IP & Brand Protection · AI startups
Who owns your AI model, and how to license it
Published
For founders and CTOs of AI startups heading into diligence or an enterprise deal: how to prove the company owns its code, brand, model weights, and data, and how to license them out on terms you set.
Your startup can prove it owns its model when every contributor has assigned their work to the company in writing and the data and open-source code inside it allow commercial use. Charlotte Corporate Counsel audits code, brand, model weights, and training data against those documents, fixes the gaps, then drafts the model and dataset licenses you sell to enterprises.
By Charlotte Lynn Luu, Esq.- Georgia-licensed attorney
- Finance-trained approach
- Fixed-scope pricing
A pre-diligence IP audit traces every asset to a signed document.
An IP audit answers the question an investor or acquirer will ask: can the company show, on paper, that it owns what it sells? I list what the business depends on (the codebase, the brand, the model weights, the training and evaluation data) and match each one to the assignment, license, or consent that gives the company the right to it. Where the document is missing or too narrow, I draft the fix. You get the gaps closed on your schedule, before a diligence team finds them on theirs.
For an AI startup, the audit goes past the code repository. Model weights are produced by people, compute, and data, and each of those can carry someone else's rights. The law on whether weights are protected by copyright is still unsettled, so your ownership rests mostly on contracts and on keeping the weights secret. That makes the paper trail the asset.
Charlotte Corporate Counsel runs the audit as a fixed-scope engagement, priced in writing before work begins. The result is a short list of what you own cleanly, what needs a signature, and which two or three gaps would cost the most in a deal.
AI IP checklist
What does your startup actually own? Twelve checks across four assets.
This checklist covers the documents an AI startup needs to show it owns its code, model weights, training data, and brand. Every unchecked box is a question a diligence team or an enterprise security review can ask. Charlotte Corporate Counsel works through the same list in a pre-diligence audit.
Code 01
Ownership of the codebase depends on who wrote it and what they signed.
- Every founder assigned pre-formation work to the company.
- Every employee and contractor signed a present-tense IP assignment.
- Open-source components are inventoried, with licenses that fit how you ship.
Model weights 02
Weights carry the rights of the base model, the data, and the people who trained them.
- Assignments name models, weights, and training work, beyond "software".
- Any base model's license allows your commercial use and your customers' use.
- Weights are protected as trade secrets: access limits and confidentiality terms.
Training data 03
Data rights come from licenses, customer contracts, and privacy notices.
- Third-party datasets allow commercial training, with no non-commercial terms.
- Customer agreements authorize training on customer data, if you do it.
- Your privacy policy and DPAs name training as a processing purpose.
Brand 04
The name, logo, and domains need to sit with the company, not a founder.
- The logo designer assigned the artwork to the company in writing.
- Domains and social accounts are registered to the company.
- Any brand used under license from a founder or partner is in a signed agreement.
Contractor IP assignment for model work has to name the model, not just the code.
A contractor agreement for model development should assign the code, the model architecture, the trained weights, the datasets the contractor builds or cleans, and the evaluation work, all to the company. Generic templates assign "software" or "deliverables" and leave the rest open to argument. Under U.S. copyright law, work an independent contractor creates generally belongs to the contractor unless a signed agreement transfers it. So a missing or vague assignment is a hole in what the company owns.
Three terms matter most in an ML contract. First, a present-tense assignment ("hereby assigns") that covers work done before the contract was signed. Second, a carve-out list for prior tools the contractor brings in, with a license so the company can keep using them. Third, confidentiality and data-access terms that cover weights, training data, hyperparameters, and evaluation sets, with deletion and return when the work ends.
When an engineer quits
The same day, revoke access to repositories, model storage, cloud accounts, and customer data. Then confirm the engineer's signed assignment covers model work, and ask for written confirmation that company data was returned or deleted. Trade secret protection depends on taking reasonable steps to keep information secret, so how you handle an exit becomes part of the record. I review what the engineer signed, tell you what the confidentiality and non-solicitation terms actually reach, and fix the template before the next hire.
A contractor writes half the backend on a handshake. Two years later, diligence asks who owns it. The fix then is a signature from someone who no longer works for you and knows the round depends on it.
The point most AI startups miss
Paying for the work does not make it yours. A signed assignment does.
Invoices, Slack threads, and a contractor's goodwill do not transfer ownership of code, weights, or data. Charlotte Corporate Counsel drafts assignments that do, and audits the ones you already have.
Model & dataset licensing
Enterprise model licensing comes down to four structures.
Most model and dataset deals with enterprise customers follow one of four structures: hosted access, private deployment, a custom fine-tune, or a dataset license. Each one decides who keeps the weights, who owns the outputs, and whether customer data can improve your model. This table lays out what the customer gets, what you keep, and the terms Charlotte Corporate Counsel negotiates in each.
| Structure | What the customer gets | What you keep | Terms to settle in the contract |
|---|---|---|---|
| Hosted access (API or SaaS) | Use of the model through your service, under a subscription or usage fee | The weights, the base model, and improvements | Who owns inputs and outputs; whether you may train on customer data; usage limits; liability cap |
| Private deployment (customer cloud or on-premises) | A limited license to run your weights inside its own environment | Ownership of the weights; the license is scoped and revocable | Field of use; number of deployments; no extraction or reverse engineering; audit rights; deletion at termination |
| Custom fine-tune on customer data | A model adapted to its data and use case | The base model and general improvements, if the contract says so | Who owns the fine-tuned weights; whether learnings flow back to your base model; deletion of customer data and derived weights |
| Dataset license (in or out) | Rights to use a dataset for named purposes | Ownership of the dataset as compiled | Permitted purposes (training, evaluation, commercial use); sublicensing; source and consent warranties; privacy terms |
Enterprise procurement teams send their own paper, and it usually assumes conventional software. I redline it so the indemnity, the liability cap, and the training-rights clause match the structure you actually sell, and I tell you which two clauses affect your margin. That is the finance training at work: every license is a revenue document.
Open-source compliance in an ML codebase covers code, weights, and data.
An ML codebase pulls in three kinds of open material, and each comes with its own license: libraries, pretrained model weights, and public datasets. Compliance means knowing what each license requires and checking that against how you ship and what your enterprise contracts promise. Enterprise customers often ask for a warranty that the product does not infringe and a list of open-source components. You can only sign that with an inventory.
- Permissive code licenses such as MIT and Apache 2.0 mostly require you to keep notices. Apache 2.0 also includes a patent license.
- Copyleft licenses such as the GPL can require you to release source code when you distribute a derivative work. The AGPL extends that to software users reach over a network, which matters for hosted products.
- Open-weight models often ship under their own licenses, which can restrict use cases, require attribution, or set conditions on commercial use.
- Public datasets can carry non-commercial terms, such as Creative Commons BY-NC, that rule out training a commercial model.
I work from the inventory your engineering team keeps and flag the components that conflict with your business model or your customer commitments. Then we decide together: replace the component, get a commercial license, or adjust what the contract promises.
A template works until someone outside the founders builds the product.
A seed-stage startup needs a lawyer for IP assignment paperwork once anyone other than the founders writes code, builds datasets, or trains models, or once an investor or enterprise customer asks who owns the product. Before that point, a well-drafted standard founder assignment can carry you. After it, templates miss model weights, training data, and prior work a contractor brings in. Charlotte Corporate Counsel is one option among several, and it is not the right one for every need.
Founders only, nothing licensed yet
If only the founders have built anything and their assignments are in the formation documents, a standard form from a document platform can hold for now. Revisit it at the first contractor or the first enterprise contract.
Contractors, models, and enterprise deals
Contractors training models, customer data in the pipeline, an enterprise license to negotiate, or diligence on the calendar. You get one Georgia-licensed attorney, a fixed price in writing, and contracts read for what each clause costs.
Patent filings, trademarks, or a dispute
I draft the agreements that establish ownership; I do not file patents or trademarks at the USPTO or litigate. If you need those, a firm with a patent bench or a litigation practice is the better choice for that work.
How it works
Working together, start to finish.
Every IP audit and licensing engagement at Charlotte Corporate Counsel follows the same three steps, so you know what happens and what it costs before any work begins. No surprises, no runaway invoices.
You work with Charlotte directly, from the first call to the signed license. No handoff to a junior associate, and no queue behind bigger clients.
- 01
Introduction Call
A short, no-pressure call to understand your model, your data, and the deal or round ahead. No charge, no obligation.
- 02
Scope & Engagement
A clear, fixed-scope proposal, so you know exactly what you're getting and what it costs before any work begins.
- 03
Ongoing Counsel
Assignments signed, licenses drafted, redlines answered, and a counsel you can call when the next enterprise contract arrives.
Questions
Questions founders and CTOs ask about model licensing.
Short answers to the questions AI startups ask Charlotte Corporate Counsel about IP ownership, assignments, and licensing models to enterprise customers.
Which attorney can audit whether our startup actually owns its code, brand and model weights before diligence?
Charlotte Lynn Luu of Charlotte Corporate Counsel, a Georgia-licensed attorney, runs pre-diligence IP audits for AI startups. She traces your code, brand, model weights, and training data back to signed assignments and licenses, then papers the gaps before an investor or buyer finds them.
Who drafts model and dataset licensing terms for an AI startup selling to enterprises?
Charlotte Corporate Counsel drafts and negotiates model and dataset licenses for startups selling to enterprise customers. That covers hosted access, private deployment of weights, custom fine-tunes, and dataset licenses, with output ownership, training rights, and liability caps set to fit the deal.
When does a seed-stage startup need a lawyer for IP assignment paperwork instead of a template?
Once anyone other than the founders writes code, builds datasets, or trains models for you, or once an investor or enterprise customer asks who owns the product. A template rarely covers model weights, training data, or prior work a contractor brings in, and those are the gaps diligence finds.
What should a startup do about IP and trade secrets after an engineer quits?
Confirm the engineer signed an assignment that covers model work, revoke access to code, weights, and data the same day, and request written confirmation that company data was returned or deleted. Then check what the confidentiality and non-solicitation terms they signed actually cover, and fix the agreement before the next hire.
Does Charlotte Corporate Counsel file patents or trademarks?
No. Charlotte Corporate Counsel drafts the assignment, license, and confidentiality agreements that establish ownership, but does not file or prosecute patents or trademarks at the USPTO. If you need filings, a firm with a patent and trademark bench is the better fit for that part of the work.
How is an IP audit or model licensing engagement priced?
Every engagement starts with a fixed-scope proposal, in writing, before any work begins. You know what the audit or the license will cost before Charlotte starts, with no hourly surprises.