How to Cleanly Spin a Vertical AI Company Out of a Big Institution: IP, Non-Competes, and Founding Equity

Spinning a vertical AI company out of a big institution cleanly comes down to four things: know what your invention-assignment agreement actually assigns, check whether a non-compete binds you in your state, never build on the employer's time or systems, and keep the founding cap table clean. Get all four right before you resign.

The operators with the sharpest vertical AI ideas are usually the ones still inside the institution the idea would serve: the bank VP who sees how underwriting really works, the insurer who lives inside claims, the health-system CFO who knows exactly where revenue-cycle work breaks. The idea is downstream of the job. That is also what makes the exit delicate, because the same employer that gave you the insight has agreements that govern what you can take with you. Spinning out cleanly is a solvable problem, but it is one you solve before you resign, not after. This is informational, not legal advice; a startup attorney should review your specific agreements and state law before you act.

The single most common mistake is treating the legal cleanup as a post-resignation task. By then your options and your bargaining position have narrowed. Everything that protects a clean spin-out happens while you are still employed and thinking clearly: reading your agreements, mapping what is yours, and drawing a hard line between what you do on the job and what you do on your own. The founders who spin out without a dispute are almost always the ones who planned the separation as carefully as they planned the company.

This is a different question from whether to spin out at all, which is a strategic call about corporate antibodies and speed covered in spinning out a vertical AI company from a corporate role instead of building it inside. Assume you have made that decision. What follows is how to do it without leaving a legal thread the institution can pull later.

Most employees at a large institution signed an invention-assignment agreement, often buried in the onboarding stack, that assigns certain work you create to the employer. The critical distinction is between the general knowledge and skill you carry in your head, which is yours, and specific inventions or code made within the scope of your job, which usually are not. An idea for how AI could transform prior authorizations is not owned by your employer. A working prototype you built on a company laptop using company data very likely is.

Here is the practical split most agreements and state laws land on, though the exact lines are set by your specific contract and jurisdiction.

The clean move is to not build anything until you have left, or to build only on your own time, your own hardware, and public information. Several states, California among them, limit how far invention-assignment agreements can reach into work done entirely on your own time without company resources, but you should never rely on that as a shortcut. Read the agreement, and if the idea is close to your day job, have counsel read it too.

Founders routinely overestimate or underestimate their non-compete exposure because they assume there is a single national rule. There is not. The Federal Trade Commission's rule that would have banned most non-competes was blocked by federal courts before it took effect and then abandoned, so enforceability in 2026 is governed entirely by state law, which varies sharply. Legal trackers such as Traverse Legal's explainer on the shifting non-compete landscape and the broader survey of developments in non-compete law since 2023 both make the same point: the map, not a federal rule, decides.

The rough shape of the map matters for where you build. A handful of states, including California, Minnesota, North Dakota, and Oklahoma, do not enforce non-competes for most workers, which is one reason so many companies form there. In roughly two-thirds of states, non-competes remain enforceable against qualifying employees within reason. Even where a non-compete would not hold, an employer can still enforce non-disclosure and non-solicitation agreements and trade-secret protections, so a non-compete being unenforceable does not mean you are free to use confidential information or poach the team. Know which agreements you actually signed, and have counsel assess them against your state before you compete in the same market.

The fastest way to turn a clean spin-out into a lawsuit is to walk out with the employer's confidential information. Trade-secret protection survives almost everywhere regardless of non-compete status, so the rule is simple and absolute: take nothing. No customer lists, no internal documents, no data exports, no code. Your edge is what you understand, not what you copied. A vertical AI company's real defensibility comes from a data flywheel it builds itself over time, the kind described in the data moat in vertical AI, not from a dataset you smuggled out, which is both legally radioactive and a weak foundation anyway.

Clean hands also protect your fundraising. When a company built on domain expertise raises or takes on a co-founding partner, diligence will probe the IP story. A founder who can show they took nothing, built nothing on the employer's dime, and left on good terms has a clean answer. A founder with a murky story spends the next year explaining it. In a regulated sector the bar is higher still, and building compliantly from the start, as laid out in how a financial services operator builds a compliant vertical AI company, begins with a clean origin.

A spin-out founder tends to be senior, well paid, and used to institutional structure, which makes the cap table the part they are most tempted to overcomplicate and most need to keep simple. The goal is a clean founding cap table: clear founder ownership with standard vesting, no early over-dilution, and no ambiguous IP contributions that cloud who owns what. Because a spin-out already carries an IP question by nature, a messy or over-diluted table on top of it is exactly the combination that makes later investors nervous.

This is where the choice of building partner shows up directly in your ownership. Traditional venture studios often take a large stake, and the economics of that are laid out in venture studio equity explained, which puts typical studio ownership around a third of the company. gAI Ventures takes a different line: it co-founds vertical AI companies in financial services, enterprise productivity, and commerce, contributing capital and a production-grade founding team while the fund and operating company together hold a combined stake near 20 percent, a cleaner cap table than the roughly 40 percent many studios take. The full logic behind those theses sits in gAI's vertical AI investment theses, and the philosophy of co-founding rather than merely funding is set out in the gAI Ventures manifesto.

The reason a spin-out is hard is not usually the idea; it is that the operator has to solve the legal separation and the build at the same time, with no engineering team. A technical venture builder removes the second problem. Rather than spending six months recruiting a technical co-founder while your non-compete clock and your market window both tick, you get a founding engineering team from day zero and can spend your energy on the clean separation and the market. You can see the kind of companies this produces across the gAI Ventures portfolio, and the people who do the building on the gAI Ventures team. More on the model and the vertical AI space is on the gAI Ventures blog, and the company itself is at gAI Ventures.

The sequence that works is boring and effective: plan the separation while employed, leave clean, validate the market before committing real money, then build with a team already in place. Done in that order, a spin-out from a big institution becomes one of the strongest starting positions in vertical AI, because you bring the domain edge the model needs and none of the legal baggage that sinks a sloppy exit.