AI-Assisted Deliverables: What Your Business Can Safely Promise Clients About Ownership When Generative AI Helped Create the Work

A branding agency signs a client agreement with a clause it has used for years: the client owns all final deliverables, free and clear, and the work is original to the agency. Nobody negotiates it.

On the project, a designer generates two dozen logo concepts with an AI image tool, picks a direction, redraws it, builds a palette and type system around it, and delivers a brand book. A developer on the same account uses an AI coding assistant for roughly a third of a client portal.

Eighteen months later, the client is being acquired, and the buyer's counsel asks for proof of chain of title. The question lands on the agency's desk: do you actually own what you assigned?

The answer is more complicated than the contract suggests, not because the agency did anything wrong, but because a promise to convey ownership cannot manufacture a property right that never existed.

Can AI-Assisted Work Be Copyrighted?

Start with the rule that is now settled. U.S. copyright protects "original works of authorship," 17 U.S.C. § 102, and the author must be human. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the D.C. Circuit affirmed the Copyright Office's refusal to register an image listing an AI system as sole author, holding that the Copyright Act requires every eligible work to be authored in the first instance by a human being. The Supreme Court denied review on March 2, 2026, ending that litigation, though a denial of certiorari is not an endorsement of the lower court's reasoning.

Here is what Thaler did not decide. Dr. Thaler disclaimed any human creative input, so the court had no occasion to say how much human involvement is enough for a work made with AI. The court said only that the human-authorship requirement does not prohibit copyrighting work made with AI assistance. Where the line falls remains open.

The most detailed answer available comes from the U.S. Copyright Office, in its March 2023 registration guidance and its 2025 report on copyright and AI. That guidance governs the Office's examining practice and carries persuasive weight with courts. It is not binding law. Treat it as the best available map, not the final word.

On that guidance, a work containing AI-generated material is protectable to the extent of the human contribution: expressive input perceptible in the output, creative selection and arrangement of human and AI-generated material, and creative modification of outputs. Purely AI-generated material is not protectable, and prompts alone generally do not confer authorship, because a prompt conveys an idea and copyright does not protect ideas. § 102(b).

This gives three categories worth keeping straight:

  • Human-authored work, with AI used for research or brainstorming: fully protectable.

  • AI-assisted work, where a person makes the expressive choices: protectable as to that contribution.

  • AI-generated material, produced by prompting without meaningful human control over the expression: not protectable.

One point reassures more than it alarms: AI involvement does not poison the whole work. Section 103(b) is the anchor: in a work incorporating preexisting or unprotectable material, copyright covers what the author contributed and does not extend to the rest. The agency above almost certainly owns something: the redrawn mark, the selection and arrangement, the brand system. What it may not own is the raw generated output underneath.

So keep records. What a human actually contributed- drafts, iterations, editorial decisions is what supports both registration and the scope of any assignment.

What Does Your Contract Actually Promise?

Four distinct things get collapsed into one clause.

A copyright assignment transfers an existing copyright. Section 204(a) requires a signed writing, and broad "all right, title, and interest" language will usually satisfy it. What it cannot do is convey a copyright that never came into existence. As to purely AI-generated material, an assignment conveys nothing because there is nothing to convey.

Work-made-for-hire does not solve this. It makes the hiring party the author of a qualifying work, but the work must still be human-authored to be copyrightable at all. And for independent contractors, work-made-for-hire reaches only nine enumerated categories in § 101 plus a signed writing. Most agency deliverables, websites, and software do not fit them, which is why assignment language matters more than the work-for-hire recital most contracts lead with.

Contract rights survive even where copyright does not. Parties can allocate delivery, use, non-assertion, and remedies by agreement regardless of whether copyright exists, and where an assignment fails, courts have found an implied nonexclusive license to use what was delivered. See Effects Associates, Inc. v. Cohen, 908 F.2d 555 (9th Cir. 1990). A client can get most of what it wants without owning a copyright.

A warranty is where the real exposure lives. Under North Carolina law, an express warranty is contractual, and liability does not depend on proof of negligence; it arises from the promise itself. So a provider who warrants that the work is "original," that it will not infringe, or that the client will "own" it is exposed when AI involvement makes that statement untrue, regardless of good faith and regardless of whether any copyright exists to fight about. The claim is breach of contract, not infringement.

Adding "to the extent protectable under applicable law" helps the ownership grant. It does not repair an originality warranty, a non-infringement warranty, or an indemnity. Those need their own treatment.

What Do Your AI Vendor's Terms Allow?

Most providers now say that, as between the provider and the customer, the customer owns the outputs. Read that sentence carefully. It allocates rights between those two parties. It does not make the output copyrightable, does not make it exclusive to you, and does not clear third-party claims. Two customers can prompt their way to similar outputs, and nothing in a vendor's terms prevents it.

Beyond ownership, five things vary enough to check for the specific product and tier you use:

  • Input licenses. You generally keep your inputs, but the provider takes a license to use them. The scope ranges widely.

  • Training and service improvement. Whether your inputs and outputs feed model training is the single most consequential tier-driven variable, and the words "train," "develop," and "improve" are not synonyms.

  • Confidentiality and retention. Are inputs treated as confidential information, and is there a defined retention and deletion practice?

  • Output indemnification. Where offered, it is typically narrow and carries exclusions — your own inputs, knowingly infringing use, disabling safety or citation features — and it is worth little unless carved out of the liability cap.

  • Use restrictions. Even where you "own" the output, terms commonly restrict sublicensing, reverse engineering, and training competing models, often through a separately linked acceptable-use policy.

Consumer, business, and API arrangements differ materially on these points, and providers revise them often. Verify the live terms for the exact product and tier before relying on any of it, and do not assume that what is true of one provider's enterprise offering is true of its consumer app or of a competitor.

Confidentiality, Trade Secrets, and Third-Party Rights

Using an AI tool does not automatically destroy trade-secret protection. Under the Defend Trade Secrets Act, 18 U.S.C. § 1836, and the North Carolina Trade Secrets Protection Act, N.C. Gen. Stat. § 66-152 et seq., protection depends on reasonable measures to maintain secrecy. A trade secret is lost when disclosed to someone under no obligation to protect it. So the question is not "did information go into an AI tool" but "who received it, and under what duty."

Entering client data into an approved business account, under terms that impose confidentiality and restrict training, with access controls and a written policy, is consistent with maintaining secrecy. Pasting the same data into a free consumer tool whose terms permit training and disclosure is not. The distinction is the account, not the technology.

Your NDAs and client agreements are a separate constraint. Many prohibit disclosure to third parties without consent, and a vendor is a third party. Check whether your agreements permit AI processing at all.

On infringement, keep two things separate. Output claims turn on whether specific output is substantially similar to a protected work; the theory that every output is an infringing derivative has been rejected as conclusory. Training-data claims turn largely on fair use, and the district courts are not uniform — one held that training on lawfully acquired copies was fair use while retaining a library of pirated copies was not; another granted judgment for the developer on a narrow record while expressly declining to bless the conduct generally; another found no fair use where the tool competed with the source. These are trial-level decisions, several on appeal, and none settles the law.

What this means for your business is narrower and more durable: a commercial user who reproduces protected expression is unlikely to be fully insulated, and a vendor indemnity runs to you from your vendor — it does not bind the upstream rights holder who has no relationship with either of you.

Two risks beyond copyright deserve mention. AI output can carry trademark or right-of-publicity exposure, which matters for logos and anything featuring a recognizable person. And AI-generated code can import attribution or copyleft obligations from open-source training material, creating license-incompatibility problems in a deliverable — a contractual risk as much as a proprietary one.

How to Protect Your Business

In your client agreements:

  • Assign the copyright you actually own, and pair it with broad delivery, use, and non-assertion rights so the client gets the practical benefit it expects.

  • Tailor the originality and non-infringement warranties. A knowledge qualifier is usually the negotiating fulcrum.

  • Decide your AI-disclosure position and state it, rather than leaving it to be discovered later.

  • Allocate third-party infringement risk deliberately, with indemnity and caps that match the warranty you actually gave.

In your operations:

  • Approve specific tools and tiers; require company accounts, not personal subscriptions.

  • Prohibit confidential client information in unapproved tools.

  • Document human creative contributions as you go.

  • Adopt a short written AI-use policy; it is also your reasonable-secrecy-measures record.

  • When registering a work containing AI-generated material, follow the Copyright Office's disclosure requirements. A registration that fails to disclaim unprotectable material can be vulnerable later.

These are starting points for your own counsel to adapt, not model clauses.

Before You Promise Ownership

Before promising exclusive ownership of AI-assisted deliverables, a business should know three things: what copyright actually protects in that work, what the vendor's terms for that tool and tier permit, and what its own client agreement guarantees. Those answers rarely line up by accident.

Isaboke Law Firm® works with agencies, consultants, developers, and other service providers on client service agreements, intellectual property provisions, and AI-use policies. If your contracts still carry the ownership and originality language you drafted before generative AI entered the workflow, we are glad to review what you are currently promising.

This article is for general informational purposes only. It is not legal advice and does not create an attorney-client relationship. The law in this area is developing, and outcomes depend on the specific work, contract language, vendor terms, and facts. Legal and vendor-terms research is current as of November 2026.

Previous
Previous

Do You Still Own Your Masters After a Music Distribution Deal?

Next
Next

Deadlock in a 50/50 North Carolina LLC