Physical AI liability: the gate that closes before the sale

When a system can move mass, the question that blocks the purchase is who answers when it causes harm. Buyers settle it before they sign.

By Harinderpal Hanspal on June 2026. Updated October 2026

A wrong software recommendation is a bad suggestion. A wrong physical action can injure a worker, damage equipment or halt a line. That changes who decides. In our engagements, the buyer's environment, health and safety (EHS) function and its legal counsel join the purchase of an autonomous physical system, and either one can stop it.

A good demo stalls in the EHS and legal review the vendor never planned for A technical demo leads into a review by the buyer's environment, health and safety and legal teams, which asks who carries insurance, what the failure-mode analysis covers, what human oversight is required and where responsibility sits when the system runs unsupervised. The purchase order sits beyond it, not yet reached. Technical demo A good demonstration does not move the deal EHS and legal review Who carries insurance What the failure-mode analysis covers What human oversight is required Where responsibility sits when unsupervised Either team can stop the purchase Purchase order The deal stalls here while the liability model is open
A good demo stalls in the EHS and legal review the vendor never planned for

Most vendor pitches were never built to answer what those two teams evaluate. "We are working through the liability model" rates as an Assumption, because it is one stated as progress. While the question stays open, a good demonstration will not move the deal. It stalls in a review the vendor did not know existed.

Verified looks different. The buyer's own legal and EHS teams have reviewed and accepted a structure that says who carries insurance, what the failure-mode analysis covers, what human oversight is required, and where responsibility sits when the system runs unsupervised.

The test is short. Ask the buyer to name the person who would sign off on the liability terms, then interview that person. If nobody can be named, the demonstration is sitting on an unvalidated hypothesis.

Go deeper: Validating physical AI before the capital is irreversible sets out the liability dimension in full, Before an AI agent writes to equipment: what the joint CISA guidance asks for shows where liability sits in a running system, An AI agent log can prove nothing changed and still not say who acted explains why a tamper-evident trail still cannot say which agent acted, and how we validate physical AI covers all three added questions.