Contracting Opportunities and Challenges for Emerging Agentic AI Companies

Defining the Agent, the Service, and the Scope of Authority

AI agents can initiate actions, interact with third-party systems, modify data, and execute a variety of different transactions with varying degrees of human oversight. Given these capabilities, the contract should clearly delineate the boundaries of what the AI agent is authorized to do and what oversight the customer is required to provide. Without a clear delineation of responsibilities, providers could be surprised to learn that they may be held responsible for autonomous actions by their AI agents, whether through services warranties, performance commitments, gap-filling contract interpretation, implied warranty theories, or negligence standards that measure the provider’s duty of care against the full range of foreseeable agent conduct.

Providers can reduce this risk by including a detailed “Agent Scope” or “Automation Authority” schedule (a construct we’ve seen also described as a “delegation of authority” subject to “policy guardrails”). That schedule would specify the categories of actions an agent may perform including, for example, the systems it may access, the triggers or conditions under which it operates, and—critically—the points at which human review or approval is required before the AI agent proceeds. Such a schedule could also clarify the extent to which customer adjustments to the tool will impact the provider’s warranties and performance commitments.

Pricing and Monetization: Moving Beyond the Subscription

Pricing and monetization approaches are evolving. Many agentic AI startups have experimented with per-seat and per-user licenses, which have been standard SaaS and software pricing practices for decades. However, per-seat or per-user licensing models fit well only if value correlates with the number of people accessing the platform. That’s true for a genAI tool that needs prompting. But an AI agent that autonomously processes transactions, triages customer inquiries, or manages procurement workflows delivers value irrespective of the number of users, might deliver value without any users whatsoever.

Fixed-price subscriptions have been a popular way for generative AI providers to price AI products. However, fixed-price subscriptions can also have little correlation with value. So, the AI provider will lose sales to low-volume users and will lose money on medium- and high-volume users. With agentic AI, there is also the complication that the AI agents are not limited by a person’s ability to input prompts.

To address these issues, many providers are experimenting with outcome-based pricing, per-task or other per-action fees, token-based or other consumption-based tiers, and hybrid models that combine a platform subscription with variable charges tied to agent activity or success. Each of these models has distinct contractual implications. Outcome-based pricing requires an agreed-upon definition of potential outcomes and mechanisms for measuring outcomes and resolving disputes. Per-action and consumption models require an agreed-upon definition of how actions and consumption will be counted and reported, along with audit rights, where necessary, that give customers confidence in billing accuracy. Customers, for their part, often push for cost predictability, which providers in some cases can offer through the use of spending caps, committed-use discounts, or advance-purchase credits. However, many providers still face the challenge that the token costs and other costs of processing agentic workflow often cannot be predicted with any certainty before contract execution.

Ambiguity in pricing provisions invites billing disputes and can undermine the provider’s ability to profit from its product. As a result, we recommend that providers reflect their chosen model or models with specificity in form contracts, including clear definitions, measurement methodologies, reporting cadences, and reconciliation procedures. In addition, we recommend that providers develop playbooks and other guidance for their contract negotiations.

Providers should also understand how other contractual provisions may impact their ability to profit from their chosen pricing models. For example, a provision that would allow a customer to terminate the contract before the agentic AI tool is able to fully realize the expected outcomes could undermine the provider’s ability to profit from the investments it makes under the contract. 

Warranties: Calibrating for Autonomous Conduct

Given the autonomous nature of agentic AI tools, customers will value assurances that the agent will perform as described under foreseeable operating conditions—assurances that go beyond those of a typical SaaS-style contract and that may look more like traditional business process outsourcing warranties.

At the same time, providers may be tempted to disclaim all liability for agentic actions and outputs. A provider might see that as following an industry standard from genAI, and also flowing down disclaimers from upstream cloud and LLM providers. Blanket disclaimers are common in pilot arrangements, which might make a provider believe that blanket disclaimers will be equally successful for AI agents that autonomously perform transactions, modify data and perform other consequential actions.

However, this sort of a disclaimer by a provider risks rendering the provider’s offering commercially unviable in competitive procurements. Sophisticated customers view blanket disclaimers as a signal that the provider lacks confidence in the reliability of its own technology or has not matured its risk framework.

Instead, providers should craft warranties that are accurate, meaningful, and bounded. A warranty that the agent will operate materially in accordance with its defined scope when deployed within specified parameters, coupled with clear exclusions such as for misuse, configuration by the customer, and third-party system failures, will be a more defensible and commercially credible position.

Providers also face the risk that their agentic AI solutions might be characterized as professional services rather than as software products. For example, because agentic AI performs many tasks traditionally done by humans, including making decisions and exercising judgment, some might claim that a “reasonable skill and care” or “workmanlike manner” standard should apply. Professional services warranties often imply accountability for results rather than mere conformance to specifications, and in regulated industries might expose a provider to claims resembling professional malpractice.

To mitigate this risk, providers could expressly disclaim professional services characterizations in their agreements, clarify that the solution is merely a technology tool, and state that the provider is not rendering professional advice in any licensed or regulated field. Where the agent operates in domains that typically require professional licensing, the contract could make it clear that outputs are informational only and do not constitute professional advice. Contractually requiring the customer to conduct a human review for consequential decisions can further reinforce the technology-tool characterization and limit exposure to professional-standard liability.

Data Ownership and Output Rights

Agentic AI raises data rights questions that go beyond conventional SaaS frameworks. Providers should establish a clear contractual taxonomy distinguishing between: (1) customer input data (owned by the customer and processed under license); (2) agent outputs (work product generated by the agent in performing its functions); and (3) training data (whether and how customer data or agent interaction data may be used to improve the provider’s products or services).

Before addressing ownership, providers should understand two threshold limitations. First, AI-generated data may lack copyright protection if there is insufficient human authorship, and courts have not recognized AI systems as patent inventors. Second, because a provider’s agentic AI tool serves multiple customers, similar prompts from multiple customers may generate similar outputs. So, while a provider can agree not to share prompts, customers cannot expect exclusive rights to what they receive from AI agents.  

With these limitations in mind, providers should expressly retain ownership of their pre-existing and independently developed materials—including underlying models, algorithms, and general-purpose methodologies. Failing to preserve these rights can result in customers claiming ownership over assets the provider brought to the engagement independently.

For agent outputs, the most commercially practical approach is often to grant the customer ownership or an unrestricted license to outputs generated from the customer’s data, while the provider retains rights in the underlying technology. However, contracts should include language acknowledging that similar outputs may arise independently and clarifying that the customer’s ownership does not restrict the provider’s or other customers’ use of such outputs.

Customers increasingly expect that their data not be used for training absent explicit consent. Providers that build this restriction into their standard terms can hold a competitive advantage and signal that they understand the risks they are asking customers to assume. Alternatively, providers may offer an opt-in or opt-out framework for training-data use, with appropriate anonymization standards.

If the parties wish to protect agentic outputs despite limitations under copyright and trademark law, the contract should include appropriate reasonable measures to preserve the secrecy of the outputs, such as confidentiality protections and access restrictions.

Navigating the Regulatory Patchwork

Providers must contract against the backdrop of an evolving and fragmented regulatory environment. The European Union’s AI Act imposes risk-classification obligations and transparency requirements that can directly affect providers offering agentic AI solutions in the European Union. In the United States, the regulatory landscape remains a patchwork of sector-specific guidance, state-level legislation and federal executive action, all of which may shift with changes in administration and congressional priorities. Internationally, jurisdictions from Brazil to Singapore are developing their own AI governance frameworks at varying speeds and with differing emphases.

For providers, this regulatory uncertainty has direct consequences. Customers will seek commitments that the provider will comply with AI laws and cooperation in the form of regulatory audits, impact assessments, and incident reporting. Providers should decide in advance which of these commitments they can and cannot support technically, operationally and financially, rather than negotiating each request from scratch. Providers should also include contractual terms that anticipate regulatory change. More specifically, contracts should carefully address whether changes to applicable legal requirements are included in the existing fees, treated as chargeable changes, or handled through a shared regulatory-change mechanism. Emerging companies should be particularly careful to avoid open-ended compliance obligations that could expose them to liability for entirely new regulatory frameworks.

Conclusion

More emerging providers are offering agentic AI solutions. Those providers need contracts that are customized for the unique aspects of the technology underlying agentic AI. These contracts, if handled well, will not only help protect those providers and further their goals, but also become part of the product and act as evidence that the provider understands the risks it is asking its customers to assume.