Skip to content
Outrich Data

Trade secrets and selling company data

Updated 2026-11-16 · 6 min read · By the Outrich Data team

Selling data can end trade secret protection if secrets are disclosed without confidentiality terms. Exclude what gives your company its edge, license the rest under use and confidentiality restrictions, and never include information you hold under an NDA. The Defend Trade Secrets Act and state laws require reasonable secrecy measures. General information, not legal advice.

This page is general information, not legal advice. Whether your company can license a given dataset depends on your contracts, your privacy notices and the law where you and your customers are. Review it with your own counsel before signing anything.

What a trade secret is

In the United States, trade secrets are protected by the federal Defend Trade Secrets Act and by state laws. Nearly every state has adopted a version of the Uniform Trade Secrets Act. The definitions are similar. Information qualifies if it has independent economic value because it is not generally known or readily ascertainable by others who could profit from it, and if its owner has taken reasonable measures to keep it secret.

The category is broad. Formulas, source code, pricing models, customer lists, supplier terms, process know-how and compiled internal data can all qualify. No registration exists. Protection lasts as long as the information stays secret and the owner keeps treating it as secret.

Trade secret law protects against misappropriation, meaning acquisition by improper means, or disclosure or use in breach of a duty. It does not protect against independent discovery or reverse engineering of a lawfully obtained product. It also overlaps with copyright and contract. A document can be copyrighted, confidential under an agreement and a trade secret at once, and each gives a different remedy.

This guide is general information, not legal advice. Trade secret disputes turn on specific facts, and the law differs somewhat by state and country. Review any planned license with counsel.

Reasonable measures, and how a sale can undo them

The reasonable-measures requirement is where data sales and trade secrets meet. Courts look at what the owner did to guard the information: access controls, confidentiality agreements with staff and partners, marking, and limits on who received it.

Disclosure does not by itself destroy a trade secret. Companies share secrets with vendors, partners and licensees routinely. What matters is the terms. A disclosure under a binding duty of confidentiality and restricted use is consistent with secrecy. A disclosure with no such duty, or to a party free to pass it on, is evidence that the owner did not treat the information as secret, and protection may be lost for good.

A data license is therefore a decision about each secret in the dataset. The owner can keep it out, or put it in under terms that preserve secrecy. The owner should not let it leave by accident.

What to exclude

The simplest protection is exclusion. AI labs buying operational data want to learn how work is done. They rarely need the specific facts that make one company more competitive than the next. The following are commonly kept out.

Secrets also hide inside ordinary records. A support ticket may quote a negotiated price. A chat thread may discuss an unannounced product. Exclusion by folder or system is a first pass, and a search of the remaining text for sensitive terms is the second.

A useful test for any category is to ask what a competitor could do with it. If the honest answer is very little, it is probably routine operational material. If the answer is that they could underbid you, approach your customers or copy your method, it belongs on the exclusion list.

  • Pricing models, margin data, discount schedules and negotiated customer terms.
  • Customer and prospect lists, and anything that reveals who buys what.
  • Core proprietary algorithms, formulas and the parts of a code base that are the product's real advantage.
  • Product roadmaps, unreleased designs and strategic plans.
  • Supplier contracts and cost data.
  • Security architecture, credentials and keys.
  • Board materials, financing documents and legal advice.

License restrictions that preserve secrecy

Where valuable confidential material stays in the dataset, the license has to carry the reasonable measures forward. The terms below are typical. None is a guarantee, and their enforceability depends on the drafting and the governing law.

A seller should also decide who the buyer may be. Approval of buyer categories lets a company keep its data away from competitors and from buyers who serve them. Outrich Data's standard commitment is that a seller approves each dataset package, buyer category and license scope in writing before delivery.

Remedies deserve attention before signing, not after. Damages are hard to prove when the harm is that a model learned something. A clause that entitles the seller to seek an injunction, and that names the governing law and court, gives a faster route. Under the Defend Trade Secrets Act, a civil action generally must be brought within three years of when the misappropriation was discovered or should have been.

TermWhat it does
Confidentiality clauseBinds the buyer to keep the dataset secret and limits access to people who need it
Purpose limitationRestricts use to named purposes, such as training or evaluation, and bars others
No redistributionBars resale, sublicensing or publication of the raw data without consent
Security requirementsSets storage, access and transfer standards for the delivered files
No extractionRequires reasonable steps so that models do not reproduce the source records verbatim
Return or deletionRequires delivered copies to be deleted at the end of the term, with certification
Audit and remediesGives the seller a way to check compliance and seek an injunction on breach

Training cannot be undone

One feature of AI licensing has no close parallel in older trade secret practice. Delivered files can be returned or deleted. What a model has learned from them cannot be removed in any practical way.

Models can also reproduce fragments of their training data, particularly text that is distinctive or repeated. A license can require the buyer to take steps against that, but it cannot make the risk zero. This is the strongest reason to exclude true secrets, as opposed to relying on contract terms to protect them. Anything a company could not tolerate a competitor learning should stay out of the dataset.

The practical consequence is a different way of sorting information. Older licensing practice asked whether the recipient could be trusted to keep a secret. Here the further question is whether the secret matters once it is absorbed in diluted form into a general model. For most operational detail it does not. For a formula, a pricing rule or a named customer relationship it might.

Other people's secrets

The second half of the problem is information that belongs to someone else. Company systems are full of it: customer documents received under NDAs, client data held under service agreements, vendor pricing shared in confidence, and material from partners under joint development terms.

Under the Defend Trade Secrets Act and state law, misappropriation includes disclosing another's trade secret without consent when the discloser knew it was acquired under a duty to keep it secret or limit its use. Courts can order injunctions and damages, and for willful and malicious misappropriation can award enhanced damages and attorney's fees. A breach of the NDA or service contract is a separate claim on top.

De-identification does not cure this. Removing a client's name from its confidential specifications leaves the confidential specifications. Material held under a confidentiality duty has to be excluded unless the owner gives written consent. Employees sometimes bring documents from former employers as well, and those do not belong in a dataset either.

Open source and licensed third-party code need the same care in a code base. The company holds that code under someone else's license terms, and those terms decide whether it can be passed on. Identify vendored and third-party components, and leave them out unless their licenses permit the use.

A short process

Before licensing, list the categories of information the company treats as secret and confirm they are excluded. List the confidentiality obligations the company owes to others, by contract, and exclude what they cover. Search the remaining records for stray sensitive terms. Then read the license against the table above, and have counsel confirm that the confidentiality and use terms are enforceable under the governing law.

Keep a record of each step. If the company's trade secret protection is ever challenged, evidence that it reviewed, excluded and restricted before licensing is evidence of reasonable measures.

Common questions

Do we lose trade secret protection by licensing data?
Not necessarily. Disclosure under binding confidentiality and use restrictions is consistent with keeping a secret. Disclosure without them can end protection. The safest course is to exclude true secrets from the dataset and license the rest under restrictive terms.
Can we license documents a customer gave us under an NDA?
Not without the customer's written consent. The NDA restricts disclosure, and the information may also be the customer's trade secret. Removing the customer's name does not change who owns the information or what you agreed.
Is our internal data a trade secret?
Parts of it may be, if it has economic value from being secret and the company has taken reasonable measures to protect it. Routine operational records often have little competitive sensitivity. Pricing, customer lists and proprietary methods often have a great deal.
Can a model reveal what was in the training data?
Sometimes. Models can reproduce distinctive passages from training data. A license can require safeguards, but learning cannot be reversed after training. Decide what to include on the assumption that it cannot be taken back.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc