AI Agent

An AI agent is an AI system that takes actions toward a goal (calling tools, making decisions, operating with varying degrees of autonomy) rather than just answering a single prompt. The agent, not the model underneath it, is what touches the real world, and that is why the agent is what creates real-world liability.

In Depth

Separate the model from the agent. People use the words interchangeably, but the distinction is exactly where liability lives. A model is the underlying capability: it takes input and returns output. An agent wraps that model in a purpose. It has a goal, access to tools and data, instructions about how to behave, and some authority to act. The same model can sit inside a harmless summarizer or inside a system that books appointments, moves money, or drafts clinical notes. The model didn't change. The exposure did.

That's because an agent is defined by what it's allowed to do. Give it read-only access and a human approving every step, and its blast radius is small. Give it tool calls, autonomy, and a high-stakes domain, and a single bad decision propagates into the world before anyone reviews it. What determines how much harm an agent can cause is its autonomy, its scope, and its domain, not the cleverness of the model.

This is the reason Ollive insures per agent rather than per company or per model. Two agents built by the same team on the same model can carry wildly different risk. Pricing them as one thing would overcharge the safe agent and undercharge the dangerous one. Writing a policy for the agent itself lets the coverage match what that specific agent actually does.

What It Looks Like

A vendor ships two products on the same foundation model. The first is an internal tool that drafts meeting summaries a human always reads before anything happens with them. The second is a customer-facing agent that answers billing questions and can issue credits on its own. Same model, same company. The first agent's worst case is a sloppy summary someone catches. The second can tell a customer something wrong, act on it, and create a binding-looking commitment with no human in the path. A risk assessment that treated them identically would miss the entire point. The agent, not the model, is the unit of risk.

Why It Matters For AI Vendors

When you sell an AI agent, you are putting something into the world that acts, not shipping a passive tool. Under the logic of vicarious liability, its actions are treated as yours, the way an employer answers for an employee. Your enterprise customer inherits that exposure the moment they deploy it, which is why their security and procurement teams now scrutinize what your agent can do, not just how accurate the model is.

Framing your product as "an agent" also reframes the risk conversation honestly. The questions that matter are agent-level ones. How autonomous is it? What can it touch? What are the stakes of its domain? Those are the questions buyers and insurers are asking.

Common Questions

The model is the raw capability; the agent is the model plus a goal, tools, data access, and authority to act. The model produces output. The agent does things with it. Liability attaches to what gets done.
More autonomy means fewer chances for a human to catch a mistake before it causes harm, so it generally does raise the risk profile. That's why human-in-the-loop design raises an agent's Agent Trust Score and lowers its premium.
Because risk lives at the agent level. One company can run a dozen agents of wildly different risk. Per-agent underwriting prices each for what it actually does instead of averaging them into one blunt number.
← PreviousAgent Trust Score (ATS) Next →AI Data Disclosure

See where your AI agents stand.

Get an Agent Trust Score, map your liability exposure, and find out what it takes to make your AI agents insurable.