How does an AI lab's data team judge a data offer?
The short answer
OpenAI's data acquisition job post (5 Jun 2026) says the role balances research value, feasibility, quality and responsible execution. Its partnership page wants large datasets that aren't already easy to find online. Its intake form asks whether you hold the rights.
I found no lab that publishes its scoring sheet. Labs' job posts come close, and they say more than you'd expect.
Everything below comes from pages the labs and data companies wrote themselves. It describes what they say they look for. It can't tell you whether a given lab will say yes.
Who reads your offer at a lab?
OpenAI's data acquisition job post (5 Jun 2026) describes its research program manager as a key interface with outside data offers. Anthropic's Research Counsel job post (3 Sep 2026) adds a lawyer who advises on data sourcing, licensing, partnerships and use.
Across the two labs' posts, three roles sit near a data offer:
- Researchers. OpenAI's post says the role works with researchers to scope data needs and define success criteria.
- Engineers. The same OpenAI post says the role works with engineers on data platforms and security practices such as access management.
- A lawyer. That Anthropic counsel advises the research and data teams, so the legal read sits beside the technical one.
So one offer has to answer a research question, a delivery question and a legal question.
What do the labs say they look for?
OpenAI's data acquisition job post lists four things it balances in data opportunities, and OpenAI's Data Partnerships page (9 Nov 2023) adds what it wants and what it isn't seeking.
What OpenAI says it looks for in a data offer, from its own pages.
| What they look for | What the lab wrote | Where |
|---|---|---|
| Four factors | Research value, feasibility, quality and responsible execution | OpenAI job post |
| Not easy to find online | Large datasets not easily accessible to the public | OpenAI partnerships page |
| Human intent | Long-form writing or conversations, not disconnected snippets | OpenAI partnerships page |
| Format | Almost any form, with OCR and transcription help | OpenAI partnerships page |
| Not sought | Sensitive or personal information, or data owned by a third party | OpenAI partnerships page |
Put plainly, the OpenAI pages ask whether your records hold something a model can't get from the public web, whether whole conversations survive, and whether anyone else's rights or personal details sit inside.
The partnership page says OpenAI can help remove personal or third-party information. Even so, the "Not sought" row above shows which of your records will raise questions first.
What does a lab's own intake form ask?
OpenAI's data partnerships interest form, the one lab form I found (live on 10 Oct 2026), asks what the data is, how much there is, whether you hold the rights to share it and what help you need to deliver it.
The form says submitting commits you to nothing and the answers stay private to OpenAI. Past your name, email, company and country, it asks eight required questions:
- your organization type, with for-profit private companies one of the options
- the data type: written text, spoken language, images, video, numeric and others
- an estimated size, in units you choose, such as words or hours
- whether the data is digital or in print
- a short description of the data
- whether you have the rights to share it
- what help you need to clean and deliver it
- whether you would share it publicly, privately or either way
It ends with an optional "Anything else?" box.
Three of those eight questions are about volume and form. One is about rights, a free-text box worth drafting with your lawyer.
What do data companies ask first?
Data companies with company-data programs publish their own first questions. These are each company's own descriptions:
- Handshake AI asks about your company and the tools you use. Its estimator asks for headcount, years in operation and country.
- Handshake AI says payouts turn on the data sources, volume and quality and its due diligence. A scoping call covers data parameters, de-identification, terms and timeline.
- Troveo's estimator asks for headcount and years of operating, and its page says value grows with the systems you license. It wants complete chat threads with participants and replies intact, and documents with authorship, versions and timestamps.
- micro1's Enterprise Data Partnership program looks for companies with 30+ employees, mature operations and English documentation, with US companies prioritized.
- That program reviews a submission for fit, then holds a discovery call on data and privacy requirements.
Handshake AI also says partners remain responsible for making sure their transfer meets legal requirements. That puts the rights work on you before anyone opens a file.
What should your first message say?
Write it for the roles above. Descriptions and counts are enough for a first read:
- What the records are. The systems, such as email, chat, tickets or a CRM, and the kind of work they show.
- How much and how far back. Rough counts and the years covered.
- Whether any of it is public. OpenAI says it wants data that isn't already easy to find online.
- Who is in it. Employees, customers or partners, and what you would leave out.
- Where the rights stand. Contracts, notices and who can sign, with your lawyer's questions listed.
- How it would export. File formats, and whether threads and metadata come out whole.
Keep the records themselves back until counsel has read the terms. Delivered data gets checked later, after the contract is signed, as data verification before payment shows, so the first message only needs to earn the scoping call.
For what labs do with the chain of work inside those records, see what data AI labs want from companies and why AI labs want real company data. The data readiness checklist turns the list above into a working sheet.
What do labs leave out?
What labs pay, how many offers they turn down and how long a review takes. Neither OpenAI's nor Anthropic's pages above say.
OpenAI's form is general, written for any organization with data. I found no lab form built for company archives such as email or chat, so the visible routes for those run through data companies. Licensing direct or through a data company weighs the two routes.
The free rights check walks through seven rights questions to take to your lawyer. Optional, paid managed readiness builds the systems inventory, questions for your counsel and buyer shortlist from descriptions and counts, quoted in writing first.
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
