Scale AI Data Partnership: What It Buys and Pays

Yes. Scale AI's data partnership licenses an established company's records from the systems its teams already run on, structured and de-identified, from data warehouses, chat, email, documents, work tools and business platforms, and Scale anonymises them inside your own environment, before any data is extracted, releasing each package only with your sign-off. Its one published figure is an illustrative $10K to $1M+ per data partnership, scaling with cadence, a range whose top is 100x its bottom, and it says 40+ employees is its preferred starting point, though it assesses every partnership case by case. Most of what a search for Scale returns is news of Meta, which invested about $14.3 billion for a 49% stake in June 2025; the programme for companies' records is a separate, newer line. What Scale does not publish is most of what a company would sign: no licence text, term, exclusivity or right to pass the data on, and no payment timing, only that its partnerships are designed for recurring revenue.

This guide is for the founder, CEO, CFO or counsel of a company that has searched for Scale AI's data partnership or been told someone wants to introduce it, and wants to know what Scale licenses, what it might pay, whether the company qualifies, what happens to its records and how to apply or be introduced. Scale publishes far less than AfterQuery or Mercor: one page, a referral form, a short FAQ and an application that opens once you create an account. This guide reads all of them as they stood on 7 November 2026, with Scale's job posts and the press about it, and says plainly where Scale publishes nothing. It ends with a comparison against AfterQuery, Mercor, micro1, Handshake AI and Turing, and a worked example.

AIData.Broker is in Scale AI's referral programme, as it is in AfterQuery's, Mercor's, Handshake AI's, Turing's and micro1's. We introduce a holder to every buyer whose published rules take its data, free for the holder, because the buyer pays us its published referral payout out of its own money: for Scale, $10,000, paid only if Scale signs a partnership agreement, the amount its referral form names (its FAQ still says up to $50,000), and the holder sees the most any buyer may pay us before we introduce it. Where a sale run by us earns more, and only for holders based in the United States, the United Kingdom and Canada, we act as the holder's agent under a mandate and the holder pays our fee only if the sale closes; every fitting partner is introduced then as well. Either way the holder itself licenses its data to the buyer, and we do not buy it, take ownership of it, or license or sell it in our own name. Every term below is Scale's own published wording, a job post or named press, cited to its record or linked to Scale's page; where a line is our reading, it says so.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. What Scale AI is, and why it buys company data
2. What Scale licenses, system by system
3. Who it takes
4. What Scale pays
5. The application, step by step
6. What happens to your data
7. The licence and payment: what is published and what is not
8. Applying directly or being introduced
9. Scale AI beside AfterQuery, Mercor, micro1, Handshake AI and Turing
10. Worked example: a managed IT provider
11. Questions people ask
12. What this means for you

What Scale AI is, and why it buys company data

Scale AI is one of the longest-running suppliers of human data to AI labs. It provides training data, RL environments and model evaluations to AI labs, governments and enterprises, and says it has paid over $1 billion to its contributors over ten years, the people who do its human data work. It had $870 million in revenue in 2024, as Bloomberg reported, and says it ended 2025 with well over $1 billion in new business, a figure for bookings rather than revenue.

The news a search for Scale returns is about its ownership. In June 2025 Meta invested about $14.3 billion for a 49% stake, valuing Scale at over $29 billion, and Scale's chief executive, Alexandr Wang, stepped down "to join Meta and help the bigger company with its AI work", staying on Scale's board; Scale said it remains an independent company ( TechCrunch). Google, then Scale's largest customer, which had planned to pay it $200 million that year, reportedly planned to cut ties after Meta's investment. Who owns Scale matters to a company thinking of licensing it records, because Scale's page does not say which AI companies receive them (see What happens to your data).

Why Scale wants your records is in its FAQ: "AI companies want to improve their models for certain corporate personas such as marketers or financiers. They need real-world environments rather than synthetically generated ones to test and improve their models" ( data partnership page). Those are RL environments, simulated workplaces where a model attempts a task and is scored on the result; Scale says nearly half of its new data training projects involve them, and its job post describes them as "sandboxed replicas of the digital spheres where real knowledge work occurs and built from the operating data of the companies that actually hold it". Its Data Acquisition Lead is hired to find the companies that hold that data and get it, from cold outreach to signature, and to define, with Scale's legal team, the first version of how these transactions get priced. The price model is not settled, in Scale's own words, which makes a rival written offer the strongest thing a company can bring.

Scale has bought company data before, at least code: its SWE-Bench Pro benchmark keeps a commercial set of 18 proprietary repositories under formal partnership agreements with early-stage startups, companies "where we contacted the company and purchased their engineering repos". The data partnership for operating records is newer, and no seller has confirmed a deal through it. Scale is a principal: it licenses your records for its own account. Our reading: since it sells data, environments and evaluations to labs, your records become material for what it sells them, and its page does not say whether labs receive the records themselves or only what Scale builds from them.

What Scale licenses, system by system

Scale says it draws "from the everyday systems your teams already run on", taking "never the raw systems themselves, only structured, de-identified records." It lists six kinds of system and what it wants kept from each:

These are examples, not limits: its application lets you type in any tool it does not list ( its application). What the six have in common is the part of each record Scale asks to keep: who took part, who replied to whom, who wrote a document and when, and a ticket's history. That is what makes records worth more to every buyer (a ticket tied to its code commit is worth more than a standalone document). It is also the part that points at people, which is why the anonymisation section below matters more than Scale's short FAQ answer suggests.

The industries on its page

Scale lists its industries of interest in five groups ( data partnership page):

Read the list as demand, not a rule: nothing on the page turns away a company for being in an industry it does not name, and the application takes any industry you type. Some names are rare among buyers: Scale is one of three buyers of company records that name staffing and recruiting, and one of five that name field-service and dispatch work. Health records are outside the HIPAA Privacy Rule only once they meet its de-identification standard ( HIPAA de-identification for AI training), and a law firm decides what is privileged and keeps it out before anything is shared.

Who it takes

Size: 40 or more, case by case

Scale's rule is one sentence: "40+ employees is our preferred starting point ... though we assess every partnership on a case-by-case basis." It is a preference, not a bar, and it is the highest of the sizes the large buyers state for every company alike: Handshake AI looks for 20+ full time employees and 3+ years in operation, micro1 for 30+ full-time employees, AfterQuery for several teams working across several systems, and a few years of history, Mercor sets a size for each kind of company on its listings, and Turing names no size at all.

Scale's application asks for employees and contractors both today and at the company's peak, and whether their work was white collar or blue collar ( its application). Our reading: a company smaller now than it was can say so, and Scale's team decides which number counts (we count people at a company's largest, since they made the records). One buyer's chief executive has said value rises with the number of white-collar employees, and desk work leaves a written trail.

Beyond size, Scale looks for years of records in the tools your teams already run on; real operational workflows, "not demos or synthetic data"; domain depth, "Specialized expertise that makes your data hard to replicate"; and a clear owner who can scope the partnership, approve packages and keep reviews moving. The last is the practical one: every package waits for that person, so name them before the first call.

Operating, winding down or closed

Scale's page says nothing about companies that have stopped trading. Its application asks "Is the company alive?" and, if not, the year it ended ( its application). Our reading: a closed company can apply and Scale's team will consider it. Records lose value fast once subscriptions lapse (Turing's intake names 30 days for Slack and 60 to 90 days for a CRM after operations cease), so export first; and since Scale anonymises inside a company's own environment, a closed company holding only exports should ask how Scale handles an archive. Someone must still have the right to sign: in Delaware a dissolved corporation continues for three years to wind up its affairs and dispose of its property, after which a court-appointed trustee or receiver is needed. Read who owns a dissolved company's data and our page for closing companies and estates.

Where the company is based

Scale's pages name no country, language or tax rule; it is one of 14 buyers of company records whose pages say nothing about where a holder can be based. Its application asks for the company's country, from a list of every country, and whether the data is in English. Asking is not a rule either way, so a company outside the US should ask at first contact whether Scale takes companies from its country now, and read US withholding tax on data licence payments and selling data from outside the US.

What Scale pays

Scale's page gives one price, in its "By the numbers" panel: "$10K-$1M+ illustrative value per data partnership, scaling with cadence". Illustrative means it is not a price list. Per data partnership means per relationship, not per record or gigabyte. Scaling with cadence means the value grows with how often data is delivered, which fits the rest of the page: "We are looking for partnerships, not transactions", built "to keep paying back" ( data partnership page), with partnerships designed for recurring revenue.

A range that wide tells a company almost nothing about its own number, and Scale's page has no calculator. What its application asks is the best guide to what it prices on. The company overview says "Estimates are fine. This helps us size and price an offer", and asks for peak revenue or ARR, staff and contractors, the data's language, what is excluded or missing, and the company's key metrics and workflows ("This is what we use to understand the value of your data. Specific numbers help the most"); the data inventory asks how much is in each system and what is in it ( its application). Our reading: size, volume, the workflows the records capture and how often new records can be delivered are the inputs, and domain depth is the multiplier.

Other large buyers publish ranges for whole companies: micro1's tiers start at $100K-500K, Handshake AI's page says $100K to $4M+, Mercor's listings run from $10,000 to $1,500,000 a company, by kind, and AfterQuery's form offers bands from under $50k to more than $500k. These are ranges across many companies, not what yours is worth. A free valuation of what you hold, with what company data is worth and how AI data buyers calculate offers, gives you a number of your own before Scale gives you its.

Scope and cadence are the levers. Scale says "Partnerships can begin with a very limited scope", and asks a company to "Bring one workflow". Licensing one department first lets you see how Scale works before you commit more, and lets Scale see your records before it prices the rest. Because Scale's value scales with cadence, the recurring part is where the money is, and no buyer prices it in public: buyers of company records talk of refreshes and ongoing partnerships but none publishes a refresh price. Settle in the first agreement what each later package is paid, how often and for how long.

The application, step by step

Scale's page offers "Apply to partner", "Refer a team" and "Book a scoping call". The application at scale.com/data-partnership/apply, read on 7 November 2026, takes a general business through five steps:

A health or life sciences company also completes a data catalog, which asks whether the data is already de-identified (by HIPAA's Safe Harbor or Expert Determination method, or not yet), what approval it rests on, whether you can license it "for commercial use and AI training", whether rights are "siloed per originating client (each must approve)", and any existing exclusivity. A data vendor completes the catalog alone. It says "Nothing here is binding."

Three things deserve attention. There is no NDA before the overview, which asks for revenue, headcount and workflows, and at least one buyer's terms say that until an NDA is signed what you submit through its site carries no confidentiality obligation beyond its privacy policy. Give estimates and categories, not client names, and ask for a mutual NDA before the scoping call goes into detail. "Have you licensed this data before?" needs an exact answer: an earlier licence or exclusivity limits what you can grant, and some buyers ask you to warrant you have not already licensed the same data on similar terms. The Drive folder is optional: Scale itself says to keep it restricted and to put nothing confidential in it.

What happens to your data

Handling is the most distinctive thing about Scale's programme. Its FAQ: "Scale anonymizes everything inside your environment, before any data is extracted. PII is stripped and records are processed so they can't be traced back to individuals or accounts." Its governance panel says "You decide exactly what leaves your walls. Everything else never moves", and lists what never leaves: "Employee and customer identities", "Raw systems, credentials, or source access" and "Anything outside your signed agreement" ( data partnership page). Its safeguards include encryption in transit and at rest, per-package approval gates, a full audit trail and on-prem and VPC options. No other buyer of company records we read says it anonymises inside the holder's environment: AfterQuery removes or replaces identifiers itself as part of its transfer step, Mercor runs its own automated pipeline and Turing removes personal data on its side, air-gapped.

That is a better starting point than a raw transfer. It leaves four questions, and Scale's page answers none of them.

What runs inside your environment, and with what access. Anonymising inside your systems means Scale's software or people work on the identifiable records where they live; credentials and source access never leave, but something has to read them. Ask what Scale installs or connects, what access it needs and for how long, who on its side can see records before they are scrubbed, and under what agreement. Then ask whether the work can run in your own cloud account (the VPC option), so that only the approved package crosses to Scale.

What "can't be traced back" means when participants are kept. Scale wants chat with threads, participants and reply structure intact and documents with their authorship. Keeping who replied to whom while making each person untraceable means consistent stand-ins for names, and enough context (a title, a project, a client) can still point back to someone. For staff and customers in the EU and the UK, the European Data Protection Board's view is that pseudonymised data stays personal data ( licensing data for AI training under the GDPR); in California, data counts as de-identified only with all three of reasonable technical measures, a public commitment not to re-identify, and contracts binding every recipient. Scale publishes no method and no re-identification test, where AfterQuery says it tests every dataset. Ask who certifies the anonymisation, and to see the result before each package leaves.

Who receives it. Scale's FAQ says the data serves "AI companies" and answers "Are AI companies going to use my data to start competing?" with "No, these companies are interested in modeling the personas within your organization" ( data partnership page). It names no company. Meta holds a 49% stake and is itself an AI developer, and nothing on the page says whether Meta, or any lab, receives your records or only what Scale builds from them. Ask which companies may receive either, whether you can exclude any, and that the no-competing promise and a ban on re-identification and onward resale be written into the agreement. Where the records hold bulk US sensitive personal data, ask too that none reaches a country of concern under the Justice Department's data security rule, which applies whether or not the data is de-identified.

Whose decision it is. Scale's anonymisation protects Scale's use of the records; it does not decide whether you may license them. Customer records raise what your privacy policy promised ( is it legal to sell customer data), staff messages your employees' expectations ( employee emails and chats in AI training), and other companies' documents your contracts with them ( shared channels and third parties' attachments are reviewed or left out). The per-package approval is where you exercise that decision, so give it to someone who knows the records; our de-identification standard is a checklist for it.

The licence and payment: what is published and what is not

Scale describes the result, not the contract: "Your data becomes an asset you're paid for", and "De-identified, structured records from the systems you already run become a new asset class, monetized on your terms, never out of your control" ( data partnership page). No licence text, term, exclusivity, sublicensing right, promise about earlier licences or payment timing is published. "On your terms" is worth taking literally: Scale has published none of its own, so the terms are what you ask for. Creating an account means agreeing to Scale's general Terms of Use, written for its website and services, so ask that the partnership agreement alone govern your records and prevail where the two conflict.

What to ask for follows from what the market prices:

• Term. Disclosed AI licences run two to three years or at least three years, and licensing counsel recommends pricing a perpetual right to models already trained separately from a finite licence to the data.
• Exclusivity. On the market's evidence an exclusive licence fetches 2.0 to 5.8 times a non-exclusive one. If Scale asks for it, it is a separate price; if not, keep your freedom to license elsewhere in writing.
• Onward use. Which companies receive the records or what is built from them, the no-competing promise, and the re-identification and resale bans above.
• The recurring part. A guaranteed minimum each year, what each later package is paid, and when each payment is due. A recurring promise is worth only what the contract guarantees.
• The end. What is deleted, what survives in models and environments already built, and by when.

On payment, Scale publishes the least of the large programmes (the comparison table below sets out the others). Ask what triggers each payment (signature, your approval of a package, or Scale's acceptance of it), whether any part is paid before data moves, and whether Scale can reject a package. Since every package needs your sign-off before it leaves, that sign-off is the natural trigger. When AI data buyers actually pay, the licence terms page and the guide to data licensing agreements explain each term.

Applying directly or being introduced

There are two ways to reach Scale's data partnerships team. Both lead to the same review, the same scoping and an agreement you sign with Scale.

Applying directly

Open the application and work through the steps above, or use "Book a scoping call" on the data partnership page. Have four things ready: your systems, how far back each goes and roughly how much each holds; your headcount now and at your largest; a short note of what you will not share (named clients, health records, privileged files, a director's mailbox); and who approves packages and who signs. Use an address at the company's own domain. A company that applies itself saves the buyer whatever it sets aside for an introducer, and buyers budget a share of a deal, or a fixed amount for whoever brings them a company, so a holder that comes direct can ask for that budget in its price, though nothing Scale publishes says it agrees to that.

Being introduced

Scale's "Refer a team" form, on the same page, says "Refer a company and earn $10,000", with bonuses paid when Scale signs a partnership agreement; its FAQ still says up to $50,000, and adds: "Multiple referrals or intros to the same company will not be paid out. Only one referral is paid out per deal based on who successfully intros Scale to the right stakeholder" ( data partnership page). The form asks the introducer for their details and how they know the team, then the company, who Scale should talk to, their email and why they fit, and whether "Scale reaches out, with you cc'd" or the introducer makes the introduction. The company pays nothing for being introduced and signs its own agreement with Scale.

This is how an introduction through us works. Value your data first: if Scale's published rules take your company (it prefers 40 or more employees), the valuation says so and offers one button, Introduce me. Pressing it is your consent to the introductions it names, and we file each one in the partner's own referral form within one business day; for Scale, its Refer a team form, with Scale reaching out to you. Scale then writes to you itself, with us copied, to arrange the next step, and from there the scoping, the agreement, the packages and the payments are between you and Scale. The same press introduces you to every other partner whose rules take your data: several answers to compare, not one. For what each route costs, read going direct or through us.

Scale AI beside AfterQuery, Mercor, micro1, Handshake AI and Turing

A company weighing Scale usually weighs the other large suppliers of human data to labs that buy company records. Each has its own guide: AfterQuery, Mercor, micro1, Handshake AI and Turing's Project Lazarus.

AfterQuery takes companies far smaller than Scale prefers and scrubs records with consistent pseudonyms and a re-identification test on every dataset as part of its transfer, where Scale scrubs before anything leaves and publishes no test. Mercor publishes the most on price and is among the fastest to pay; it extracts first and anonymises on its side. micro1 takes the broadest licence of the group, perpetual and irrevocable, with the right to sublicense through multiple tiers; Scale's "on your terms" is an easier place to start for a company that wants control. Handshake AI binds itself in writing to training and evaluation only, the kind of use restriction Scale's FAQ promises but does not publish as a term. Turing shares Scale's interest in ongoing value but pays a share of every licence it sells of your record, nothing upfront.

Read side by side, Scale is the most protective of the large buyers on how records leave a company, the most demanding on size, and among the least specific on price, licence and payment. For a company of 40 or more that cares most about handling, it is worth an offer, in writing, beside at least one other buyer's. The guide comparing every company-data buyer programme covers a dozen more buyers, the buyer board shows each one's published rules, and the buyer match tool checks your company against them.

Worked example: a managed IT provider

Example. A managed IT services provider in Columbus, Ohio, has 70 people and 12 years of records: every client ticket with the technician's notes in its PSA, client documentation and runbooks in IT Glue, scripts in its remote monitoring tool, Microsoft 365 email and Teams, HubSpot and QuickBooks. About a fifth of its clients are medical practices.

Does it fit? Scale names managed IT services (MSP), the firm has more than the 40+ employees Scale prefers, and it has years of records from systems in daily use, real workflows and domain depth: diagnosing other companies' networks is hard to replicate. Its service manager, who already approves every change, is the natural owner for package approvals. It also meets Handshake AI's 20+ full time employees and micro1's 30+ full-time employees.

What might it earn? Scale will not say until it has reviewed the application. What its questions reward is what the firm has: tickets with a technician's notes and a resolution, which are how an agent learns IT support, and a stream of new tickets every month, the cadence Scale's range scales with. The owner gives estimates on the overview and gets an independent valuation first.

What must it settle first? Its tickets are about its clients' systems, and its client agreements decide what it may do with what it learned ( third parties' material is reviewed or left out). Its medical clients signed business associate agreements, and a vendor holding a health client's data may de-identify it for its own purposes only if that agreement permits, so those clients stay out. Its documentation tool holds client passwords and network details, and Scale's process would run inside the very environment where they sit, so the owner keeps that tool out of scope, removes secrets from tickets before Scale's process runs, and asks for the VPC option. Our page for managed IT providers and the guide to selling IT service desk data set out the rest.

What does it ask Scale? A mutual NDA before detail; what Scale runs inside its systems and with what access; who certifies the anonymisation of each package; which companies receive the records, with the no-competing promise in writing; the licence's term, exclusivity and end; a yearly minimum and the trigger for each payment; and that the partnership agreement prevails over the Terms of Use. It puts the same records before Handshake AI and micro1, whose terms answer some of those questions in writing already. It may sign with Scale. It will know what it signed.

Questions people ask

Does Scale AI buy company data?

Yes, as a licence rather than a purchase of ownership. Its data partnership takes structured, de-identified records from data warehouses, chat, email, documents, work tools and business platforms, which it says become an asset you are paid for, on your terms. It has bought startups' code before, for its SWE-Bench Pro benchmark.

How much does Scale AI pay for company data?

Its page shows an illustrative $10K to $1M+ per data partnership, scaling with cadence, meaning with how often data is delivered, and says partnerships are designed for recurring revenue. It has no calculator and is still defining how these deals are priced. Get an independent valuation before you answer its questions.

Does my company qualify for a Scale AI data partnership?

Scale says "40+ employees is our preferred starting point ... though we assess every partnership on a case-by-case basis", and looks for years of records in systems in daily use, real workflows, domain depth and a clear owner. Its industries run from software and law to freight and field services, and it turns no industry away.

Does Scale AI take closed companies?

Its page does not say. Its application asks whether the company is alive and, if not, the year it ended, so a closed company can apply; that is our reading of the form, not a published rule. Someone must have the right to sign for it.

How does Scale AI protect personal data?

It says it anonymises everything inside your environment, before any data is extracted, that employee and customer identities never leave, and that every package needs your sign-off. It publishes no method or re-identification test and keeps participants and authorship in what it takes, so ask who certifies the result.

Does Meta get my data?

Scale's page does not say which AI companies receive the records or what is built from them. Meta holds a 49% stake in Scale, which said it remains independent. Ask which companies may receive your records and put the answer in the agreement.

How do Scale AI referrals work?

Anyone can refer a company through the "Refer a team" form on Scale's page, which says "Refer a company and earn $10,000", paid when Scale signs; its FAQ still says up to $50,000. One referral is paid per deal. The company pays nothing and signs its own agreement with Scale.

Is Scale AI legitimate?

It is a long-established vendor to AI labs, with $870 million in revenue in 2024 and Meta as a 49% shareholder. Its data partnership is newer, and no seller has confirmed a deal through it. In October 2026 noyb filed five GDPR complaints against Scale over how its Outlier platform tracks and verifies its workers; they are complaints, not findings. Sign an NDA before you share detail, and read the agreement rather than the page.

What this means for you

Scale keeps more control on your side of the wall than any other large buyer: it anonymises inside your environment and lets nothing leave without your sign-off. It also prefers the largest companies and publishes the least about price, licence and payment. It rewards a company that arrives prepared.

1. Check the fit honestly. Count your people now and at your largest, list your systems and how far back each goes, and find your industry on Scale's list. The buyer match tool runs the same check across every buyer's published rules.
2. Get your own number first. Scale asks for revenue, data sizes and workflows before it names a price. Value your data and answer with estimates.
3. Name the owner and the scope. Who approves packages, which teams, systems and years go in, and what stays out. Use the readiness check and the pages for your systems under rights.
4. Ask the handling questions. What runs inside your systems and with what access, who certifies the anonymisation, which companies receive the records, and whether the VPC option fits; our de-identification standard is the checklist.
5. Write the terms Scale leaves open. Term, exclusivity, onward use, a yearly minimum, each payment's trigger and date, and what survives the end. Put the written offer through the offer check and decide where to stand with the negotiation page.
6. Hear from more than one buyer. Apply to Scale directly, or be introduced to it and every other buyer whose rules take your data with one press after your valuation, and compare price, licence, handling and payment together on the buyer board.