Trade Secrets and Non-Compete

Most of an AI company's core assets never appear on a patent list. Model weights, training data, prompts and tuning know-how, customer structures and pricing models are rarely patentable in time, and are not suited to protection by disclosure. For these assets there is one route — trade secrets, and the confidentiality architecture and personnel constraints built around them.

The assets that never reach a patent list

Most of an AI company’s core assets never appear on a patent list. Model weights, training data, prompts and tuning know-how, customer structures and pricing models are rarely patentable in time, and are not suited to protection by disclosure. A patent trades disclosure for protection; a trade secret trades confidentiality measures for protection. They cover different parts of the same asset, and the choice depends on how easily the technology can be reverse-engineered and whether disclosure would still leave a defensible position.

In practice the two routes run together: inference optimisation and training methods — the parts an outsider can observe — go to patents; weights, data recipes and tuning records go to trade secrets. The difficulty is not the choice. It is whether the confidentiality measures taken will support a claim later.

Classifying what is protected

Technical secrets and business secrets carry different evidentiary burdens. Algorithm structures, training-data processing methods and parameter-tuning records are technical secrets, and can usually be evidenced through code repositories, training logs and version history. Customer lists, pricing models and channel policies are business secrets, and require proof that the information is not publicly known, has commercial value, and has been subject to confidentiality measures. Classification is about allocating protection cost: not every piece of information deserves the full set of measures, but once trade-secret protection is chosen, the measures cannot be backfilled after a dispute begins.

Three layers of confidentiality

The first layer is contractual: confidentiality clauses in employment contracts, standalone NDAs, and undertakings covering contractors, advisers and interns. The second is access control: tiered access to code and data, account revocation on the day someone leaves, separation between production and training environments. The third is record-keeping: logs of downloads and outbound transfers of confidential documents, and activity records for key personnel.

One of the most common reasons a trade-secret claim fails in practice is that the rights holder cannot prove it took corresponding confidentiality measures. An agreement without access separation, or separation without records, leaves a gap at the point of proof.

The boundaries of non-compete

Non-compete obligations apply only to senior management, senior technical staff and others under a duty of confidentiality; the term may not exceed two years, and monthly compensation must be paid during that term. Three factors decide whether a clause is enforceable: an overly broad definition of covered personnel will be held invalid; a missing compensation standard, or compensation not actually paid, makes the clause difficult to assert; and geographic or business scope drafted too generally may be narrowed in interpretation.

For AI companies the real difficulty is defining a competing business. Whether a computer-vision company competes with an embodied-intelligence company should be settled in the agreement by describing the business, not by naming companies. The labour dispute judicial interpretation (II) in force since September 2025 addresses the application and compensation of non-compete clauses, and existing agreements are worth reviewing against that standard.

Handling the exit

Departure is the point of highest trade-secret risk. What needs to be completed is device and account recovery, recorded handover of code and data, an exit interview record, written notice of non-compete obligations, and confirmation of when compensation starts. The value of this step is not restricting movement; it is fixing, at that moment, the evidence that a later claim will depend on.

Choosing the enforcement route

A breach-of-contract claim for violating confidentiality or non-compete terms is relatively direct on evidence. A tort claim under the trade-secret provisions of the Anti-Unfair Competition Law (2025 revision) requires proof of secrecy, commercial value and confidentiality measures. Serious cases may engage Article 219 of the Criminal Law. The evidentiary requirements rise along these three routes, which is why the chain has to be built in day-to-day management rather than after a dispute arises.

Common points of failure

Claiming a trade secret without having taken confidentiality measures; agreeing non-compete without paying compensation; confidentiality coverage that does not match what the person actually accessed; key people exposed to secrets as advisers or contractors without being brought within the constraints; and non-compete clauses in equity incentive documents mixed with statutory non-compete, leaving both the conditions and the remedies unclear.

Related services: IP Protection, Employment Compliance, Equity Incentives and Option Pool, Regulatory Investigations and Criminal Risk.

Frequently Asked Questions

Do trade secrets need to be registered?

No registration is required, but you must show confidentiality measures and records to prove the information was kept secret, otherwise a claim is hard to sustain.

How much non-compete compensation is required?

There is no uniform statutory rate; practice commonly uses 30% to 50% of average monthly salary paid monthly, and amounts below local minimum standards may invalidate the clause.

How do we secure evidence when an employee leaves?

Recover accounts and devices, document the handover, conduct an exit interview and give written notice of non-compete duties, and preserve download and outbound logs.

Further Reading

Rule changes, enforcement cases and practical judgments around this service, listed in publication order.

Let's Talk About Your AI Company