November 17. 2026

The Risks and Disputes that May Arise in Agentic AI

Share

auf einen Blick

This Legal Update is designed to help readers anticipate risks and disputes to address in contracting for agentic AI. Disputes will likely focus on performance failures, injuries to third parties by agentic AI actions, disputes over IP rights, data privacy and cybersecurity issues, and violations of regulatory obligations.

The Rise of Agentic AI and Risks It Creates

Agentic AI models are becoming increasingly available and are anticipated to become widely used by corporate customers. Because agentic AI possesses planning, reasoning, execution, and write-access capabilities, the risks shift from merely “bad text output” (from generative AI) to “destructive real-world actions” (from agentic AI). And because of the speed of execution and potentially iterative nature of AI agents, the impact of those real-world actions can be magnified quickly. Agents providing customer support might authorize returns or offer concessions that exceed their established authority (or guardrails). AI agents directed to conduct market research might engage in hacking of competitors or customers. AI agents operating across competitors might learn to collude and fix prices or output. AI agents might be susceptible to prompt injection attacks that lead them to disclose corporate information to attackers. AI agents stuck in recursive loops might drive up cloud computing or token costs in a matter of hours. AI agents used to generate debugging scripts may inadvertently introduce new errors into an IT system. And AI agents engaged in outreach campaigns may be accused of commercial slander of competitors.

Even with the contract provisions described in other Legal Updates, we can expect an increase in the number of disputes between companies and their agentic AI service providers. Some of those disputes will be similar to the disputes that arose in the context of outsourcing and SaaS agreements, while other disputes are likely to be new and unique to the risks posed by agentic AI.

Breach of Contract Claims for Failure to Perform as Promised

One likely area for disputes regarding agentic AI will be claims regarding the failure to deliver on promised capabilities. For example:

  • If an AI agent confuses an instruction to purchase 10 chairs at $100 apiece and executes a loop buying one chair 1,000 times, the user of the AI may assert that such an error resulted from the gross negligence of an agentic AI provider, permitting the user to recover beyond a set damage cap.
  • If a company entered a three-year contract to use an agentic AI CRM tool and the tool regularly fails to schedule meetings properly, spams company contacts repeatedly, or fails to properly alert company employees about customer issues, the company may seek to terminate the contract early on the grounds that the tool wholly failed to deliver as promised, and may bolster that argument with a claim that it was fraudulently induced to enter the agreement by promises that the agent could perform functions that it could not perform.
  • If an AI provider distributes an update that either degrades the accuracy of the agent or slows its overall performance (e.g., downgrades the model version on which it is working), users might assert that such conduct breaches the duty of good faith and fair dealing even if the contract does not expressly address the subject.
  • If an AI agent fails to fundamentally perform its task (for example, due to frequent hallucinations or infinite loops), users may assert that the breach frustrated the fundamental purpose of the parties’ agreement, therefore permitting them to terminate the agreement. A similar frustration of purpose argument might be made if regulations are adopted that limit or prevent the intended use of the agentic AI.

Third-Party Injuries

One of the key differences between agentic AI and generative AI (or other SaaS or outsourcing arrangements) is that agentic AI may take actions directly, including interacting with third parties. Those direct interactions open the company deploying the AI agent to potential tort claims from third parties who are adversely affected. This in turn will lead to disputes between the company and the agentic AI provider to determine where responsibility should lie.

One can imagine a range of conduct that might result in such third-party injuries. For example, an AI agent designed to market a product may make comparative claims in private or public fora that the competitor alleges are disparaging. This may trigger a legal claim of commercial slander or disparagement and lead to a dispute between the company using the AI agent and its developer about who bears responsibility. Similarly, an automated customer service system might induce customer purchases or behavior based on inaccurate representations, which may also lead to third-party claims for which the user seeks indemnification from the agentic AI developer. Additionally, as third parties adopt provisions in their site usage agreements intended to preclude access by agentic AI, questions will arise regarding whether an AI agent violated those third-party terms of use. Users of agentic AI models may argue that the model developers should have built guardrails to ensure agents do not access such sites, while the model developers may point the finger back at the user.

Depending on the scope of the AI agent’s actions and the extent of third-party injury, such third-party claims could quickly become very significant. Disputes over which party bears the fault or allocation of liability will likely turn on interpretations of the parties’ contract, technical questions about the design, training and usage of models, and potential testimony about the appropriate industry standards and technical solutions. For example, it might focus on the kinds of guardrails that were (or should have been) implemented, the extent to which the company modified or tailored the AI agents in implementation, and a review of the prompts and direction given to the AI agent before conducting the market research.

IP Claims

The potential for generative AI to develop content (images or written output) that infringes on others’ copyrights has long been recognized.

Agentic AI creates the potential for a different kind of IP claim. Agentic AI models that scrape third-party websites, gain access to copyrighted materials, build new capabilities based on patents, or gain unauthorized access to commercial trade secret information, for example, may lead to claims by the third parties that the user of the agentic AI misappropriated the third parties’ IP. Similar to the potential for third-party injuries noted above, such allegations by third parties will trigger disputes over whether the user of the agentic AI bears responsibility for the attack, or whether the attack should have been prevented by guardrails built into the agentic AI.

A different IP issue will also likely arise where the agentic AI, working at the direction of a corporate licensee, creates new intellectual property (e.g., new code, or new capabilities for servicing customers). Although agentic AI licenses may anticipate and address some such jointly created IP, disputes about whether any IP right even exists, and if so what content it covers, may make it difficult for either the AI user or developer to identify potential breaches of the IP rights allocation, let alone enforce their rights.

A related IP and data security issue may arise in the context of contractual provisions preventing the AI service provider from having access to, retaining, or using data of the AI user to develop new models. Take, for example, the potential of an agentic AI deployed by a third-party IT help desk company that provides help desk functions for its customers and which has access to thousands of user service calls. The parties agree that the AI service provider may not use that data to train its AI models. Assume that a year after the agreement is entered, the AI service provider develops and markets an “IT Fix It” model, targeted at helping to identify and fix IT issues before they generate service calls. Such product may materially impact the company’s business. But it may be difficult, if not impossible, for the company to prove that the IT Fix It model was developed and trained, at least in part, on the company’s data. While discovery might ultimately disclose the training data used on the model, the company may find it difficult to identify its own data within the training data set. Moreover, given that a complaint must be drafted without access to such discovery, it would be difficult to have sufficient information to allege such conduct even on information and belief.

Data Privacy/Cybersecurity

Additional issues are also likely to arise from agentic AI in the area of data privacy and cybersecurity. As part of any agentic AI system, the agent will need access to data owned by the company that deploys the agent. Such access to corporate data inherently creates the potential risk that private or sensitive corporate data may be inadvertently used or disclosed by the agentic AI. Making matters worse, agentic AI is intended to interact with other parties and/or other agentic AI in performing its functions. This exposes the agentic AI service to the risks described in our prior Legal Update on cybersecurity contract provisions.

For example, questions will proliferate around which party holds responsibility for a malicious prompt injection. Should such injections have been anticipated by the agentic AI developer and effectively blocked? Or did the company that used the agentic AI assume the risk that the agent might be attacked in that manner—in the same way that companies effectively assume the risk that their own employees might be phished or scammed and thereby expose corporate servers to attack? While contracts between companies and agentic AI developers may anticipate some of those questions, given the unique nature of each deployment of agentic AI, it is likely that the allocation of responsibility may be resolved only on a case-by-case basis.

Regulatory Violations

An additional area of risk is the increased potential for regulatory violations. While generative AI platforms might recommend actions that would result in regulatory violations, companies typically have a “human in the loop” to monitor the actions for potential legal violations. Agentic AI platforms may create a range of regulatory violations without human intervention, and the responsibility for such violations would be ascribed in the first instance to the company. Allocation of responsibility between a company and its providers may be difficult to fully address in contracts for such services, leaving a hole to be filled with litigation.

For example, the news recently noted that an AI-powered online retailer directed to cease and desist selling scooters capable of traveling over certain speeds to customers in New York. One could imagine a company’s AI agents pursuing sales goals by encouraging consumers to purchase scooters that would violate those regulations. The regulators would hold the company responsible. Absent contractual language addressed to that risk, the company and its AI provider may dispute who should have been responsible for such decisions.

A second example might be an AI agent that a company uses to source products at the best available price, which then communicates with AI agents deployed by other companies toward the same task. One could imagine those AI agents agreeing not to pay above a certain price for such products, violating antitrust provisions. Questions about whether responsibility for such conduct rests with the AI user or developer could again lead to increased disputes between the parties.

Beyond the need to avoid or prevent transactions or agreements that violate the law, agentic AI users may also need to explain the rationale for decisions that the agentic AI made. Such obligations might be addressed through audit rights and verification tools, but users of agentic AI will increasingly insist that the models maintain a log explaining the steps and decision-making process of the AI.

Takeaways

Contracts for agentic AI involve genuinely new risks regarding performance failures, injuries to third parties by agentic AI actions, disputes over IP rights, data privacy and cybersecurity issues, and violations of regulatory obligations. The extent of these risks is not yet known, because the technology is still developing and the litigation is yet to come. Lawyers contracting for agentic AI can help clients maximize value and avoid costly pitfalls by considering the implications of the technology within the existing legal frameworks.

Contracting for Agentic AI

Read more chapters from Contracting for Agentic AI, the latest book in our annual series on technology and outsourcing topics.
mehr lesen

Stay Up To Date With Our Insights

See how we use a multidisciplinary, integrated approach to meet our clients' needs.
Subscribe