From idea to a defensible AI patent: a practical path

Start from the method, not the product

“The moat moved” explained why patents matter more now that AI agents make software cheap to build. This note is the how: what I have learned from six filings, three inventions and three issued patents about turning an idea into claims that hold.

The first lesson is that a product is not an invention. “An AI search engine” cannot be patented; a specific way of computing the result can. My own optimization patent is not “personalized search”. It is a method that represents a user’s mandatory constraints and partially ordered preferences as a directed acyclic graph, assigns weights with a separation guarantee, so that no combination of lower preferences can outweigh a single higher one, and returns results with a mathematical guarantee of optimality, incrementally as new candidates arrive. That sentence is the invention. Everything around it, the interface, the data sources, the model that reads the user’s preferences, is product.

So the first artifact is a one-page invention disclosure, written by the inventor before anyone talks to a lawyer: the problem, how it is solved today, the new step, why a skilled engineer would not have arrived at it by default, the technical effect it produces, and enough detail that the same engineer could build it. If nothing survives that page, there is nothing to file, and that is a cheap thing to learn early.

Novel, non-obvious, and enabled

Three tests decide whether a method can be claimed, and the inventor can run the first version of all three before spending a euro on counsel.

Novelty: nobody has published it. Search Google Patents and the research literature yourself, with the vocabulary a stranger would use, before your attorney does the formal search. Non-obviousness: the step is not a routine combination of known things; the strongest evidence is a technical result the known approaches could not produce, such as a guarantee of optimality where heuristics had none. Enablement: the description teaches how to build it. For software inventions there is a fourth question, whether the claims are tied to a concrete technical improvement rather than an abstract idea, and that is where counsel earns the fee: the same method can be claimed in a way that survives examination or in a way that does not.

File or keep secret

Not everything defensible should be patented. A patent is public and time-limited, but it stops even someone who re-invents your method independently. A trade secret lasts as long as it stays secret, and protects nothing once it does not.

The rule I use is about visibility. If the method shows in the product’s behaviour, a ranking that users can see, a verification anyone can test, a guarantee you advertise, then a competitor can observe it, rebuild it, and you have no recourse without a patent. If it cannot be observed from outside, a training recipe, a data pipeline, an internal heuristic, a filing teaches the world more than the claims would protect, and a secret is the better instrument. Most products need both, on different parts, and the disclosure page is where you decide which is which.

Work with counsel like an engineer

I am not a patent attorney: I am an inventor of methods and systems in computer science, engineering and artificial intelligence — spanning blockchain and mathematical optimization — and I work with patent counsel on the filings.

The division of labour is clear once you have been through it. Counsel drafts the claims, chooses the jurisdictions and the route, and argues with the examiner. The inventor supplies the disclosure, the drawings and flowcharts, worked examples, the argument for why the step is not obvious, and every technical answer during examination, when the examiner cites prior art and asks what is different. Those answers are engineering, not law, and they are the part nobody can do for you. The timing rule from the first note still applies: in Europe a public disclosure before filing can destroy novelty, so the demo, the deck and the paper come after the priority date, never before.

Build so the claims hold

A patent on a method nobody built is a certificate. A patent on the method your product actually runs is a moat.

That means building the product on the claimed method rather than on something that resembles it, and being able to show it. In practice: keep the claimed method as a recognizable component in the code, with a boundary and a name, so that “the product practises the claims” is a fact an engineer can demonstrate rather than an opinion; keep the design documents and the version history; and write the research paper about the industrial application after the filing, with the benchmarks, because it is both the reputation and the evidence. My optimization method runs as the engine behind PerfectFinder.ai. The chatbot patent I co-invented at App2Check is the method that built EVA’s knowledge base from FAQs and documents. In both cases the claims describe what the running system does.

The path, in order

Seven steps. The order is most of the value.

1. Write the disclosure.One page: problem, current approaches, the new step, why it is not obvious, the technical effect, how to build it.
2. Search the prior art yourself.Google Patents and the literature, in a stranger’s vocabulary. Finding the killer reference on day one is cheap; finding it during examination is not.
3. Decide patent versus secret, per part.Visible behaviour gets filed; invisible machinery stays secret. Write the decision down.
4. File before you show anything.Priority date first; then the demo, the deck and the paper. Counsel picks the route and the countries.
5. Build on the claims.A named component that practises the method, with design documents and history to prove it.
6. Publish after the priority date.The paper with real benchmarks, at a research venue and a business one. It cannot hurt the patent now, and it builds the brand.
7. Budget for examination.Office actions will come, and the technical answers are the inventor’s job. Plan the time as you would plan a release.

The receipts

The public records of the filings above.