AI in the Boardroom, Shareholders in the Courtroom: When AI Strategy Becomes Fiduciary Risk

Introduction

As organizations invest heavily in large language model (LLM)-based artificial intelligence (AI), a new category of shareholder litigation is emerging: one that targets not the technology itself, but the business decisions behind its adoption. The scale of that adoption is hard to overstate: use of AI in professional services nearly doubled to 40% from 2025 to 2026,¹ and 97% of C-suite executives reported deploying AI agents within their company in the last year.²

Shareholders are now suing boards over how that adoption happened. Three derivative complaints filed over the past several months each alleged that officers and directors adopted an unlawful business plan to build AI products on datasets containing copyrighted works, and each pled that decision as a breach of fiduciary duty.³ This article examines what those complaints alleged, the underlying legal theories they drew on, how new and emerging AI technologies may create further exposure, and what companies should do now to mitigate it.

The Three Derivative AI Suits and What They Allege

The three complaints, filed from April to July of this year, all pled the same theory. Each alleged that the board approved an AI strategy built on training data the plaintiff said was pirated, and each treated that approval, rather than the underlying infringement, as the breach of fiduciary duty. Two distinguishing features are worth noting. First, they named directors and officers individually. Second, they pled the company’s own governance documents and public disclosures as the standard against which the board’s conduct was measured.

Adobe: Data Provenance Traced to an Impermissible Source

SEIU Pension Plan Master Trust filed the first of the three complaints on April 24, 2026, in the Northern District of California, derivatively on behalf of Adobe against its named directors. According to the complaint, Adobe executives breached their fiduciary duties and violated the Securities Exchange Act by adopting an AI strategy that used copyrighted works without authorization. The complaint pointed to a research paper published on Adobe’s website, which states that its SlimLM small language models are “pre-trained on SlimPajama-627B,” and that a third-party vendor built SlimPajama by copying, cleaning, and deduplicating another vendor’s RedPajama dataset. Following that chain further, RedPajama incorporated the Books3 dataset, which consisted of roughly 196,640 books taken from the Bibliotek shadow library and was removed in October 2023 because it was “defunct and no longer accessible due to reported copyright infringement.”⁴

Against this factual pattern, the complaint then pointed to Adobe’s disclosure in its 2025 proxy statement that Firefly was trained on “licensed and public domain content … [that] “doesn’t infringe on third-party intellectual property rights,” and also on Adobe’s marketing claim that it does “not mine content from the web to train Adobe Firefly.”⁵ To build the argument for damages, the plaintiff pointed out that Adobe stock fell from $354.66 to $264.67 across the filing of two copyright class actions directed at the company, and that Chief Executive Officer Shantanu Narayen resigned during this period. The plaintiff alleged that the directors’ conduct in approving the use of the alleged infringing data set in a hasty rush to adopt AI at the company exposed Adobe to settlement costs, legal fees, lost goodwill, reputational harm, and lost customers in those class actions.

Microsoft: Countering Passive-Investor Defense

Microsoft shareholder Eric Anderson filed the second complaint on June 30, 2026, in the Western District of Washington, derivatively on behalf of Microsoft against its named directors and officers, over the company’s investment in an AI developer that later faced copyright liability. Anderson recited that Microsoft invested more than $13 billion in OpenAI, designed and built a bespoke Azure supercomputing environment “in collaboration with and exclusively for OpenAI,” optimized it for training large language models on internet-scale datasets, and continued to operate it during both training and inference. On that record, the plaintiff pled that Microsoft was a co-developer and commercial partner in the creation, training, deployment, and monetization of the systems now facing copyright suits from the Authors Guild, individual authors, and nearly 400 newspaper publishers.⁶

Anderson also relied on Microsoft’s proxy statements for the 2024 and 2025 annual meetings when making claims against the board. Those statements told shareholders that “[t]he Board maintains direct oversight over AI strategy risk” and that “[f]or major initiatives, such as our approach to AI, the Board engages with management on strategic vision, investments, partnerships, capital requirements, and risks.”⁷ The complaint added that the directors recommended that shareholders vote against proposals seeking accountability for AI data sourcing, and that they continued to approve the training practices at issue after Judge Alsup’s June 23, 2025 order⁸ in Bartz v. Anthropic distinguished lawfully acquired books from pirated ones. Because a derivative claim seeks recovery for the company rather than for shareholders, the complaint pled two ways in which Microsoft itself lost money. First, Microsoft paid Chief Executive Officer Satya Nadella $224,115,510 in compensation from 2023 through 2025, which the complaint pled as unjust enrichment on the theory that the awards rested on an AI strategy the plaintiff alleged was unlawful. Second, Microsoft repurchased its own shares at prices the plaintiff alleged were inflated by the same disclosures, so the company overpaid for its own stock.⁹

Nvidia: The Theory Extends to Video, Voice, and Circumvention

Shareholder Jessica Berliner filed the third complaint on July 31, 2026, in the Northern District of Illinois, derivatively on behalf of Nvidia against its named directors and officers. Berliner’s complaint pled that Nvidia trained its NeMo Megatron language models on The Pile, a dataset containing the Books3 corpus, and separately downloaded the SlimPajama dataset. When publishers declined to license books at the speed and volume Nvidia sought, the complaint alleged the company turned to Anna’s Archive, was told those collections were illegally acquired and maintained, but nevertheless gave “the green light” within a week.¹⁰

The complaint made a separate set of allegations about Nvidia’s Cosmos video model. Berliner alleged that Nvidia employees used the yt-dlp downloader together with virtual machines that rotated IP addresses to extract YouTube videos, and that because YouTube provides no lawful download mechanism, each clip-level retrieval was a separate circumvention under Section 1201(a) of the Digital Millennium Copyright Act.¹¹ To establish knowledge, the complaint relied on internal Slack messages. When an employee asked whether the company had legal approval for one video dataset, a research leader replied that “[t]his is an executive decision. We have umbrella approval of all the data.”¹²

Berliner also alleged that Nvidia’s voice synthesis and voice cloning models ingested hundreds of thousands of hours of speech recordings and extracted source speakers’ voiceprints without notice or written release, in violation of the Illinois Biometric Information Privacy Act, and that Nvidia published the resulting model weights on Hugging Face for download worldwide. The damages theory was similar to the Microsoft complaint. Nvidia paid President and Chief Executive Officer Jensen Huang $36.3 million for 2025, $49.9 million for 2024, and $34.2 million for 2023, and separately spent $13.258 billion repurchasing its own stock at prices the plaintiff alleged were inflated by the same disclosures.¹³

Data Provenance Is the Common Element

All three complaints worked backward from a released AI product to a dataset tied to an allegedly unlicensed source, then attempted to establish scienter with a public statement that contradicted the company’s actions. None of the causes of action asserted in the three complaints required the plaintiff to prove that the underlying copyright infringement actually occurred. Because these are derivative fiduciary duty claims, the question is whether the boards knew of the provenance risk and consciously failed to address it. Each complaint used the company’s own proxy disclosures, board reports, and strategy approvals to plead that knowledge, then pointed to the gap between public statements and actual data-sourcing practices to allege bad faith. Adobe’s Code of Business Conduct, for instance, required personnel to confirm the company held appropriate rights before copying copyrighted materials, and Nvidia’s Code instructed personnel not to use company resources to acquire, share, or store copyrighted material that is not properly licensed.

Other Important AI Risks for Boards

These complaints arose from decisions about training data, but the risks could go further. Three technology trends already embedded in enterprise AI products are creating new categories of exposure that the next wave of complaints is likely to target: agentic AI systems that act on company data without human review, retrieval-augmented generation that copies outside material each time a query runs, and token consumption that grows without contractual guardrails to control it.

Agentic AI Systems

Agentic AI systems are LLM-enabled systems that “can take goal-directed actions with minimal human oversight, make contextual decisions, and dynamically adjust plans based on changing conditions.”¹⁴ Analysts predict that within two years, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from zero in 2024.¹⁵

Agentic AI differs from generative AI in four ways that bear on where legal exposure may lie. First, it takes actions and makes decisions rather than generating content, which raises the question of which business decisions have been delegated to an agent, what limits apply to that authority, and where an unusual case gets escalated to a person. Second, according to the National Institute of Standards and Technology, agents are discrete actors within an organization’s IT environment, each with its own permissions and credentials.¹⁶ Because an agent holds credentials the way an employee does, the questions mirror those for personnel: how access is granted, how often it is reviewed, and how promptly it is withdrawn when a role changes or ends. Third, agents coordinate with one another over machine-to-machine protocols, sharing context and delegating subtasks without human intervention; where those exchanges are not logged, a chain of autonomous decisions may be difficult to reconstruct in response to a books-and-records demand.¹⁷ Fourth, agents can invoke external tools to act outside the company’s systems, and every autonomous action taken through a third-party integration is an action the company may be held to have authorized.¹⁸

Unlike traditional cybersecurity threats, which involve discrete breaches by identifiable actors, agentic systems that are not properly deployed and monitored create a risk of unpredictable data leaks that may form the basis of a future derivative complaint. For instance,, agents can overshare full documents through “context dumps,” and “fuzzy retrieval” can surface sensitive material that was never part of the agent’s task. Those spills carry intellectual property consequences: an uncontrolled disclosure can undermine trade secret protection by showing a failure to maintain secrecy, and disclosure of proprietary technical information through agent communications can jeopardize patent rights if it precedes filing.

Web Scraping and Retrieval-Augmented Generation

Retrieval-augmented generation (RAG), also called grounding, systems pull material from a designated source each time a query runs, whether a proprietary repository, a curated document set, or content scraped from the open web, and adds that material to the prompt, which creates the potential for copyright infringement or privacy-related claims.¹⁹

Courts have not settled how fair use applies to the copying involved in building and operating these systems, and no appellate court has ruled on this point. The decisions to date address training data rather than retrieval, and how fair use applies to the copies a RAG system makes during query remains an open question.²⁰

Tokens and Context

A “token” is the basic unit of text an AI model processes, roughly a word fragment, and “context” is the full set of tokens supplied for a given task, including the prompt, any retrieved material, and prior exchanges. Agentic and RAG systems consume far more of both than traditional generative AI tools, because a model given autonomy spends tokens not only answering a question but planning, retrying, summarizing, calling tools, and checking its own output.

That consumption has a direct cost. Most enterprise AI vendor contracts price access by token volume, and without a consumption cap or audit right, the company is writing a blank check it cannot verify. If usage outruns the contracted allotment, the company faces a binary choice: pay whatever the vendor demands or lose access to a system the business already depends on. In light of this, companies should be scrutinizing AI vendor contracts closely and should confirm that their contracts and budgets can absorb the consumption growth that comes with deploying more agents and more retrieval sources. The Adobe complaint shows the shape such a claim can take, because it pled corporate waste as its second count, alleging that the directors caused Adobe to incur substantial expenditures as a result of an AI strategy they approved.²¹

Recommendations for Businesses Adopting AI

The risks described above—from agentic autonomy to undiligenced retrieval sources to uncontrolled token spend—are real, but they are also manageable. Each recommendation below maps to a specific allegation or governance gap identified in the derivative complaints. The goal is not to eliminate risk, but to build the documented record of reasonable oversight of AI adoption, which would help answer claims future shareholder derivative plaintiffs may make.

Recommendations:

1. Reconcile public statements against actual practice. The derivative complaints discussed here used the companies’ own governance documents and public statements as the measuring stick. Adobe’s Code of Business Conduct required personnel to confirm the company held appropriate rights before copying copyrighted materials. Nvidia’s Code instructed personnel not to use company resources to acquire or store improperly licensed copyrighted material. And the Adobe complaint pled the company’s marketing claim that it does “not mine content from the web to train Adobe Firefly” as the basis for counts under Section 14(a) of the Securities Exchange Act of 1934 and Rule 10b-5, turning a product-page statement into a securities claim. In each case, the plaintiffs did not need to construct a standard of care; the company supplied one and then allegedly failed to meet it. Companies should consider every public-facing AI commitment, including codes of conduct, audit committee charters, proxy statements, earnings calls, press releases, and marketing materials, and weigh them against what the company is actually doing, including with respect to its agentic and RAG agents, as a commitment the company cannot verify is a commitment a plaintiff may use against it. All things being equal, fewer, enforceable commitments are better than aspirational ones that create a gap between promise and practice.²²
2. Limit agent access and report the strategy to the board. Each AI agent should be limited to the data necessary for its assigned task, with restricted categories (such as trade secrets, privileged communications, and personally identifiable information) segregated so that such restricted information never enters general agent storage. But technical controls alone do not satisfy the board’s oversight obligation. A board should also implement a reporting system that surfaces mission-critical agent-created risks at the board level. Further, AI governance should be a recurring item on the board calendar, with documented reports in the minutes, not only because it is best practice, but because the absence of that documentation is a pleaded allegation in every one of the above mentioned derivative complaints.
3. Treat token spend as a contracting issue, and escalate it when the numbers warrant. Most enterprise AI vendor contracts price access by token volume, and agentic and RAG systems consume far more tokens than traditional generative AI tools because they plan, retry, summarize, and call external tools as part of every task. Companies should consider whether their AI vendor contracts include consumption caps with defined overage rates, audit rights that let the company verify actual usage, and data portability provisions that prevent lock-in if the relationship needs to end. Whether consumption itself belongs in board reporting may depend on the size of the commitment relative to the company. Still, where AI spend has become material, companies would be wise to periodically review these relationships by pressure-testing consumption caps, overage rates, and portability provisions against actual usage patterns and current market terms—and having a written policy in place detailing this review process helps build a favorable record.
4. Diligence training data and retrieval sources separately—and by name. A training-data license does not cover the separate reproductions a RAG system makes each time a user query triggers retrieval from an outside source. Diligencing one does not resolve the other. Companies deploying RAG-based tools should map every external knowledge source the system retrieves and verify whether that knowledge source has been appropriately licensed.²³

Conclusion

Companies find themselves caught between investor pressure to move quickly on AI and shareholder litigation over how companies do so. The record a board builds today is the record it will be judged on, and “we moved fast” is not a viable defense. Questions remain, including whether these complaints survive motions to dismiss, but there are steps a board can take now to build its company’s record of AI governance and reduce its exposure.