Short answer. The company records AI labs pay most for are the ones that let a stranger follow a piece of work from the first request to the final result: the email thread, the chat, the ticket, the file revisions and the outcome, in one traceable chain. Industry matters second. It decides how many buyers are asking and how much of the record you are allowed to license.
This guide covers the type of record that carries value, what a full chain looks like in four kinds of business, which industries buyers name most, which are in demand but hard to license, which records look valuable but are not, and which cannot be licensed at all. It is general information, not legal advice.
What kind of company data do AI labs want?
Operational records. The working files a company produces while it does its job, as opposed to the finished product it hands a customer.
The reason is how labs now train. They build simulated workplaces and let an AI agent practice inside them. TechCrunch describes these reinforcement learning environments as "training grounds that simulate what an AI agent would be doing in a real software application." HFS Research calls them "digital replicas of business operations containing workflows, documents, policies, systems and exceptions." A Surge AI executive told TIME that these environments "are only as good as the data that populates them," and TIME notes that the data needed is "largely private, buried in corporate emails and chats."
A simulated workplace needs three things from a real one:
- The request. What someone asked for, and the constraints they gave.
- The reasoning. The back and forth: the question, the objection, the revised number, the approval.
- The result. What was delivered and whether it held up.
Buyers say the same thing in their own words. AfterQuery wants "Who asked for what, what was decided and how it turned out." Pewter says "Records that show skilled work step by step are in highest demand." Polyshares says "What we are paying for is captured judgment."
The clearest illustration comes from a buyer. Ryan Lock of Polyshares said on The NGMI Podcast that his own former solar company opened a ticket, resolved it on a phone call and closed the ticket. "None of it had any value," he said. His co-founder's company ran remotely in Slack, so everything the team did was recorded in context.
So the first test for any company is simple. Pick one job from three years ago. Can someone who was not there follow it from start to finish using only what you kept?
What does a full chain of work look like?
It differs by business. The examples below are illustrative. The record types in the last column are the ones buyers name on their public pages.
| Business | One chain, start to finish | What buyers name |
|---|---|---|
| Product manufacturer | Customer complaint, engineering change request, drawing revision, supplier quote, released change notice, updated work instruction | "CAD and design histories, quality processes, supplier coordination" (AfterQuery). "Production records, quality reports, work instructions" (Handshake AI) |
| Commercial contractor | Request for information, revised estimate, change order, owner approval, schedule update, closeout | "RFIs, submittals, schedules and drawings" (Collinear) |
| Freight broker | Rate quote, shipment paperwork, dispatch, a late pickup, the claim, the invoice adjustment | "Rate quotes, shipment documents, dispatch records, delivery exceptions and claims" (Pewter) |
| Software company | Support ticket, team chat, bug report, code change and review, release, reply to the customer | "Private repositories with pull requests and linked tickets" (Collinear). "Technical documentation, issue histories, development workflows" (Handshake AI) |
Each chain has a decision made under real constraints, and the decision is visible. That is what a lab can turn into a practice task with a known right answer. MIT Technology Review reports that buyers value a set by "research theses" about how many useful tasks it can yield, which is why the same records draw different prices from different buyers.
Two practical points follow.
The chain usually crosses systems. The request is in email, the argument is in chat, the approval is in a ticket and the result is in a folder. Collinear describes its best fit as "the same work visible across several tools". A company with one system has one link of the chain.
Revisions are worth more than final versions. In Vaquero's conversations with buyers in November 2026, the most specific request was for native CAD files in preference to PDF exports. A PDF shows the final answer. The native file and its revisions show the work.
Which industries are buyers asking for right now?
Vaquero reviewed the public pages of 48 buyers on October 30, 2026. Thirty-five publish a specific industry list. This is how many of the 35 name each industry. One buyer can name several, and the grouping is ours.
| Industry | Buyers naming it (of 35) |
|---|---|
| Software and technology | 25 |
| Finance and insurance | 24 |
| Healthcare | 21 |
| Manufacturing | 19 |
| Logistics, freight and distribution | 18 |
| Legal | 18 |
| Retail and e-commerce | 15 |
| Professional services and consulting | 13 |
| Accounting | 12 |
| Construction | 10 |
| Engineering design and CAD | 10 |
| Trades and field services | 8 |
| Real estate and property | 7 |
| Hospitality | 5 |
| Marketing agencies | 4 |
One buyer ranks demand outright. Labelbox Prism labels Software and Internet "URGENT," Financial Services, Healthcare and Life Sciences, and "Manufacturing, Logistics, and Telecom" each "HIGH DEMAND," and "Retail, Consumer, and Hospitality" "STEADY."
A list tells you who is asking. It does not tell you who can sell. The industries where both line up, because the records are about the company's own work:
- Manufacturing and engineering. Drawings, change requests, quality records and supplier threads belong to the company that made the product. Buyer-published examples include a 70-person engineering firm and a 760-employee industrial manufacturer.
- Software. Code, tickets and design discussion are the company's own. It is the most named industry, and also the one labs already hold the most of. Public code is in the models already, so buyers want private repositories with the history around them. DataVendor says such material sells best when it "cannot be found in public".
- Freight, logistics and distribution. Quotes, exceptions and claims are judgment calls, written down daily.
- Construction and the trades. Estimates, change orders and field notes. Of the four deals Polyshares publishes, three are a roofing company, a general contractor and a solar company.
Demand moves with lab projects. A list read in October is a guide for this quarter and should be checked again next quarter.
Which industries are in demand but hard to license?
Finance, healthcare, legal and accounting sit near the top of the table. Buyers want them because the work is expert judgment. Mode sets a lower size floor for them: 6 or more employees for legal and 10 or more for accounting, finance and medical, against 30 or more for everyone else. Labelbox Prism marks "Financial reasoning traces" "URGENT." DataVendor says its strongest demand is for enterprise codebases and for cybersecurity, legal and finance data.
The problem is ownership. Most of what these firms hold is about someone else.
| Industry | Why buyers want it | What limits it | What usually remains |
|---|---|---|---|
| Law firms | Research, drafting and negotiation are high-judgment work | Client files and privileged communications. AfterQuery: "Privileged material is never used." | Internal procedures, templates, training, firm operations. Handshake AI names "Research workflows, process documentation, reusable templates" |
| Accounting firms | Workpapers and reconciliations show step-by-step reasoning | Client confidentiality. The Journal of Accountancy: "Redaction does not, by itself, turn confidential client information into nonconfidential information." Tax return information carries criminal exposure under IRC Sec. 7216 | Firm methods, review checklists, internal operations |
| Healthcare | Clinical and administrative decisions | Patient records. Labelbox Prism requires "IRB approval or qualified de-identification" for medical records | Scheduling, billing operations, staffing, training. Handshake AI names "Operating procedures, training materials, administrative workflows" |
| Banks, lenders, advisors, insurance agencies | Underwriting and valuation reasoning | Customer financial data under federal privacy rules and client agreements | Models, credit policy discussion and process documentation with customer data removed |
| Agencies, consultancies, staffing firms | Project work is judgment-heavy | The deliverables and the client correspondence belong to the client. Polyshares: "Data you hold for your own clients is out of scope." | Playbooks, internal planning, how the firm runs. One of Polyshares' four published deals is a global marketing agency |
| Contract manufacturers and job shops | Same engineering value as product manufacturers | Customer prints and specifications may belong to the customer | Process sheets, quoting, quality records, the shop's own tooling designs |
Two rules sit behind the table. Frankfurt Kurnit's Daniel Goldberg puts the first one plainly: "A company that holds data does not necessarily have the right to sell it." The second is that removing names does not fix it. In his words, de-identification "addresses whether a record can be linked to an individual, not whether its content is confidential."
These firms are not a hard no. They are a narrower yes, and the contracts have to be read before any buyer prices the scope.
Which records look valuable but are not?
Owners often lead with the wrong asset. The common cases:
- Customer lists and contact databases. Buyers strip identities out. Scale AI lists "Employee and customer identities" among the things it never takes. Labelbox Prism describes its program as "not selling customer data."
- Finished outputs without the work behind them. A final report, a signed contract, a set of financial statements. The answer is there and the reasoning is not.
- Exports with no narrative. A CRM or accounting export on its own, or tickets with an open and close date and no notes. Polyshares' pages say support tickets, internal threads and quote revisions "tend to carry more of the reasoning" than CRM and accounting exports do.
- Loose documents. Collinear lists "Loose documents with no company around them" as not a fit. A folder of files without the email and chat that produced them is a pile, not a chain.
- Public, scraped or AI-generated material. Labs have the public web already. Collinear excludes "AI-generated code, or a public framework with a thin layer on top." Scale AI asks for real operational workflows, "not demos or synthetic data."
- A large headcount that works in person. Restaurants, retail floors and single-site hospitality can employ hundreds and write very little down. Buyers price the written record. Labelbox Prism puts retail, consumer and hospitality in its lowest demand tier.
- Records mostly outside English. Collinear lists "Data that is mostly not in English" as not a fit.
A note on industries no buyer lists. In our review, no buyer named franchising, event production or membership clubs. That means nobody has published a price for them. It does not mean they are worth nothing. A franchisor's head office, with years of operating manuals, training and support threads across locations, may pass the chain test better than a larger company that kept less. Ask before assuming either way.
Which records cannot be licensed at all?
Some records stay out no matter how well they show the work. Our owner's guide covers the law behind each. The short list:
- Anything you hold for a client or that a contract or NDA restricts
- Patient and health records
- Payment card data
- Attorney-client privileged communications
- HR and personnel files, payroll and tax records, health and benefits information
- Tax return information held by a preparer
- Export-controlled drawings and technical data. This one matters for manufacturers, because the same CAD files buyers want most can fall under it.
- Recorded calls without a documented consent record
- Records collected under a privacy policy that promised not to sell them
Frankfurt Kurnit recommends excluding these categories outright instead of relying on de-identification to cure them.
How do you tell where your company lands?
Three questions, in order.
- Can a stranger follow a job from start to finish in what you kept? If the work happened in email, chat and files over several years, yes. If it happened on the phone and in person, mostly no.
- Is the work yours or a client's? Your own products, projects and operations are clean. Work done for clients narrows the scope to your internal side.
- Is anyone asking for your industry this quarter? The table above is the starting point. A broker who talks to buyers every week knows what changed since.
| Where you land | Typical examples | What to expect |
|---|---|---|
| Strong chain, own work, in demand | Manufacturers, engineering firms, software companies, freight and distribution, contractors | Several buyers will look. Compare offers. |
| Strong chain, client-owned work | Law, accounting, healthcare, financial firms, agencies | A narrower license on internal records, after a contract review. |
| Weak chain | In-person businesses, ticket-only systems, short history | Little to license today. Worth fixing how work gets recorded. |
| Restricted | Regulated, privileged or export-controlled records | Stays out. |
Find out where your records land
The Vaquero estimator takes office headcount and years of records and returns a range built from published deals. A person replies within one business day with a view on which buyers are asking for records like yours this quarter. No files, no names, nothing to sign. The buyer pays our fee at closing. You pay nothing at any step.
Rivers Colyer is a co-founder of Vaquero Data Co. in Austin, Texas. Vaquero represents companies licensing their operating records to AI labs through licensing houses. This guide is general information, not legal or tax advice.
Sources
Read November 19, 2026 unless another date is given.
- TechCrunch, "Silicon Valley bets big on 'environments' to train AI agents," Sept 21, 2025. Definition of reinforcement learning environments.
- TIME, "Why Google Bid $10 Million for a Failed Airline's Data," Aug 25, 2026. Environments only as good as their data; private workplace records.
- MIT Technology Review, "AI labs are buying data from dead startups to train their models," Oct 13, 2026. Valuation by research theses on tasks per dataset.
- HFS Research, "The experience data embedded in your business is AI's next gold rush," Oct 10, 2026. Digital replicas of business operations.
- The NGMI Podcast, Oct 16, 2026. Ryan Lock on his former solar company's tickets. Heard October 17, 2026.
- AfterQuery, data partnerships. Who asked for what; CAD and design histories; privileged material.
- Pewter Data. Skilled work step by step; record types by industry.
- Collinear, data acquisition. Construction and codebase records; same work across several tools; not-a-fit list.
- Handshake AI, Data Partnerships. Record types by industry.
- Labelbox Prism and Prism categories. Demand labels by industry and by data category; medical record conditions.
- Scale AI, data partnership. What it never takes; real workflows, not synthetic data.
- Mode. Size floors by industry.
- DataVendor documentation, "What you can sell". Strongest demand categories; non-public material.
- Polyshares, "How much is company data worth", "Sell business data to AI labs", intake and track record. Captured judgment; which records carry reasoning; client-held data; four published deals. Read November 12, 2026.
- Frankfurt Kurnit Klein + Selz, Daniel M. Goldberg, "Selling Company Data for AI Training: What to Know Before You Say Yes," Oct 1, 2026. Right to sell; limits of de-identification; categories to exclude.
- Journal of Accountancy, Kelly D. Mullins, "Asked to sell firm data for AI training? Be very cautious," Nov 2, 2026. Redaction and client confidentiality; IRC Sec. 7216.
- Vaquero buyer review, October 30, 2026. Industry counts from the public pages of 48 buyers, 35 of which publish a specific industry list.
