Articles

AI in the Deal Room: Why Intellectual Property Diligence Matters More in Acquisitions of AI-Enabled Products

Date: July 1, 2026

Key takeaways

  • AI does not make intellectual property less important in acquisitions. It usually makes the ownership, protectability, and risk questions more specific.
  • Patent diligence should not stop at a patent-count review. Buyers should ask whether the claims map to the product, whether human inventorship is documented, whether AI-assisted work created inventorship issues, and whether the patent position supports the deal thesis.
  • Copyright diligence should separate human-authored software and content from AI-generated or AI-assisted outputs, then test whether the target has protectable rights, adequate assignments, and training-data provenance.
  • Trade secrets often carry much of the value in AI-related deals, but only if the target can show reasonable secrecy measures, access controls, provenance records, and contractor/employee restrictions.
  • Purchase agreements should translate these diligence findings into representations, covenants, indemnities, exclusions, schedules, holdbacks, or closing conditions rather than treating AI IP as a generic technology-risk item.
     

Introduction

Acquirers are increasingly looking at businesses whose products use AI, depend on AI, or sit close enough to AI that the buyer expects future value from data, automation, software or model-enabled workflows. The target may not be an “AI company” in the headline sense. It may be a medical-device company using machine-learning outputs in a diagnostic workflow, a software platform embedding generative AI in a user interface, a manufacturer using computer vision in quality control, a services business with proprietary datasets and automation, or a consumer brand relying on AI-created marketing assets.

In each case, the acquisition thesis often assumes that some combination of software, data, patents, brand, confidential know-how, workflows, or customer-specific learning will transfer cleanly to the buyer. That assumption deserves pressure testing. AI can make a product faster to build and easier to demonstrate, but it can also obscure who created what, what is protectable, what was licensed, what was copied in and what rights actually move in the deal.

Intellectual property diligence in an AI-related acquisition should therefore be more than a schedule of registered assets. The practical question is whether the target owns and can transfer the technology position that supports the valuation.


Patent diligence: who invented the claimed technology, and what do the claims actually cover?

Patents can matter significantly in AI-enabled acquisitions, particularly where the target’s value depends on a technical workflow, device, platform architecture, model-training improvement, inference pipeline, data-processing technique, human-machine interface, semiconductor or sensor integration, robotics system, or regulated product. But a buyer should resist two shortcuts.

The first shortcut is assuming that any “AI patent” is valuable because it uses AI language. Patent value depends on claim scope, specification support, product mapping, competitor relevance, remaining term, family strategy, prosecution history, enforceability, challenge posture, and the availability of continuation practice. In an acquisition, the relevant question is not “does the target have AI patents?” It is “do the claims protect the technical advantage we are buying?”

The second shortcut is ignoring AI-assisted inventorship records. U.S. law remains anchored in human inventorship. In Thaler v. Vidal, the Federal Circuit held that an AI system could not be listed as an inventor under the Patent Act. The USPTO has also made clear that AI systems are tools, not inventors, and that the use of AI in the inventive process does not create a separate inventorship standard. The USPTO’s February 2024 guidance stated that AI-assisted inventions are not categorically unpatentable if a human made a significant contribution. The USPTO’s revised November 2025 guidance rescinded the prior guidance but reaffirmed the core point: only natural persons may be named as inventors, and the ordinary inventorship standard applies regardless of AI use.

For deal teams, that means the diligence file should include evidence of human contribution. Who conceived of the claimed subject matter? What did the AI tool contribute? Were prompts, lab notebooks, engineering tickets, design reviews, model outputs, prototype tests, and decision records preserved? Did employees and contractors assign their rights? Are all named inventors current or former employees, consultants, university collaborators, founders or third parties? Does the target have written assignments from each inventor, and are assignments recorded where appropriate?

Patent subject-matter eligibility is another diligence issue for AI and software assets. The USPTO’s 2024 guidance update on patent subject-matter eligibility, including AI inventions, does not eliminate the need for careful claim analysis. A buyer should ask whether the claims are directed to a practical technical application or merely recite results, data analysis, business logic, or generic automation. In a transaction, weak eligibility posture may affect value, enforcement leverage, negotiation strategy, and whether the buyer should require continued prosecution, continuation filings or narrower representations.


Copyright diligence: what was authored, what was generated, and what was used to train or build it?

AI-related acquisitions also raise copyright questions. Software code, user-interface text, documentation, training materials, datasets, marketing content, images, videos, reports, product designs, and customer-facing outputs may all be part of the value story.

The first issue is authorship. The U.S. Copyright Office’s 2023 registration guidance for works containing AI-generated material requires applicants to disclose and exclude non-human-generated material where appropriate. Its 2025 report on copyrightability addresses the protection of outputs created using generative AI. The practical diligence question is not just whether an asset looks polished or important. It is whether protectable human authorship exists and whether the target owns it.

That distinction matters in several common deal settings. A target may have used AI tools to create marketing images, product mockups, training content, software snippets, test data, customer reports, or design materials. Some of those outputs may be commercially useful but thinly protected or unprotected by copyright. Some may be protectable because humans selected, arranged, modified, or authored expressive elements. Some may be risky because a contractor used AI tools under terms that restrict commercial use or because the target lacks records showing who did the human creative work.

The second issue is input-side risk. The Copyright Office’s AI initiative and Part 3 report on generative AI training reflect the central unresolved question: when does using copyrighted material to train or build an AI system require consent, compensation, or another legal basis? Litigation is developing. Thomson Reuters v. ROSS Intelligence, a 2025 District of Delaware decision involving use of Westlaw headnotes in connection with an AI legal-research tool, shows that training-data and fair-use issues can move from policy debate to concrete deal risk.

A buyer does not need to resolve every open copyright question before closing. But it should know what it is buying. Diligence should ask for a data-source inventory, licenses, vendor terms, model documentation, training-data provenance, customer-data permissions, open-source reviews, takedown or demand letters, pending disputes, and records of AI-tool use in product and content development. Where the target cannot support its rights story, the buyer may need more tailored representations, indemnities, exclusions, escrow, price adjustment or post-closing remediation.


Trade secrets and data: much of the value may be invisible

In many AI-related deals, the most valuable assets are not registered rights. They are confidential datasets, feature-engineering choices, model weights, prompts, workflows, evaluation sets, annotations, deployment methods, customer feedback loops, security architecture, integration know-how, and the practical experience of making the system work in a specific market.

Those assets may be protectable as trade secrets, but only if the target can show that the information derives economic value from not being generally known and that it used reasonable measures to keep it secret. The federal Defend Trade Secrets Act defines trade secrets and misappropriation around those concepts. In deal diligence, the question is concrete: what did the company do to keep the important information confidential?

Good answers usually include more than a form NDA. They include role-based access, source-control permissions, vendor restrictions, employee and contractor confidentiality agreements, offboarding procedures, logging, clean data-room practices, policies governing AI-tool uploads, customer-data controls, security measures, and a map of who had access to sensitive model, data, and product materials.

AI can stress those controls. Employees may paste confidential code, designs, customer data or unreleased product information into third-party AI tools. Vendors may train on submitted data unless the customer selected the right enterprise configuration. Contractors may use public generative tools to produce code or creative assets without leaving a clean provenance trail. A target may claim a proprietary dataset but lack documentation showing where it came from, what rights attach to it, or whether it includes third-party confidential information.

A buyer should treat trade-secret diligence as an operational inquiry, not just a legal document review. The best question is often: “Show us the evidence that this secret stayed secret.”


Assignments, contractors, universities, customers, and vendors: title still matters

AI-related products often emerge from a messy collaboration stack. Founders experiment with public tools. Contractors build the first version. Employees improve it. Customers provide implementation data. Universities or accelerators contribute research. Cloud vendors supply model services. Open-source packages and pre-trained models sit underneath the product. Marketing agencies create launch materials. Investors receive technical decks.

That history can produce real title issues. Patents require written assignments from inventors. Copyright ownership depends on authorship, employment status, work-made-for-hire rules, and written transfers. Trade-secret ownership depends on confidentiality and control. Data rights depend on contracts, privacy terms, customer permissions and vendor terms. Open-source and model licenses can impose use, attribution, disclosure or field-of-use constraints.

For acquisitions, the diligence team should test the chain of ownership around the assets that matter most to the deal thesis. If the buyer is acquiring a model-enabled medical product, it should know who owns the training data, validation materials, code, documentation, and inventions. If the buyer is acquiring an AI software platform, it should understand open-source dependencies, model-service terms, plugin rights, customer-data restrictions, and whether any contractor retained rights. If the buyer is acquiring a consumer or media business using AI-generated creative assets, it should know whether the brand, creative library, and product materials are actually protectable and transferable.

The transaction documents should reflect the answers. Broad representations that the company owns all IP may be insufficient if the risk is concentrated in a dataset, model pipeline, or contractor-built module. Targeted schedules and exceptions are better than vague comfort.


How IP diligence should change deal terms

The point of AI-related IP diligence is not to create a longer checklist. It is to improve deal decisions.

For buyers, diligence can affect valuation, structure, reps and warranties, indemnities, escrow, holdbacks, closing conditions, integration planning and post-closing remediation. A buyer may decide that a target’s patent portfolio supports a premium because the claims map to a hard-to-replicate technical feature and the inventorship/title records are clean. Or it may decide that the real value is not patents but a confidential dataset and customer-specific deployment know-how, making trade-secret controls and customer contracts central to the deal. Or it may discover that key outputs are not strongly protectable, requiring a different valuation story and tighter covenants around future development.

For sellers, the lesson is to build the diligence file before the buyer asks. That means preserving invention records, recording assignments, cleaning up contractor agreements, documenting AI-tool policies, maintaining data provenance, tracking open-source and model licenses, cataloging trade secrets, and preparing product-to-IP maps. A seller that can explain what it owns, why it matters, and how it has been protected will usually tell a stronger value story.


Practical diligence questions

Deal teams should consider asking the following questions early:

Patents and inventions
  • Which patent claims map to the products, workflows, datasets, models, or technical features driving the acquisition thesis?
  • Were AI tools used in conception, experimentation, drafting, testing, or reduction to practice?
  • Who made the human inventive contributions, and what records show that contribution?
  • Are assignments signed and recorded, and do any universities, prior employers, government grants, contractors, or collaborators have rights?
  • Is the patent family positioned for continuation filings or claim refinement as the product changes?
     
Copyright and software
  • Which software, content, documentation, images, reports, and product outputs are human-authored, AI-assisted, or AI-generated?
  • Are contractor and employee assignments sufficient?
  • What open-source packages, pre-trained models, APIs, and model licenses are used?
  • Do any AI-generated materials have limited protectability, usage restrictions, or attribution obligations?
     
Data and training materials
  • What data was used to train, fine-tune, evaluate, or operate the system?
  • Was the data licensed, scraped, customer-provided, employee-created, publicly available, synthetic, or purchased?
  • What contracts, consents, privacy notices, or terms authorize the target’s use?
  • Has the company received any notices, demands, opt-out requests, or claims relating to data use?
     
Trade secrets and operational controls
  • What does the company treat as its core trade secrets?
  • Who has access to those materials?
  • What technical, contractual, and procedural controls protect them?
  • Are employees or contractors allowed to upload sensitive materials into third-party AI tools?
  • What records show that confidential information was kept confidential?
     
Deal documentation
  • Do the representations address AI-generated assets, training data, model licenses, contractor rights, customer data, open source and trade-secret controls specifically?
  • Should known AI/data risks be handled through special indemnities, exclusions, holdbacks, covenants, or closing deliverables?
  • Does the integration plan preserve the assets that justified the acquisition?
     

Conclusion

AI does not replace traditional IP diligence. It makes the traditional questions more important and adds new ones. In an AI-related acquisition, deal value may depend on a patent claim, a training dataset, a model workflow, a contractor-built software module, a customer-data right, a trade-secret control, or a creative asset whose copyright status is not obvious.

The buyer’s task is to connect the IP record to the deal thesis. The seller’s task is to make that connection easy to verify. The lawyer’s task is to translate the answers into practical diligence, valuation, risk allocation, and post-closing action.

For companies building with AI, the best acquisition preparation starts long before a letter of intent. Keep the invention records. Clean up assignments. Document data provenance. Control trade secrets. Track AI-tool use. Map product value to protectable rights. That discipline does not just reduce legal risk; it can make the company easier to buy.
 

ABOUT WHITEFORD | SchellIP

Whiteford and SchellIP help companies and acquirers evaluate intellectual property in the context of a transaction—assessing whether patents, copyrights, trade secrets, and data rights actually support the deal thesis, and translating those findings into practical valuation, risk allocation, and closing terms. The goal is a clear understanding of what a company owns, what transfers, and how those rights affect leverage, value, and business outcomes.

Selected authorities and sources