Guide
How to prepare your records for AI data licensing
Short answer: Prepare by building an inventory, not by exporting data: list each system, the years it covers, the file types and approximate counts, who owns the records, and what must be excluded (client-confidential, personal, HR, controlled or defence work). Then pick one to three representative sample projects to share under NDA. Buyers say this level of detail is what gets a real offer.
What to inventory, what to exclude, and what buyers want to see before they price, written for firm owners, not engineers.
Last updated · Sources linked inline · Buyer figures as published, observed Oct 16, 2026 (micro1, Mercor and GDC re-checked Nov 3, 2026)
What goes in a records inventory?
One line per system: what it holds, years covered, users, approximate volume, and whether you can export it.
- Every system that holds work records: email, chat (Teams, Slack), file shares, document management, ticketing, project management, CAD/PDM vaults, source control, accounting and CRM
- Years covered by each, and how long you've used it (buyers value long, consistent use of the same tools)
- How many staff use it day to day (desk staff)
- Approximate volume: file counts by type, mailbox sizes, ticket counts
- Whether you're still subscribed, and whether exports are possible
- Main language of the records
How should I count files and record types?
By type and by project. A rough count per format beats a single total.
For engineering, architecture and manufacturing archives, count native files by format (.dwg, .rvt, .ifc, .step, .sldprt and so on), PDF drawing sets, and the project paperwork around them (RFIs, submittals, change orders, inspection and QA records). Polyshares' co-founder describes a 15-year industrial architecture firm that submitted 300 projects and 220,000 files, broken down as 25,000 .dwg, 8,750 .step, 2,000 .sldprt and 20,000 PDF drawings, plus 3 full sample projects (post; self-reported). That's the level of detail to aim for. For software companies, count repositories, years of commit history, linked tickets and design docs. For professional firms, count matters or engagements, procedures and template libraries.
What makes a good sample set?
Complete, representative and de-identified to the buyer's standard, shared only under NDA.
- Complete: a whole project from brief to close-out (models, drawings, markups and correspondence), not a folder of finished PDFs.
- Representative: typical work, not your single best job.
- Clean of the excluded: no client-confidential terms, personal data, credentials or controlled technical data.
- Under NDA: ask the buyer for its de-identification requirements before you share.
What must be excluded before anything is shared?
Client-confidential material, personal and HR data, credentials, privileged content, and defence or export-controlled work.
- Client files protected by a professional code or contract
- Current designs and pricing (GDC says these stay with the company, as published)
- Defence, controlled-goods or ITAR/EAR technical data (why)
- Tax returns, tax IDs and tax-authority correspondence
- Bank and card details
- HR, payroll and medical files
- Passwords, keys and security configs (source code often contains secrets: scan for them)
- Privileged legal advice
Who owns the records you want to license?
Check before you count. Client contracts often give the client ownership of deliverables.
Your engagement letters, MSAs and terms of service decide who owns drawings, models, reports and code. Work you authored for internal use (procedures, templates, standards, internal correspondence) is usually firm-owned. Deliverables may not be. If ownership is mixed, the scope should say so. The intake asks about ownership for this reason. See contracts and tool terms.
What gets a submission ignored?
Big totals with no breakdown, no samples, and records already shopped around.
The same co-founder says a "Mixed CAD" submission of 300,000 files with no samples "wasn't even considered" (post; self-reported). A broker commentator says "some labs won't even entertain a dataset if it's already being shopped around" (post; opinion). Lead with a breakdown and samples, and go to one buyer at a time. More on that in exclusivity and licence terms.
Sources: Polyshares co-founder on intake (X) · termsheetinator on exclusivity (X) · The General Data Company: data partnerships · micro1: data partnerships
Part of our guide series. Start with the pillar: how to license your firm's records to AI.
What do firms ask about preparing records?
Do I need to export my data before talking to a buyer?
What is a sample set?
Should I clean up or delete things first?
How long does preparation take?
Ready? Tell us what you hold
Three answers give you a rough range based on what buyers publish. The rest helps a buyer take you seriously: every question after step 1 is optional. No files, logins or client names.