IP Protection and Layout
For an AI company the core asset is intellectual property — and protection has to start on day one of development. Patents, software copyright and open-source compliance work as one system, not three separate filings.
Filing is not the same as protecting
Four questions come up in every financing diligence, and none of them is answered by the fact that a registration certificate exists: can the core algorithm be patented at all; is there copyright exposure in the training data; does the GPL component in the pipeline reach the product; and who owns the code an engineer wrote before leaving?
IP protection for an AI company is not a filing exercise. It is a design exercise that runs across three layers at once — the technical architecture, the commercial model, and the legal instruments that hold them together.
What we do
- Patent portfolio: patentability analysis, claim drafting, portfolio management, PCT and national-phase filing
- Software copyright: registration of core codebases and each material release, building a dated chain of ownership evidence
- Trademark: domestic filing, Madrid international registration, defensive coverage in adjacent classes
- Open-source review: component inventory, licence risk grading (GPL / AGPL / LGPL / MIT / Apache), remediation plan
- Trade secrets: confidentiality architecture, non-compete where appropriate, exit knowledge management
What is different for an AI company
- Patent or trade secret is a choice, not a default. A patent must disclose enough to practise, and once model weights and training detail are public, enforcement is hard; a trade secret stays closed but is difficult to prove against reverse engineering and staff movement. Route by “can it be reverse-engineered” and “will it be licensed out”.
- Employee invention and university-derived IP must be separated first. A model built at a previous employer or in a lab may not belong to the founder. Financing and filing both ask; a new agreement fixes the future, not a transfer that already happened.
- Open-source licences are tiered by how you use them. AGPL and SSPL add obligations when a service is offered over a network; Apache 2.0 carries a patent grant and termination clause; MIT requires only notice retention. A component schedule and licence matrix is cheaper than replacement later.
- Training data does not create ownership. Data itself is hard to title; what can be protected is the database compilation, annotation output and model parameters. Set boundaries with licences and confidentiality in deals rather than claiming “the data is ours”.
- Overseas filings follow market priority. The US scrutinises abstract algorithms, Europe looks to technical effect, and PCT and Paris Convention deadlines decide whether a gap can be cured. File where you sell first, not after the product spreads.
Typical situations
Situation 1: The core algorithm was never patented, and the valuation reflects it
An inference-optimisation method is treated by the team as “an implementation detail” rather than an invention, so nothing is filed. In Series A diligence the investor finds the core method unprotected and directly replicable — no technical barrier, no negotiating position. The valuation is discounted accordingly.
Our approach: identify patentable subject matter layer by layer from the architecture outward — training methods, inference optimisation, data pipelines, application-layer interactions — and build a portfolio whose purpose is coverage of the competitive surface, not a certificate count.
Situation 2: An AGPL component in the training stack reaches the whole product
A component under AGPL-3.0 is used in the training framework. AGPL extends copyleft to software offered over a network: if the product is a service built on a derivative work, the corresponding source must be made available. Diligence finds it, and the investor requires either replacement or disclosure.
Our approach: run open-source review before launch, not before financing. Every component is registered with its licence, graded by obligation strength, and AGPL-licensed dependencies are replaced or architecturally isolated so that the obligation does not attach to proprietary code.
Situation 3: An engineer leaves with part of the training code
A core developer resigns and joins a competitor, taking part of the pipeline. There is no NDA beyond boilerplate, no non-compete, and no copyright registration on the code. Enforcement is practically impossible.
Our approach: three instruments used together — an explicit IP assignment in the employment contract, copyright registration of core code to fix ownership in time, and a non-compete with defined scope and liquidated damages for designated positions. Any one alone is weak; together they are enforceable.
Process and deliverables
| Stage | Work | Deliverables |
|---|---|---|
| 1 | Architecture review, IP asset inventory, protection strategy | IP asset register + strategy |
| 2 | Patent drafting, copyright registration, trademark filing | Application set + registration certificates |
| 3 | Open-source review, licence grading, remediation | Compliance report + risk-graded component list |
| 4 | Trade secret system, confidentiality documents | Policy set + NDA and non-compete templates |
| Ongoing | Portfolio management, annuity reminders, enforcement support | Deadline alerts + enforcement support |
AI-specific deliverables
- Algorithm claim mapping table: claims drafted against concrete technical applications and checked line by line against patent-eligibility guidance.
- Training-data and open-source ownership checklist: provenance, licence type and copyleft scope recorded as the basis for diligence and enforcement.
Frequently asked
Can an AI algorithm be patented? Yes. The Patent Examination Guidelines as revised in 2024 set out the standard for AI-related applications: the claim must solve a technical problem by technical means and produce a technical effect. A bare algorithm fails; an algorithm bound to a concrete technical application can pass. Claim drafting is where that distinction is won or lost.
How much does open-source compliance matter? It is a standard diligence question, and the answer can reach the business model. A single AGPL component inside the core product can require source disclosure of the derivative work. Review belongs before launch, because remediation after the fact means replacing components under commercial pressure.
What does copyright registration actually give us? It is the most direct evidence of ownership of code, it is cheap, and it is fast. Diligence asks for it. Registering each material release builds a timeline that a dispute or a diligence can rely on.
When should the trademark be filed? At naming, not at launch. Registration takes roughly nine to twelve months, and a name that gains recognition before filing is a name someone else may already have claimed. For products sold overseas, file in the target markets before entering them.
Legal basis
- Patent Law (2020 amendment): invention, utility model and design patents; term and scope
- Patent Examination Guidelines (2024 revision): examination standards for AI-related applications
- Regulations on Computer Software Protection: copyright in software, ownership and term
- Trademark Law (2019 amendment): registration, opposition, renewal
- Anti-Unfair Competition Law Art. 9: trade secrets
- Open-source licences: GPL-3.0, AGPL-3.0, LGPL-3.0, MIT, Apache-2.0, BSD — obligation strength differs materially and must be graded per component