy.co, accessed November 2026
- Section 7216 frequently asked questions, Internal Revenue Service, accessed November 2026
- Guidance regarding methods for de-identification of protected health information, US Department of Health and Human Services, accessed November 2026
- 15 U.S. Code section 6802: obligations with respect to disclosures of personal information, Legal Information Institute, Cornell Law School, accessed November 2026
- Why is storage of sensitive authentication data after authorization not permitted?, PCI Security Standards Council, accessed November 2026
For AI teams
Ready to source?
Book a consultation to brief us. We shortlist vetted sellers, each with a vetting file covering rights evidence, provenance and a representative sample.
Related guides
- De-identification for AI training data
Legal and compliance
- How to buy AI training data
Buying data
- AI training data due diligence checklist
Buying data
- How to evaluate a dataset sample before you buy
Buying data
- Data provenance explained for AI teams
Buying data