Deal documents evolve to address AI defensibility and intellectual property risks

Investors and acquirors are pressing AI companies on a question that now shapes valuations and deal terms: what makes the business defensible? As foundation models become commoditized and publicly available tools let competitors replicate products quickly, the answer increasingly determines whether a company's advantage will last. Transaction documents are already adapting, with dedicated AI-specific representations and warranties moving from novelty to expectation.

What AI defensibility means

AI defensibility measures how well a company's AI capabilities, data assets, and models are insulated by legal protections and strategic moats. The concept mirrors traditional competitive advantage but applies it to a landscape where rapid technical advances can quickly erode a product's uniqueness. If a competitor can recreate your AI product using off-the-shelf models and public data, defensibility is weak.

Why it matters now

The dotcom era showed that participation in a technology shift is not itself a moat. Investors are asking whether a business model can be easily replicated, what barriers prevent another startup or big tech from producing the same thing, and whether a market correction is inevitable. "The value will undoubtedly accrue to companies that convert AI capability into durable competitive advantage," the authors write. For founders, legal advisors, and boards, the question is not whether the company uses AI, but whether that use creates a defensible position.

The pillars of AI defensibility

Companies and their advisors evaluate defensibility across several dimensions:

• Proprietary data and algorithms. The strongest positions come from data that is both exclusive and properly obtained. This includes trade-secret datasets, protected algorithms, and rigorous data-quality processes. Critical questions include whether the company has permission to use its data, whether data sources are fragmented or subject to non-compete restrictions, and the constraints imposed by open-source dependencies.
• Operational dependence and stickiness. When AI is embedded in customer workflows, switching costs rise. Feedback loops-where the algorithm trains on usage data to improve over time-make the AI more tailored to each customer, deepening the relationship and raising the cost of moving to a competitor.
• Scaling, distribution, and network effects. Speed-to-market advantages become durable when paired with exclusive distribution channels and established relationships. Network effects amplify this: as value grows with each participant, displacing the product requires rebuilding an entire network, not just building a better tool.
• Industry and regulatory expertise. Deep knowledge of regulated sectors-healthcare, finance, data privacy-creates meaningful barriers. Specialized focus signals sustainable positioning and makes it harder for generalist entrants to compete.

What the law actually protects

From an IP standpoint, AI defensibility is narrower than the broader business concept. Patents can protect novel, non-obvious improvements to AI methods, system architectures, or specific applications that solve a technical problem. But abstract ideas implemented on a generic computer using off-the-shelf models are not patent-eligible. Trade secrets are where the real moat resides. Companies can protect proprietary datasets, data labeling methods, model architectures, training pipelines, and optimization techniques. Under the Defend Trade Secrets Act of 2016, trade secrets require no registration and last indefinitely, but they vanish if exposed or reverse-engineered.

Copyright offers limited protection. It covers source code, documentation, and some original training data, but likely does not protect ideas, methods, functionality, or model outputs. Contracts and licensing, often overlooked, can create effective defensibility where IP law is weak: API terms of service, data usage agreements, and enterprise contracts all play a role. The uncomfortable reality is that most AI products are weakly defensible from a pure IP standpoint unless they rely on trade secrets or truly novel inventions. The strongest IP-based defensibility typically combines trade secrets, selective patents, and contracts. As the authors put it: "AI defensibility is not about the model, it's about what surrounds the model: data, users, workflows, and positioning."

How deal documents are catching up

Transaction documents are moving beyond general IP and technology representations to include standalone AI-specific provisions. New definitions-such as "AI Technology," "AI Inputs," and "Company AI"-anchor representations that address:

• Ownership and outputs. Sellers may be asked to represent exclusive ownership of or valid licenses to all Company AI and its outputs, and to confirm that AI usage has not harmed the ownership or enforceability of any proprietary IP.
• AI inputs and data governance. Representations may require that all AI Inputs are obtained and used lawfully, and that no copyrighted, proprietary, or confidential material has been fed to third-party AI providers without authorization.
• Regulatory compliance. Companies must represent compliance with applicable AI laws, risk management frameworks, bias mitigation protocols, transparency requirements, and algorithmic accountability standards. These obligations can arise from truth-in-advertising laws, the Fair Credit Reporting Act, FDA rules on AI-enabled medical devices, and more.
• Third-party AI services. Investors expect disclosure of all third-party foundation models, LLMs, model APIs, and fine-tuning services used in products, along with compliance with terms of use. A key representation is that no such usage grants the provider any ownership of the company's prompts, outputs, embeddings, or confidential information.
• Performance claims. All material written claims about AI capabilities must be true, accurate, and backed by reasonable substantiation, with known limitations, failure modes, and bias disclosed.

Why this matters for legal professionals

AI-specific diligence and representations are rapidly becoming standard. General IP and technology reps often fail to cover AI-specific risks, so a standalone AI section allows for targeted disclosure, appropriately scoped knowledge qualifiers, and clearer risk allocation. For deal teams, the message is clear: if your transaction involves an AI company, you need to examine ownership of training data, third-party model dependencies, and whether the company's AI usage jeopardizes its own IP. Staying current on these evolving standards is essential. Professionals looking to deepen their expertise can explore AI for Legal Professionals Courses or targeted resources like AI Intellectual Property Courses to navigate the intersection of AI, IP, and transactional practice.