AI data licensing for operating companies
Clean data commands a higher price.
AI data brokers are approaching operating companies with six- and seven-figure offers to license their operational data for model training—employee e-mail, Slack, support tickets, sales records, and imagery. The buyer prices that data on provenance. A dataset with a documented rights story is priced on what it contains. A dataset with an undocumented one is priced on the buyer’s worst case, and the discount comes out of your number.
We inventory what you hold, clear what can be licensed, exclude what cannot, and negotiate the agreement. The cleared dataset is smaller than the broker’s intake list. It also sells for more.
Why provenance sets the price
Every AI buyer now carries litigation risk on the data it trains on. That risk is priced. When a broker’s counsel cannot tell where a dataset came from, whose confidentiality obligations attach to it, or whether the people in it were told it would be sold, the buyer discounts for the chance it has to pull the data later—or indemnify someone for it. Brokers who resell unauthorized data know this; it is why that data sells for pennies.
The seller controls provenance. The vendor agreements, the employee notices, the privacy policy as it read when the data was collected, and the NDAs sitting in the inbox are all on your side of the table. Documenting them before the broker’s intake call is the single highest-return piece of work in the deal, because it converts the buyer’s uncertainty into your price.
The dataset as the broker’s intake describes it
- Every system in the business: e-mail, Slack, CRM, POS, Airtable, shared drives
- Vendor and counterparty correspondence under NDAs, bundled in
- Personal information of sole-proprietor vendors and customers, undeclared
- “Anonymization” left to the broker, against no stated standard
- Rights and consents represented by you, on the broker’s paper
The same dataset after rights clearance
- A data map by source, with the governing agreement identified for each
- NDA-bound and trade-secret material excluded, with the exclusion documented
- Personal information de-identified to a named standard, or excluded
- A no-reidentification covenant that flows down to the model builders
- Rights representations narrowed to what the inventory supports
Sellers ask whether the inventory will shrink the deal. It removes sources that could not have been licensed in any case. What remains is cleaner, and cleaner data commands a higher price.
The broker’s intake is mapping your entire enterprise
The intake questionnaire does not ask for a dataset. It asks for every system where the business collects information: employee e-mail accounts, Slack, Airtable, the CRM, point-of-sale reporting, delivery-platform data, customer and e-mail lists, and event imagery. That is the company, and most of it is governed by obligations that have nothing to do with anonymization.
The founder usually reads the offer as a one-time licensing decision with found money attached. For an operating company it is a one-time decision layered on top of vendors, employees, and partners who will still be there after the deal closes. The question is which sources can be licensed without breaching a promise you made to one of them.
Why “anonymized” does not settle it
Brokers lead with anonymization because it answers the question founders ask first. It does not answer the questions that create liability. If your vendor agreements contain confidentiality provisions, those provisions survive the removal of names. If your inbox holds correspondence with a counterparty under an NDA, the NDA does not release the material because the broker stripped the signature block. If a vendor’s pricing, volumes, or product plans sit in your sales reporting, that is the vendor’s trade secret regardless of whether the vendor is identified.
There is also a technical problem. Operating-company data is small-n. A dataset with 40 vendors in one market can be de-identified in name and still identify every one of them by product, volume, and week. De-identification that holds up is defined against a standard, performed before transfer, and warranted by the party who did it. The agreement should say who de-identifies, when, against what standard, and what happens if the standard is not met.
When a license is a “sale”
If your company is subject to California’s consumer privacy law, a paid license of personal information is a sale under that statute, and a sale carries notice and opt-out obligations whether the buyer is an advertiser or a model developer. The exemptions that once kept business-contact and employee data outside the law have expired, so the identifiable data of a sole-proprietor vendor or of your own staff is personal information for this purpose.
The practical consequence is that the privacy policy that was in force when the data was collected governs what you may do with it now. If that policy did not disclose licensing for AI training, the data in question is either de-identified to the statutory standard before it goes anywhere, or it does not go.
The employee inbox problem
Employee e-mail is the source brokers want most and the source with the most people in it who never agreed to anything. Every inbox contains third-party correspondence: customers, vendors, candidates, counterparties, and your own lawyers. Licensing it transfers communications from people who are not parties to the deal, and in the case of communications with counsel, risks waiving privilege by handing them to a third party.
New York requires private employers to give written notice before monitoring employee e-mail and internet use. A company-property clause in an employment agreement establishes that the company owns the account. It does not by itself tell employees their messages would be sold for model training, and it says nothing to the people on the other end of those messages. Employee e-mail is usually the first source we recommend excluding or heavily filtering, and sellers are often surprised at how little of the offer depends on it.
The broker’s compliance failure becomes your investigation
Founders ask how a regulator would ever find out. The realistic path is not an employee complaint. It is a state regulator investigating one of these brokers for operating without registering as a data broker—California, Texas, Oregon, and Vermont each maintain a registry—and your contract surfacing in the production. The penalty to your company in that scenario may be modest. The cost of responding to a regulator, collecting documents, and producing them is a separate and larger number, and it arrives whether or not you did anything wrong.
The inventory produces that exposure figure in writing, by source, so the decision to license is made against a known downside rather than a broker’s assurance that compliance is handled.
Training is irreversible, so termination rights do little
A conventional data license lets the licensor terminate and demand deletion. Once a model has been trained on your data, deletion of the copy does not remove what the model learned, and no termination clause changes that. The protections that matter are agreed before transfer: the scope of the license, the standard for de-identification, a no-reidentification covenant that flows down to every model builder in the chain, representations narrowed so clearance risk does not sit entirely with you, indemnity that runs in both directions, audit rights over the broker’s downstream agreements, and clear treatment of the data after the term ends.
The broker’s form will not contain most of these. The seller who arrives with an inventory and a position on each of them negotiates from the stronger side, because the buyer wants the provenance file as much as the data.
How we work
The engagement is built around the inventory, because nothing else in the deal can be done well without it.
1. Inventory and rights clearance
- A data map by source across every system the intake covers
- Review of the agreements that govern each source: vendor, NDA, employee, and the privacy policy as collected
- A rights-clearance matrix: licensable as-is, licensable only aggregated or de-identified, or excluded, with the reason for each
- A written exposure assessment covering regulatory, contractual, and vendor-relationship risk
- A go/no-go recommendation by source you can put in front of the broker
2. Agreement negotiation
- “Anonymized” tied to a defined de-identification standard
- No-reidentification covenant with flow-down to model builders
- Rights and consents representations narrowed to the inventory
- License scope, exclusivity, indemnity, audit, and post-term handling
- Handled across one or several broker agreements
3. Clean footing for the next sale
- Vendor agreement language that permits licensing going forward
- Privacy policy and notice-at-collection updates
- Employee monitoring notice and account-ownership terms
- So the second offer closes on a cleaner record than the first
Fees. Every engagement is fixed-fee, quoted in writing after scoping, and payable without regard to whether the deal closes. We do not take a percentage of the proceeds, because advice on what to exclude should not come from counsel paid on what stays in.
Minimum. Our minimum engagement is $6,000. If the offer in front of you is below roughly $100,000, the legal spend is unlikely to pay for itself, and we will say so.
Who does the work. The inventory and the negotiation are handled by Yelena Ambartsumian (CIPP/US, AIGP) with 1 senior teammember. The firm has taken a company’s first sale of a proprietary dataset from inventory through closing, and advises AI-native companies on the buyer side of training-data provenance regularly.
Questions sellers ask
Will the inventory shrink my deal?
It removes sources that could not be licensed without breaching an obligation you already have, which is the same thing the buyer’s counsel would remove at a discount later. What remains is documented and de-identified to a stated standard, and cleaner data commands a higher price per record. In our experience the cleared dataset is smaller than the broker’s wish list and is worth more to the buyer, because the buyer can defend having bought it.
Do I have to tell my vendors?
It depends on what your vendor agreements say and what the data reveals. If a vendor’s confidential information sits in the source, the agreement usually requires consent before disclosure, and anonymization does not satisfy that requirement. If the source is your own transaction data with no vendor confidential information in it, notice may not be required. The inventory answers this vendor by vendor rather than in the abstract.
How would a regulator even find out?
Through the broker. State regulators have begun investigating data brokers for registration and disclosure failures, and a broker’s seller contracts are among the first documents requested. Your agreement surfaces in someone else’s investigation, which is why the agreement needs to read well on its own.
What is the realistic penalty?
For a seller, the direct penalty in a broker investigation is often in the tens of thousands of dollars. The cost of responding—collecting documents, preparing a production, and dealing with the regulator—is a separate figure of similar size or larger. The exposure assessment puts both numbers next to the offer so the decision is made with the downside quantified.
Can I license photos of my customers or event attendees?
Images of identifiable people are personal information, and in several states a person’s likeness carries its own rights. Whether you can license them turns on what your event terms, photo release language, and privacy policy said at the time, and on whether faces can be removed without destroying the value of the imagery. Event and product imagery is often licensable; attendee imagery often is not.
The broker says it handles compliance. Is that enough?
Read the agreement for what “handles” means. In most broker forms it means the broker registers itself, de-identifies after transfer against its own standard, and takes a representation from you that you had the rights to send everything you sent. The compliance risk that matters to you is allocated to you in that structure. The negotiation moves it.
Do you represent buyers as well?
We advise AI-native companies on training-data provenance, de-identification warranties, and indemnity structures from the buyer side, and that work informs how we negotiate for sellers. We do not represent both sides of the same transaction, and we run conflicts on every broker and platform before a scoping call.
If you have an offer in hand
Send the name of the broker or platform so we can clear conflicts, the size of the offer, and the systems the intake covers. We will confirm whether an engagement makes sense at the offer size and, if it does, send a fixed-fee scope in writing. Please do not send agreements or data until an engagement letter is signed.
Start an inquiryThis page is general information. It is not legal advice and does not create an attorney-client relationship. Outcomes depend on the specific data, agreements, and jurisdictions involved. AMBART LAW PLLC, New York, NY.