AI Startups: Protect Everything EXCEPT the...

Short version: One dividing line decides the patent question for artificial intelligence (“AI”) companies. The same feature “done with AI” is not worth patenting—keep it a trade secret. A feature that was never before possible—the detectable, tangible result the customer buys—is the patent worth owning, and it carries more weight as software gets cheaper to copy.

The uncomfortable truth about “AI use-case” patents

Most AI patents in the wild are patents on a use case: “using AI for [X].” They share three defects:

Undetectable. You can’t see what’s happening inside a competitor’s graphics processing unit (“GPU”) cluster. If infringement can’t be detected from the outside, you can’t enforce it in practice.
Narrow by design. “Using AI for [use case]” leans on conventional tools. Examiners force the patent claims into tight corners that are easy to design around^[1]. What looks like “broad” coverage is, in truth, very narrow.
Disclosure that helps competitors. You publish a roadmap of how you did it. They learn from your specification without taking on your cost or risk.

Put bluntly: thin “AI use-case” patents backfire. You pay to teach your competitor while getting almost no protection.

The dividing line: a use case is not a feature

The patent decision for an AI company turns on one question: did the AI change what the customer gets, or only how you deliver it?

The same feature, done with AI: nothing worth patenting

If your product does what earlier products did—built faster, run cheaper, staffed leaner—the AI is an implementation detail. The customer buys the same result they could buy before. There is nothing worth patenting here: a patent on “doing X with AI” is the use-case patent above—undetectable, narrow, and a gift-wrapped roadmap for your competitor. The idea of “AI to solve X” is easy to imagine. The heavy lifting that made it work belongs in trade secrets, not in a published specification.

A feature that was never before possible: where a patent earns its keep

If the AI work produced a capability no product could offer before—a control the customer sees on the user interface (“UI”), a new call in your application programming interface (“API”), a behavior a user can trigger and watch happen—patent that feature. The patent is on the detectable, tangible result the customer pays for, not on the implementation underneath it.

The new feature passes the detectability test by definition: your customers have to see it, or they would never pay for it. A competitor who copies it infringes in plain view. And because the patent is on the new function, it covers a competitor who builds the same feature with a different model, a rules engine, or a room full of contractors. Patent “AI that does X,” and you have narrowed the coverage to one implementation while handing over the recipe.

This is the same discipline we apply to every software patent: put the patent claims on the interfaces of the product—the API, the user experience, the administrative screens—where infringement is visible from the outside. Detectability is a Key Factor for Patent Value describes the interface approach, and Method Claims and Undetectability shows why the form of the patent claims matters as much as the feature.

A few other cases sit on the same side of the line—externally observable, and blocking a capital-intensive alternative:

Hardware or sensor designs with testable outputs
Cryptographic watermarking or verifiable artifacts
Algorithms that produce a distinct, measurable, externally evident behavior
Systems where you can buy and reverse-engineer a shipped product

The question for the invention disclosure meeting is simple: does the customer get something that was never before possible, and can you see it in a competitor’s product? When both answers are yes, the AI underneath does not matter—patent the feature.

Where the moat lives on the other side of the line: data and operations

We never like the “big idea patents” that mistakenly try to be broad. The valuable positions are the roadblocks—the specific mechanisms a competitor cannot avoid if they want to deliver the same value. Ask what was the hardest technical thing you did. Where did the time, money, and frustration go? What solution unlocked your value proposition? Those answers are the “slide to unlock” of your product.

In an AI business, most of that hard work never crosses the dividing line—it is invisible from the outside. It is not the AI model; it’s everything around it:

Collection: How you source raw data and secure rights to use it.
Curation: Cleaning, de-duplication, labeling, harmonizing, schema design.
Process: Feature/embedding choices, evaluation harnesses, guardrails, feedback loops, post-processing heuristics.
Integration: How the model plugs into workflows, contracts, and service levels customers rely on.

As with any intellectual property (“IP”) protection, you always default to trade secret^[2]s first. Only get patents where the trade secrets cannot be protected.

Your methods for data collection, curation, processing, and integration are all easily protected by trade secrets. Should you get a patent on any of these? NO – because it is easy to design-around any of these processes, a patent would be undetectable, and you give away your hardest-earned information.

When the hard work does surface as a customer-visible feature that was never before possible, it crosses the line—patent it.

Trade secrets, not patents

If you can’t detect infringement, don’t patent it. Keep it secret.

What to keep as trade secrets

Data sets and labeling recipes
Curation tooling and pipelines
Evaluation/evasion tests, red-team suites, safety filters
Retrieval schemas, prompt orchestration, and post-processing
Performance heuristics and deployment runbooks

How to protect them

Contracts: Strong invention assignment, non-disclosure agreements (“NDAs”), confidentiality, and data-rights language with employees, vendors, and customers.
Access controls: Least-privilege repositories, role-based data access, audit logs, and secrets management; restrict exports.
Process discipline: Code reviews, change control for pipelines, and documented handling of sensitive data.
Vendor posture: Data processing agreements (“DPAs”), clear IP ownership, and exit/transition clauses.

Trade secrets are the only route that actually protects the advantage here because they don’t require detectability^[3]^[4].

Plan for competition—the economics changed

Assume others will build their own dataset, their own schema, and their own curation process. They can, and some will. The cost of building software is collapsing, which means a well-funded competitor can replicate your product faster than ever. Bake that reality into the business model.

“AI use-case” patents will not keep them away. Your protection is layered: trade secrets on the invisible work, execution on the business, and enforceable patents on the features that were never before possible. When a competitor can duplicate your product overnight, those feature patents are the barrier left standing.

Beyond the patents, win on things a rival can’t copy overnight:

Distribution: Channels, partnerships, and placement.
Contracts: Long-term agreements, integrations, switching costs.
Customer base: Adoption, support, renewals, references.
Performance: Speed, accuracy, coverage, reliability, compliance.
Brand and sales motion: Clear positioning, repeatable demos, proof that it works in production.

A thin “AI use-case” patent cannot carry that load. Build the moat^[5] where it holds.

Valuation: two very different perspectives

1. Valuation to investors: revenue is the only metric for value

For equity investors, valuation rests on revenue and profitability. Revenue is proof that customers choose your product over alternatives, and profitability is proof that you can deliver it efficiently. Nothing signals competitive advantage better than a customer willing to pay, and keep paying.

Investors look at the company. They want realized value—cash flow tied directly to product-market fit. That’s why your best story for investors isn’t a patent—it’s recurring revenue, renewals, and growth. In theory (but rarely in practice), the venture investor is supposed to price a company as a reflection of its future discounted cashflows.

See: The Company as the Product.

2. Valuation for lending: we only look at cashflow but sometimes data

When we look at a company for IP-backed lending, the lens shifts. A patent, standing alone, has almost no resale value—especially in AI, where the data and processes are tightly coupled with the business and rarely transferrable.

In this context, our lending underwriting is completely dependent on cashflow.

We cannot give credit for any “AI on X” patents, and we typically cannot give credit based on specialized trade secrets. In the “AI for X” businesses, your trade secrets are hard-fought and hard-won for you, but they often have very little resale value. Any acquirer will want the ‘package’ of the entire system, and you cannot separate out the IP from the company.

There is an exception: there are some businesses where the data can be resold on its own. There are emerging markets for data, especially carefully curated, cleaned, and updated data. This may be a separate line of business for the AI startup company, but it also may undermine the company’s products. In the resale market, data can be sold, but it is very difficult and you should not expect high valuations.

See: Patents that Protect a Business Advantage.

Action checklist for founders and boards

None of this is a founder side project. These items are part of what a Chief Intellectual Property Officer (“CIPO”) handles. The founder’s job is to make sure someone owns them.

Stop the “AI use-case” patents. Undetectable patent claims burn cash and teach your competitors.
Map the data. What you collect, where it comes from, what rights you hold, what gaps to close.
Protect the trade secrets. Employee IP assignment, NDAs, vendor data agreements, least-privilege access, and pipeline discipline.
Patent the features. A new customer-visible capability that no product had before is a patent worth getting—no matter how you built it.
Spend the savings on go-to-market. Distribution, integrations, renewals, expansion.
Pitch the real moat. Lead with data rights, process advantage, distribution, and revenue—not a thin patent.

Bottom line

The dividing line decides the patent question. The same feature “done with AI” gives you nothing worth patenting—keep the work secret and put your money into your data, your process, your distribution, and your customers. A feature that was never before possible—the tangible result the customer buys—is where a patent earns its keep, and as software gets cheaper to copy, it carries more of your defensibility, not less.

Patent the new. For everything else, focus on everything but the patent.