Do I Own My AI Product?

Less of it than your financing documents are about to say you do. Every priced round asks the company to represent that it owns or has the right to use its intellectual property, and for an AI company that one sentence is doing four separate jobs: the people who built the product, the material your own tools generated, the model you sit on top of, and the data you describe as a moat. Each one fails differently. Three of the four are fixed with paperwork you could start signing this month.

The sentence you are going to sign

Somewhere in your stock purchase agreement is an intellectual property representation. You will represent that the company owns, or has sufficient rights to use, everything the business runs on, and that nobody else has a claim on it. Your investor's counsel will spend more diligence time on that one sentence than on almost anything else in the package, for the good reason that it is the representation most likely to be wrong.

I have sat on both sides of this part of a financing, and until recently it was close to mechanical on a software company. Assignment agreements, open source scan, close the file. AI companies have made it interesting again, and the standard diligence list has not fully caught up. That gap is worth something to the founder who knows where the soft spots are.

The people who built it

Start with the people. This is the oldest gap in the file and still the one that most often slows a round down, and I have written about it before without it getting any less common.

An employee's work inside the scope of the job belongs to the company automatically. A contractor's does not. The copyright statute lets you commission a "work made for hire" in only nine listed categories, and software is not one of them, which means the freelancer who built your first data pipeline owns what he wrote until he signs an assignment. The same goes for your designer, and for the friend who stood up the retrieval layer one weekend in exchange for equity nobody ever papered.

Then run it at the founders. Who was still employed somewhere else while the prototype was being written, whose laptop was it on, and what did that employer's invention assignment agreement say, a document most people have not opened since the day they signed it. The version I see most often is a prototype built by a technical co-founder who was two weeks from his last day at a large company, on a laptop that company had issued him.

None of this is interesting, and all of it is the first thing a diligence lawyer checks.

The code your tools wrote

Copyright protects what people create. The Copyright Office has held that line, and its January 2025 report on copyrightability put the point about as plainly as an agency ever does: prompts alone do not give a user enough control over the expressive result to make that user the author of the output. What you contributed is protectable. The rest is not, and not for anybody else either.

So look at how your product got built. If a real share of your codebase came out of an assistant with light human editing, nobody owns that share, including the competitor who copies it. Registration runs on the same rule. You cannot bring an infringement suit on an unregistered U.S. work, and registering something with AI-generated material in it means disclosing that material and carving it out of the claim.

What this changes in practice is the question you get asked. Diligence used to ask whether you own the code, and the answer lived in a stack of signed agreements. Now it asks how the thing was built, and that answer lives in a written AI-use policy and a record of which tools were used where. Teams that have neither spend an afternoon reconstructing it from memory, in front of people taking notes.

The model underneath

Your provider's contract almost certainly handles the ownership question cleanly. OpenAI's services agreement says the customer owns all Output and assigns over whatever rights OpenAI has in it. Anthropic's commercial terms say much the same, and add that Anthropic will not train on customer content. Good drafting, and it should reassure you, and it still cannot deliver what the provider does not have. Where material has no human author there is no copyright to assign, to you or to anyone. What you actually hold is protection against a claim by your vendor, which is worth having and is a smaller thing than an exclusive right.

Two other clauses in that same contract are worth ten minutes. Most provider agreements restrict using outputs to develop a competing model, with narrow carve-outs for classifiers and for fine-tuning through the provider's own service, so if your roadmap involves distilling your own model from a frontier provider's outputs, read that language before it goes in the deck. And open weights are not the public domain. Llama's community license wants "Built with Llama" displayed, a specific attribution notice carried in a NOTICE file with any distribution, and the word "Llama" at the front of the name of any model you fine-tune with it. Other open-weight licenses restrict commercial use outright. Diligence will ask for the list, and "we did not realize that was a condition" is a weak answer when the fix was a line of footer text.

Fine-tuning deserves one more line, because founders spend real money on it. A fine-tuned hosted model is usually not a portable asset. The customization lives on the provider's platform, under the provider's terms, and it can be a genuine advantage without ever becoming a thing you own and can carry away.

The data you call a moat

If your deck says the product improves with every customer interaction, somebody is going to read your customer agreements to see whether you are allowed to do that. Read them first.

Look at what your first ten contracts say about using customer data to improve the service. Silence is not permission, and sophisticated buyers now strike those rights on purpose. A moat built from data you had no contractual right to accumulate is a product story your paper does not support, and at exit it stops being a diligence finding and becomes somebody's integration problem. Then run the question the other way: where did your training and evaluation data come from, and can you show it. If you are already revisiting your customer-facing terms, do both in one pass.

What is actually yours

For most AI startups the real asset was never the copyright anyway. It is the trade secret: the pipeline, the evaluation sets, the retrieval architecture, the prompts that took nine months to get right, the failure modes you learned about expensively. Trade secret law has no human authorship requirement. What it has instead is a requirement that you took reasonable steps to keep the material confidential, and a lot of early companies have not.

In practice that means confidentiality agreements that reach this material specifically, access controls that exist somewhere other than a policy document, and some restraint about how much of your architecture goes on a conference stage. It costs very little. It protects the part your competitors actually want.

Before the data room opens

Four things, and you can start all of them this week without calling a lawyer.

Get signed assignment agreements from every founder, employee, contractor, and advisor. Write a one-page AI-use policy for your engineering team and start keeping a record of what got used where. Build a schedule of every model and open-weight license the product depends on, with the conditions each one imposes. Read your ten largest customer contracts for what they let you do with customer data.

All of it is cheaper this month than it will be with a data room open and a closing date on the calendar. I have watched more than one team do this work during the raise instead of before it. Same work. It costs more, and it happens in front of an audience.

Frequently asked questions

Do I own the code an AI coding assistant wrote for my startup?

Not as copyright, at least not the parts the model produced on its own. Copyright protects human authorship, and the Copyright Office has concluded that prompting alone does not make the user the author of the output. The practical question in a financing is less about ownership than about description: investors want to know how the product was built, which tools were used where, and what a human reviewed.

My AI vendor's contract says I own the output. Isn't that enough?

It is worth having and it is not the whole answer. Providers like OpenAI and Anthropic assign customers whatever rights they hold in the outputs, which protects you against a claim from the vendor. It cannot create an exclusive right that does not exist. If the material has no human author, there is no copyright for anyone to transfer, and nothing stops a competitor from producing something similar.

Do I still need assignment agreements if everyone on the team is an employee?

Yes, and you need them most from the people who were not employees at the time. Work created by an employee within the scope of employment belongs to the company automatically. Work by a contractor does not, because software is not one of the nine categories the copyright statute allows to be commissioned as a work made for hire. Founders who built a prototype before incorporation, or while employed elsewhere, are the usual gap.

Does building on an open-weight model create a diligence problem?

Only if you ignore the license. Open weights come with conditions, not with the public domain. Llama's community license, for example, requires displaying "Built with Llama," a specific attribution notice in a NOTICE file with any distribution, and beginning the name of any model you fine-tune with Llama with the word "Llama." Some other open-weight licenses restrict commercial use entirely. Keep a schedule of every model in the product and what its license requires.

Can my startup use customer data to improve its model?

Only if the customer contracts say so. Silence is not permission, and enterprise customers increasingly negotiate improvement rights out on purpose. If the pitch describes a data advantage that accumulates across customers, the customer agreements need to actually support it, both for the financing and for any eventual acquirer.

If copyright does not protect an AI product, what does?

Usually trade secret, and it is a better fit anyway. The pipeline, the evaluation sets, the retrieval architecture, and the prompts that took months to tune are all protectable as trade secrets, and trade secret law has no human authorship requirement. It does require reasonable steps to keep the material confidential, which means real confidentiality agreements and real access controls rather than a policy nobody follows.