Agent Liability Gap
The agent liability gap is the situation in which harm caused by an AI agent cannot be clearly attributed to any single party, because model provider, tool operator, integrator and deploying company each carry only partial responsibility.
The agent liability gap describes the unresolved responsibility when an autonomously acting AI agent causes harm — accountability is often unclear between provider, operator, integrator and user.
Explanation
Classic liability concepts assume an acting person or a defective product. An agent is neither cleanly: it combines a third-party model, third-party tools, your prompts and context data into behaviour nobody fully predicted. Terms of service that largely exclude provider liability and the chaining of agents across system boundaries make it worse. In practice the gap is closed organisationally: documented mandates with clear boundaries, human approval for externally visible actions, complete logs enabling reconstruction, contractual liability and audit rights, and insurance for residual risk.
Marketing Relevance
For companies this is a go-live criterion: no agent should perform externally visible actions until attribution is settled, because in practice the operator usually ends up carrying the loss.
Example
A service agent promises a customer a refund the contract does not provide for. Provider and integrator point to their terms of service and the deploying company pays — a textbook lesson in missing action boundaries.
Common Pitfalls
A common mistake is logging only technical errors. Without recording context and decision basis you cannot later reconstruct why an agent acted the way it did.
Origin & History
As the first agent systems went into production from 2025, cases of erroneous orders, false promises to customers and unintended data disclosure accumulated. In 2026 the attribution question moved to the centre of compliance teams' and insurers' attention.
Marketing Use Cases
Written agent mandate defining permitted actions, limits and escalation paths.
Human approval for every action with legal or financial external effect.
Complete decision logs including context sources to allow later reconstruction.
Contractual clarification of liability, audit rights and notification duties with providers.
Frequently Asked Questions
Is the model provider liable for agent errors?
Under most standard contracts, no. Terms of service largely exclude consequential damages and shift responsibility to the deploying company, which defines the agent's boundaries and operating conditions.
How do you close the gap in practice?
Through tight action boundaries, mandatory approval for externally visible steps and complete logging. Together these turn unclear attribution into documented operational responsibility.
Is insurance available for this?
Yes, specialised AI risk policies have emerged since 2026. They typically require documented governance, logging and defined action boundaries as a precondition for cover.
Related Services
Go deeper: Agentic AI Hub · Governance & compliance