How AI Labs Buy Training Data: Formats, Paths and What Owners Actually Do

AI labs do not all want the same thing when they buy training data. Some want raw records, exactly as your business generated them, and they will structure the tasks themselves. Others want records already shaped into workflows, decisions and step-by-step reasoning traces. Understanding how AI labs buy training data, and which format fits your records, changes both the value of your asset and the work required to sell it.

In short:

• Published 2026 deals range from about $110,000 to $1.2 million for business operating records, including email, dispatch logs, workpapers, CAD files and SOPs.
• AI labs follow two main paths when acquiring training data: raw records they structure themselves, or records pre-shaped into tasks and workflows. The path that fits your data affects price and packaging.
• De-identification, contract review and a written offer come before any records leave your building. Nothing is licensed until you sign.

---

What "format" actually means when AI labs buy training data

When owners ask about training data formats, they usually mean: what does the file look like when it arrives at the lab? But format is really two questions. First, what is the file type (email threads, PDFs, structured database exports, chat logs)? Second, how much structure has been added to the content before delivery?

File type matters less than most owners expect. Labs can ingest plain text, PDFs, spreadsheets, JSON exports and most common formats. What matters more is whether the content shows skilled reasoning in a way the lab can use directly, or whether the lab will need to do significant work to extract that reasoning.

Raw records are unprocessed exports: your email archive, your shared drive, your dispatch system pulled into flat files. The lab receives them and builds its own annotation or structuring layer on top. This path requires less work from the seller. The lab absorbs the structuring cost itself.

Task-shaped records are records that have been organized to show a decision or a workflow from start to finish. A freight broker's file that shows the original load request, the rate quotes considered, the carrier selected, the exception that arose and how it was resolved is already a training task. An accounting firm's workpaper that shows the client question, the research steps, the reconciliation and the final position is already a reasoning trace. Labs that train agents on real-world tasks look for records in this form, because much of the structuring work is already done.

---

The two paths, and which one fits your records

Path one: raw data acquisition. Some labs, particularly those building large general-purpose models, want volume and variety. They buy raw operating records, clean them internally and build their own task structure. For this path, the seller's job is to produce a clean, complete export. The value comes from the breadth and authenticity of the records, not from any packaging the seller adds.

Records that fit this path well include large email archives, broad document repositories and database exports that cover many topics across many years. The lab wants the signal that comes from real business communication, not a curated sample.

Path two: task and workflow acquisition. Other labs, particularly those building agents that reason through multi-step problems, want records that already show a workflow. They are looking for the equivalent of a worked example: here is the problem, here are the steps taken, here is the outcome. For this path, records from firms that do skilled, process-heavy work fit best, and some light organization by the seller (or a licensing intermediary) can make them easier for a buyer to use.

Records that fit this path well include dispatch logs with exceptions and resolutions, engineering calculations with revision histories, accounting workpapers with client questions and final positions, and construction project files that show scope changes and decisions. A licensing firm reported a $1.2 million deal for an industrial distributor's communications, product knowledge, order history, supplier records and operating documents. A licensing partner reported $450,000 for an engineering firm's CAD drawings, calculations, technical reports, revisions and specifications. These records show skilled work step by step, which is exactly what agent-training labs need.

---

What owners actually do, in order

Knowing how labs think about format is useful, but most owners want to know what they personally have to do. Here is the sequence.

1. Inventory what you have. Before any conversation with a buyer or intermediary, know roughly what systems hold your records: email (and which platform), shared drives, project management tools, accounting software, dispatch or ERP systems. Note how many years of records exist and whether they are still accessible.

2. Get a size estimate. Buyers and intermediaries will ask about headcount and years in business before anything else. These two numbers give a rough sense of record volume. The estimate calculator at /#estimate asks only those two questions and returns an indicative range from a licensing partner's payout model. Treat it as a scenario, not a valuation.

3. Understand which path fits your records. If your firm did process-heavy, skilled work (logistics, accounting, engineering, construction, distribution), your records likely fit the task-and-workflow path. If your records are broad and varied but not organized around specific decisions, the raw-data path is more likely.

4. Allow a sample review. Before an offer is made, a licensing intermediary may ask, with your permission, to review a small sample to check quality and format. This is not a commitment. It is a quality check that lets the buyer confirm the records are what they appear to be.

5. Review the written offer. A written offer will specify what records are covered, the license terms (exclusive or non-exclusive, perpetual or time-limited), the price and the payment timeline. Only four of the fifteen published 2026 deals are described as perpetual, exclusive licenses. Terms vary. Read every term before signing.

6. Review de-identification. Identifying details about clients and staff are removed or replaced with consistent placeholders before delivery. No de-identification method removes every risk. You should review the process and a de-identified sample before signing, and you can exclude regulated, privileged or third-party records. Ask your lawyer about any records that may be subject to confidentiality obligations or professional privilege. This is a legal question, not one an intermediary can answer for you.

7. Approve the export and coordinate delivery. The full export happens only after you approve the terms in writing. Coordinate with your IT person. Once records are delivered and training begins, that training cannot be undone.

---

Questions to ask before you sign

---

Frequently asked questions

Do I need to convert my records into a special format before approaching a buyer?

Not necessarily. Most buyers and intermediaries can work with standard exports from common platforms: email archives, shared drive exports, PDF document sets, CSV or JSON exports from business software. If your records fit the task-and-workflow path, some light organization may increase value, but a licensing intermediary can advise on that after reviewing a sample.

Does it matter if my firm is closed or winding down?

No. Published deals include firms that were winding down at the time of the deal. The records need to exist and be accessible. Operating, winding down and closed firms all qualify if those conditions are met.

Can I exclude certain records from the license?

Yes. You approve the scope in writing before anything is exported. You can exclude regulated records, privileged communications, records covered by confidentiality agreements or any other category you choose. Talk to your lawyer about which records you are obligated to exclude.

How long does the process take?

Payment timelines vary by deal. Three published deals reported payment in five, seven and nine days from bid. Do not treat those as typical. Ask for the timeline in writing as part of the offer.

What if my records are a mix of formats and systems?

That is common. Most established businesses have records spread across email, shared drives, project tools and accounting software. Pewter Data sizes the records, brings a written offer from its licensing partner and coordinates the deal through payment, which starts with working out which systems are worth including.

---

The free guide at /before-you-sell covers rights, privacy, contract terms and an 18-question checklist that walks through the decisions above in detail. If you want a rough sense of what your records might be worth, the two-minute estimate calculator at /#estimate is a reasonable starting point.