How AI data deals are structured

Most owners first hear about data deals as a single number: "a lab paid a company X for its Slack history." That's the headline. Underneath it is a structure, and the structure decides how much risk you carry, how much of your data leaves the building, when you get paid and whether you can do it again next year.

This guide walks through how deals for de-identified company data are usually put together, the pieces that vary, and the trade-offs between them. It's written for owners and COOs of companies with 20 to 200 employees who want to understand the shape of an agreement before they ever see one.

General information, not legal advice. Have counsel review any agreement before you sign.

The five building blocks of every data deal

Whatever the paperwork looks like, almost every agreement answers the same five questions.

Everything else, warranties, liability caps, deletion terms, is there to protect one side or the other within those five choices.

Scope: the part you control most

Scope is where owners have the most leverage and where the biggest mistakes happen. A good scope is specific: named systems, a date range, a list of exclusions and a description of the de-identification applied.

A weak scope says "company communications." A strong one reads more like this:

Slack channels in the Operations, Dispatch and Customer Escalations categories, January 2018 through June 2026, excluding direct messages, private channels, the #hr and #leadership channels and any shared channels with outside organizations; de-identified per Schedule B.

That level of detail protects you in two ways. The buyer can't later claim they were promised more, and you can see exactly what's leaving your systems. It also tends to help value: buyers can evaluate a well-defined dataset faster and with more confidence.

For help building the list of what you have, see how to inventory your company data.

Rights: what the buyer can do

The grant of rights is the heart of the agreement. In practice most grants for training data land on a version of: a non-exclusive license to use the dataset to train, evaluate and improve the buyer's models.

The points worth negotiating:

• Who can use it. Just the buyer, or affiliates too? Contractors working for the buyer?
• What it can be used for. Training and evaluation are standard. "Any purpose" is not.
• Resale and sublicensing. Most owners prohibit resale of the raw dataset.
• Publication. Buyers sometimes want to publish small samples in research. You can prohibit it or require approval.
• Re-identification. Should be flatly prohibited, with notice obligations if it happens anyway.

Exclusivity: the biggest lever after scope

Exclusivity means you agree not to license the same data, or similar data, to anyone else. It's the term most likely to move an offer, and the term most worth thinking hard about.

Here's the insider view: owners often assume exclusivity is expected. It usually isn't. Many deals are non-exclusive, and giving exclusivity away without a clear premium is one of the most common avoidable mistakes we see.

Deliveries: one archive or a relationship

A one-time delivery is a single archive covering a fixed date range. It's simple, finite and easy to explain to a future acquirer.

A repeat-delivery structure adds later refreshes: the next six or twelve months of data, delivered under the same terms. It can add value for both sides, but it creates ongoing obligations: someone at your company has to run exports, and de-identification has to happen each time.

If you choose repeat deliveries, make sure the agreement says:

1. How often deliveries happen and what each contains
2. Whether you can stop future deliveries, and with what notice
3. Whether pricing for later deliveries is fixed now or negotiated each time
4. What happens to future deliveries if your company is sold

Payment: timing matters as much as amount

We don't publish prices and won't promise one, because every dataset is different. But the mechanics of payment are fairly standard.

• Lump sum on acceptance. You deliver, the buyer has a set window to confirm the delivery meets the spec, then pays.
• Milestones. Part on signing, part on delivery, part on acceptance. Useful for large or staged scopes.
• Per delivery. In repeat structures, each delivery is paid on its own acceptance.

The clause to read carefully is acceptance. A tight acceptance clause, with objective criteria and a short window, protects you from open-ended delays. A vague one gives the buyer room to hold up payment.

For how offers are actually priced, see how AI data pricing works.

Protections that belong in every structure

Regardless of shape, these terms should be in the agreement:

• A precise dataset definition with exclusions
• A ban on re-identification and on attempts to contact people in the data
• Security requirements for storing the dataset
• Retention limits and a deletion certificate at the end of the term, where applicable
• Mutual confidentiality, including about the deal itself if you want it
• Limits on publicity and use of your company name
• Reasonable, mutual liability caps

Our licensing agreement checklist covers each clause in detail.

Three example structures

To make it concrete, here are three composite structures for companies of similar size.

Where a broker fits

A broker's job is to turn your data into a clear, comparable description, get it in front of several buyers, and help you compare structures, not just numbers. Two offers with the same headline can be very different deals once you look at exclusivity and acceptance terms.

DataOffer takes no upfront fee and shares nothing until you approve the buyer, scope, price and terms.

Red flags in a term sheet

A few terms deserve a second look whenever they appear:

• "Any purpose" grants. Training and evaluation are standard. Open-ended use is not.
• Acceptance with no deadline. Payment that depends on an acceptance window with no end date.
• Exclusivity without a premium. If a buyer wants exclusivity, it should be reflected in the offer.
• Uncapped indemnities from you. Liability should be proportionate and mutual.

Frequently asked questions

Can a data deal be structured around a pending sale of my company?

Yes. Non-exclusive licenses with clear use limits are the easiest to carry through a sale. Tell your deal counsel and the buyer about any LOI before you sign.

What is the most common structure for an AI data deal?

A non-exclusive license of a defined, de-identified dataset, delivered once and paid after acceptance. Repeat deliveries and exclusivity are less common and usually negotiated for a specific reason.

Do AI data deals include royalties?

Royalty structures are uncommon for training data, because it's hard to trace value from one dataset to one model. Lump sums and milestone payments are more typical.

Can I license the same data to more than one buyer?

Yes, if your agreements are non-exclusive. That is one of the main reasons owners prefer non-exclusive licenses.

Who owns the models trained on my data?

The buyer does. Your agreement controls the dataset itself: how it can be used, stored and deleted.

How long is a typical license term?

It varies. Some licenses run for a fixed term with deletion afterward; others are perpetual for training already done. Make sure the term and what happens at the end are written down clearly.

Is a data deal taxed like a sale of an asset?

Tax treatment depends on structure and your situation. Ask your accountant before you sign, not after.

Getting started

If you'd like to see what structures buyers would offer for your data, share rough estimates of your systems and history. We'll come back with an illustrative range and, when you're ready, competing offers you can compare side by side. Get an offer.

This guide is general information, not legal, tax or financial advice. Figures and ranges are illustrative; talk to qualified advisors about your situation.