Selling Data Direct to an AI Buyer vs Using a Broker

If an AI company has contacted you about buying your business data, the first question is usually "is the number good?" A better first question is "compared to what?" With one buyer, there is nothing to compare to. The buyer sets the price, the scope of use and the contract terms, and you decide only whether to say yes.

This guide explains what changes when several buyers compete for the same data, which terms tend to tilt toward the buyer in a one-on-one deal, how de-identification and packaging affect the outcome, and when going direct is a reasonable choice. It is written for companies, not for lawyers, and it is not legal advice. SellYourAIData is a broker, which means we have a stake in this question. We say so up front and try to show our sources for every claim about other companies.

The short version#

• A buyer with no competition decides the price.
• Public price tiers are anchors the buyer chose, not offers.
• Exclusivity is a separate product with a separate price.
• Contract terms (scope, exclusivity, indemnity, payment timing) can move more money than the headline number.
• Personal information handling is part of the deal, not an afterthought.
• Going direct is fine in some cases. Get an independent read first.

Single-bidder pricing versus a competitive process#

In a single-bidder deal, the buyer knows you have no alternative on the table. You may know the data is valuable, but you have no price that proves it. Buyers value data on their own theses. MIT Technology Review reports that AI labs value archives from failed companies on differing research theses, so prices vary widely.¹ That means two buyers can look at the same data and honestly come to very different numbers. One buyer alone gives you only one of those numbers.

What a real competitive process looks like#

The best-documented public example is the sale of Spirit Airlines' internal business data. Under a bidding procedures order dated June 22, 2026, Spirit ran a virtual auction in its bankruptcy case.² Google was named successful bidder on August 14, 2026 with a $10.0 million bid. Mercor bid $7.5 million and was named alternate bidder, meaning it steps in if the Google sale does not close. micro1 filed a notice that it intended to bid.² Time reported that micro1 later put in a higher figure after the auction closed, which shows how bidders can keep moving once they know others are in the room.³

The package was large. Reporting on the court filing describes about 100 million emails, 500 million Microsoft Teams messages and 516 code repositories.⁴ It is also a reminder that competition does not make a deal simple. The sale has not been approved. The court hearing was adjourned five times and is now set for November 6, 2026, with unions, software vendors and lawmakers raising objections.²⁵

Two cautions about reading this case. First, Spirit was a distressed seller with a court-run clock, which is not how a healthy company sells data. Second, the final price is one data point, and MIT Technology Review's reporting makes clear that valuations depend on who is looking.¹ The lesson is not "you will get $10 million." The lesson is that a process with several interested buyers produced a spread of numbers and a negotiating position, and a single-buyer approach would have produced one number.

How a private competitive process differs#

Most companies will never go through bankruptcy court, and you should not want to. A private process borrows the useful part. You prepare the same summary of the data, show it to several qualified buyers on the same timeline, and ask each for written terms. Buyers are then comparing their offers against what they believe others might do, and you are comparing real documents. This is the role a broker plays. SellYourAIData describes its process as taking a summary of your data, without your company name, to AI companies that may want it.⁶

A competitive process also costs time and attention, and it does not guarantee a higher number. If your data is narrow or ordinary, several buyers may simply agree it is worth little. Even then you learn that before you commit, which is worth something.

Published price tiers are buyer-set anchors#

Several buyers now publish what they say they pay. These pages are the closest thing to public pricing for ordinary business data, and they are worth reading. They are also marketing, and they are written by the party that benefits from you accepting them.

micro1's data page lists tiers: $100,000 and up for qualified enterprise data partnerships, $500,000 and up for large-scale operational datasets or participation across multiple teams, and $1 million and up for highly unique proprietary operational data. It says compensation depends on quality, uniqueness and value.⁷ The Next Web reported micro1's October 2026 announcement of a $1 billion, 12 month commitment to acquire company data.⁸ Handshake AI's data licensing page shows a payout range of $100K to $4M or more, and describes its estimate as indicative only and not an offer to purchase. It says final payout depends on data sources, volume, quality, due diligence and acceptance of licensing terms.⁹

Notice what the tiers do and do not tell you:

• They give a range spanning orders of magnitude, so they do not tell you where your data lands.
• They are conditional. The words "qualified" and "due diligence" mean the figure can change after you have shared samples and, often, stopped talking to other buyers.
• They are one buyer's view. A tier on one site says nothing about what a different buyer would pay for the same records.
• Mercor, for comparison, states no payout amounts. Its data page says pay depends on data volume, tools connected and data depth.¹⁰

An anchor is a number that shapes the conversation. If the first number you hear is a published tier, you will tend to negotiate around it. The way to loosen an anchor is to have other numbers.

Exclusive versus non-exclusive licensing#

Exclusivity decides who else can use your data, and it changes the price.

A non-exclusive license lets you license the same data to other buyers. An exclusive license says the buyer is the only one who may use it, usually for a stated field, term and territory. Once you grant exclusivity, you cannot sell the same records to anyone else within that scope, and you may be limited in using the data yourself for AI work.

A Malaysian law firm note on AI data licensing advises settling exclusivity at the outset and, for licensors, keeping the right to license the data to third parties and to use it to train their own AI.¹¹ The same note says that removing data from a trained model may not be feasible, so licensors may have to leave datasets in the model.¹¹ That second point matters: a license can be hard to unwind once training has happened.

On price, the public evidence is thin. Epoch AI reports, citing two founders it interviewed, that exclusive deals for reinforcement learning environments cost roughly four to five times as much as non-exclusive ones.¹² That is one segment, two sources and a different kind of product from a company's email archive, so treat it as a sign that buyers pay for exclusivity, not as a multiplier you can apply to your data.

Licensing the same data to several buyers#

If your data has value to more than one kind of buyer, non-exclusive licensing can let you be paid more than once. A support-ticket archive might be useful to one buyer for training customer-service models and to another for evaluation. That only works if:

1. Each license is clearly limited to a purpose (field of use).
2. No license grants exclusivity or a "most favored" clause that stops you selling elsewhere.
3. The data is packaged so each buyer gets only what its license covers.
4. You have the underlying rights to license to several parties in the first place. Customer agreements, vendor terms and employee policies can restrict this.

A single direct buyer has little reason to propose this structure, because it is cheaper for them if you stay exclusive and cheap. This is one of the clearest differences between direct and brokered sales.

Contract terms buyers write in their favor#

Headline price gets the attention. These terms often move more value.

Scope of use#

Is the buyer licensing the data to train one model, any model, or to resell? Is the license perpetual, or does it run for a term? A sample agreement memo from terms.law recommends a perpetual license for training and use of the trained models, paired with a finite license for ongoing dataset access, with extension to other models for an additional fee.¹³ That is a practitioner view, not a market standard, but it shows the points on which fees can be negotiated: which models, how long, training versus access. A buyer's first draft usually asks for the broadest version of each.

Exclusivity#

See above. Check also for clauses that act like exclusivity without using the word, such as a promise not to license "similar data" to others or a right of first refusal on future data.

Indemnity and warranties#

Law firm Sidley says provider representations that you have the right to license the data, and indemnity for third-party claims, are increasingly standard in well-negotiated AI development agreements.¹⁴ In plain terms, you promise you had the right to sell, and if someone says you did not, you pay. The Spirit case shows why that matters. Software vendor Springshot argued that data generated by its software inside Spirit's systems was not Spirit's to sell.⁴ A seller who signs a broad warranty without checking vendor and customer contracts takes on that risk personally.

Payment timing#

When and how you are paid is as important as how much. Mercor's data page says it typically wires payment within 2 to 4 weeks of the signed agreement.¹⁰ Handshake says payment depends on due diligence and acceptance of terms.⁹ Neither is a warning sign by itself. What matters is that timing, any holdbacks and what happens if due diligence shrinks the scope are written into the contract.

Public reporting shows what happens when pay terms go wrong, though not in a data-licensing setting. Business Insider reported in March 2026 that five contractors on OpenAI projects said Handshake AI suspended their accounts without warning, and four said they were not paid for their work. Handshake support reportedly told one contractor the decision was final with no appeal process, and a Handshake representative declined to comment.¹⁵ That story is about individual contractors, not companies licensing data, and the claims are allegations. It is still a useful reminder to get payment terms, dispute rights and termination consequences in writing before you hand anything over.

Security and who holds your data#

A buyer's security record matters because your data may sit with them. TechCrunch reported that on March 31, 2026 Mercor admitted a data breach it attributed to a compromise of the open-source LiteLLM tool, and that a hacker group claimed 4 terabytes of stolen data. Mercor had not commented on the claim's authenticity at the time.¹⁶ In a June 25, 2026 update, Mercor said its investigation was complete and that a very limited subset of its nearly five million experts had sensitive information affected, with no evidence of fraudulent use.¹⁷ That incident concerned its expert platform, not its data licensing program, but it is exactly the kind of fact to weigh before you connect business systems. Ask for security documentation, audit rights and deletion duties in the contract.

De-identification and packaging#

Business data nearly always contains personal information: customers, employees, vendors. Someone has to remove it or restrict it, and the contract should say who, how and to what standard.

The Spirit case shows the steps. The parties excluded transaction, refund, in-flight purchase and WiFi systems, and the privacy ombudsman's October 5 report recommended approval after that carve-out. Tonic.ai is to deidentify the remaining corporate email before each transfer.²¹⁸ A flight attendants' union still objected, arguing the agreement requires links across datasets to be preserved and that information about individuals could be reconstructed, which Spirit disputes.⁵ The ombudsman's mandate covered consumers and not employees, so the worker issue stayed open.² The Next Web separately reported on who picked and paid the firm doing the anonymizing, which shows how much control over this step can sit with the buyer.¹⁹

What this means for you:

• Decide scope before the buyer does. Which systems are in and out is your decision, and it is easier to narrow the package before an offer than after.
• Agree who deidentifies. If the buyer does it, you give up visibility into the standard. If you do it, you carry the cost and the risk if it is imperfect.
• Treat anonymization as a claim to test. The UK Information Commissioner's Office publishes guidance on how to check that anonymisation is effective, which is the kind of standard a contract can point to.²⁰
• Package by use. A buyer that wants customer-service transcripts does not need your payroll files. Smaller, cleaner packages are easier to de-identify and easier to license more than once.

Anonymity until terms#

Another difference between direct and brokered sales is what the buyer learns about you before there is a deal. In a direct approach the buyer usually knows who you are and what you have from the first conversation, which is also how the buyer learns what you might accept. A broker can show buyers an anonymized summary of the data first, and reveal your identity only once terms are agreed. SellYourAIData describes its own process this way.⁶ You should confirm it in writing with any broker, including us, along with who sees raw data and at what point.

When going direct can make sense#

Going direct is not wrong. It is sensible in several situations:

1. The dataset is small or simple. If the work of running a process would cost more than the likely difference in price, a clean direct deal is reasonable.
2. A buyer has already approached you with a strong written offer. If the price, scope and terms are in a document and look fair, that is real information. You can still check it.
3. You have a long relationship with the buyer and trust is part of the value, for example a vendor that already uses your data under contract.
4. Speed matters more than price. A shutdown or a deadline can make a quick sale the right sale.
5. You already have several offers. Then you have your own competitive process.

Even in these cases, get a second opinion. An independent review of a single offer is cheap compared with granting a broad, exclusive or perpetual license. If a buyer pushes you to sign quickly, or asks you not to talk to anyone else, treat that as a reason to slow down.

If you are already talking to a buyer, we review offers free.

An illustration: one offer versus three#

Imagine a company with a support-ticket and CRM archive. In Scenario A, one buyer makes an offer. In Scenario B, three buyers make offers after seeing the same anonymized summary.

Three things stand out, even with made-up numbers.

First, the lowest offer in Scenario B is lower than the single offer in Scenario A, so more offers do not mean every number is higher. The benefit is that you can see the range.

Second, the headline price is not the whole comparison. Offer 2 pays less than Offer 3 but is non-exclusive, so the seller could license the data to others during the term. Whether that is worth more depends on whether other buyers exist, which is exactly what a competitive process tests.

Third, the single offer in Scenario A has the worst terms on exclusivity, payment and indemnity, and you would not have noticed that without something to set it against. In Scenario B you can ask the buyer in Offer 3, with the highest price, to move on the uncapped indemnity, because you can point to two offers that cap it.

None of this means the competitive route always wins. It means you have information, and information is where negotiating power comes from.

How to compare offers#

Whether you have one offer or five, line them up on the same questions:

1. What exactly is licensed, and for which uses?
2. Is it exclusive, and for how long and in what field?
3. What is the total price, how much is paid up front, and when is the rest due?
4. Who removes personal information, to what standard, and who pays?
5. What do you warrant, what indemnity do you give, and is it capped?
6. What happens to the data and any derived datasets or models when the license ends?
7. Who are the downstream recipients, and can you object?
8. What security evidence will the buyer show you?

If a buyer will not answer these in writing, that tells you something.

What to do next#

• Take stock of what you have and what it is. Our assessment is a short way to start.
• Read through the deals database to see how past transactions were structured and priced, and what went wrong.
• Read our guide to AI data buyers to see what each buyer publishes about its terms.
• If you are already in conversation with a buyer, send us the offer. We review offers free.

This page is general information and not legal advice. Talk to a lawyer before you sign any license that touches personal data, customer contracts or employee records.