Quick answerSet your liability cap as a defined multiple of fees paid (commonly one to twelve months, depending on deal size and risk category), carve out gross negligence and willful misconduct from any cap since courts frequently refuse to enforce caps against those anyway, and tie the cap tightly to a clear definition of what counts as an agent "error" so you're not negotiating against an undefined term. This is the reverse of negotiating a vendor's SLA as a buyer; here you're the one setting the ceiling, and the leverage runs the other way.
This is the mirror image of a familiar negotiation, not the same one
Most teams have already been through negotiating SLA and support terms as the buyer of a vendor's AI agent. Setting the liability cap your own company offers to ITS customers is a structurally different exercise, because you're now the party with something to protect, and the customer's leverage, urgency, and objections will look different from what you experienced as a buyer. Don't assume the terms you fought for as a buyer are automatically the terms you should offer as a seller; the risk allocation runs the opposite direction.
This is also distinct from budgeting for liability insurance, which is about what you spend on coverage. The liability cap is a contractual term about what you're exposed to in the first place, independent of whether or how much of that exposure is insured.
Anchor the cap to a defined multiple of fees, not an open-ended promise
The standard commercial pattern is capping liability at some multiple of fees paid under the contract, often over a trailing twelve-month period. For AI agent errors specifically, err toward the higher end of what's typical in your category if the agent handles anything with direct financial consequences for the customer (payments, contractual commitments, regulated advice), and toward the lower end for lower-stakes use cases like general support deflection.
Carve gross negligence and willful misconduct out of any cap
Almost every jurisdiction with meaningful commercial law will refuse to enforce a liability cap against conduct that rises to gross negligence or willful misconduct, so negotiating as if the cap covers everything is negotiating against a term that won't hold up anyway. Building the carve-out in explicitly, rather than getting surprised by a court striking the cap down later, keeps your actual risk model honest.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Define "error" before you cap liability for it
A cap is only meaningful if "error" is defined tightly enough that both sides know what it covers. Vague language inviting the customer to argue that any unsatisfactory outcome counts as an "error" erodes the value of the cap entirely. Tie the definition to something measurable: a factually incorrect answer that caused quantifiable harm, an unauthorized transaction, a policy violation directly caused by the agent, rather than "the agent didn't help me."
What customers will actually push back on
Enterprise customers, in particular, will push for a higher cap or for full liability on specific categories like payment errors. Decide in advance which categories you're willing to carve out to a higher cap for the right deal size, and which are non-negotiable regardless of deal size, so your team isn't improvising the answer live in a negotiation.
FAQ
Should the cap differ by customer segment or be a fixed policy? A fixed floor with room to negotiate upward for larger deals is the common pattern; a fully case-by-case cap with no floor invites inconsistent exposure across your customer base.
Does insurance change what cap you should be willing to offer? It can support offering a higher cap on categories your policy actually covers, but the cap and the insurance coverage should be reviewed together, not set independently, since a generous cap unsupported by matching coverage is real uncapped-in-practice exposure.
Should the cap be the same across your entire customer contract, or specific to the AI agent's functionality? Increasingly, companies carve out a specific AI-agent-related liability section distinct from general contract liability, since the failure modes and risk profile of agent errors differ enough from general service liability to warrant separate terms.
How does this interact with the agent's actual transaction authority? The higher the dollar value of transactions the agent can autonomously execute, the more scrutiny the liability cap for that category deserves, since uncapped or loosely capped exposure scales directly with what the agent is allowed to do on its own.