Contracting With Agentic AI Vendors: What Belongs in the Agreement

AI vendor contracting depends on the specific tool, its access, and your risk tolerance. This post is informational. For a specific agreement, consult a licensed attorney.

There is a meaningful difference between an AI tool that answers a question and one that takes an action. The first generates a draft you review. The second sends the email, moves the funds, updates the record, books the meeting, or files the ticket, on your behalf, sometimes without a human looking at each step. That second category, often called agentic AI, is showing up fast in software your company already uses.

The contracts, in most cases, have not caught up. Vendors are shipping tools that act with real authority under agreements written for tools that only produce output. When an agentic system does something wrong (and it will), the question of who is responsible gets answered by a contract that never contemplated the tool taking action at all. Here is what a serious agentic-AI agreement should address.

Scope of authority

The first question is not technical, it is legal: what is this system actually authorized to do on our behalf? An agent that can draft a reply is different from one that can send it; one that can recommend a refund is different from one that can issue it. The agreement (and the configuration it references) should define the boundaries of the agent’s authority explicitly: what actions it can take, up to what thresholds, and what requires human approval. Undefined authority is the root of most agentic-AI disputes.

Audit logs and traceability

When an autonomous system takes thousands of actions, you need a record of what it did and why. The agreement should require the vendor to maintain and give you access to audit logs: a traceable record of the agent’s actions, the inputs it acted on, and the decisions it made. Without that, you cannot reconstruct what happened after an incident, you cannot demonstrate compliance to a regulator or customer, and you cannot prove the agent acted outside its authority.

Liability for autonomous actions

This is the provision most standard AI agreements handle badly. If the agent takes an action that causes harm (sends confidential data to the wrong recipient, makes a commitment you did not authorize, executes a transaction in error), who bears the loss? Many vendor agreements disclaim responsibility for the tool’s output entirely and cap liability at a trivial amount. That allocation might be defensible for a tool that only suggests. It is a much harder sell for a tool that acts. The liability framework should reflect that the vendor built a system to take actions, and should not leave you absorbing the full cost of actions the vendor’s system took wrong.

Data, training, and IP

Agentic tools often need broad access to your systems and data to function. The agreement should address what the vendor may do with that access: whether your data can be used to train the vendor’s models, how long it is retained, and who owns the outputs the agent produces. Broad operational access plus permissive data terms is a combination worth catching before signing.

Human oversight and the off switch

Finally, the agreement and the product should give you meaningful control: the ability to require human review for defined categories of action, to set thresholds, and to stop the agent quickly if it starts behaving unexpectedly. “Human in the loop” is not just good practice. It is a term you can and should negotiate, because it directly limits the scope of harm an autonomous system can do before someone intervenes.

The bigger picture

Agentic AI is not a reason to avoid these tools. Used well, they take real work off your team. But the shift from AI that suggests to AI that acts is a genuine change in your legal exposure, and it deserves a contract that reflects it rather than one recycled from the last SaaS deal. (For the broader picture of how this fits alongside AI use policies and compliance, see the firm’s AI Governance and Technology Transactions practices.)

If you are evaluating or renewing an agreement with a vendor whose tool takes actions on your behalf, talk to MB Legal before you sign.

This post is informational and does not constitute legal advice. AI vendor agreements vary widely by tool, access, and use case. For guidance on a specific contract, consult a licensed attorney.