Should an AI customer service agent ever speak to another AI?

When the customer is represented by an AI agent (a personal assistant, a procurement agent, a price-comparison bot), the support-side AI is now speaking to another AI, not a person. The conversation can be more efficient, but identity, authority to act, and audit get harder. The right pattern is to treat AI-to-AI contacts as a defined category, not a special case of the human one.

A customer's personal AI assistant contacts an airline's support AI to rebook a flight. The two negotiate, settle on an option, and confirm. No human was in the conversation. The customer's AI says it had authority; the airline's AI took its word for it. Later the customer disputes the rebooking. The question of who agreed to what becomes a question about two systems talking to each other.

What people in the field are saying

The Agentic Commerce Frontier's issue tracking agentic commerce argues that AI-to-AI customer service is moving from edge case to baseline faster than most teams expect. The infrastructure for it is being built now.

What changes when the other side is an AI?

Three things. The conversation is faster and more structured: no small talk, no ambiguity, no emotional content. The authority question becomes central: did the customer authorise this AI to bind them on this kind of action? The audit becomes machine-to-machine: there is no human transcript to read, only two logs.

Should the AI know it is talking to another AI?

It should, for two reasons. Some actions should require human confirmation specifically because the customer is not in the loop. Some disclosures and risk frames are different when the counterpart is a machine: a marketing message intended to surprise and delight does not work; a price negotiation might work better.

How does authority work?

The customer's AI needs to present credentials (delegation tokens, signed mandates) that prove the customer authorised it to act on their behalf on this kind of decision. The support AI checks those credentials and the policy under which they are issued. Without that, the support AI is acting on the say-so of a system whose connection to the actual customer is unverifiable.

Where does this go wrong?

Two systems agree to something neither customer's owner actually wanted. The support AI accepts a counterpart's claim of authority without verifying. Audit logs on the two sides disagree, and there is no third record to settle it. The optimistic case (faster, cleaner conversations) and the failure case (cross-AI errors with no human in the loop) both look like the system working.

What is the practical first step?

Define the category. Decide which contact types your AI may handle when the counterpart is another AI, and which require human confirmation regardless. Log the credentials and the policy basis on which authority was accepted. The infrastructure for the rest will take time; the policy decisions do not.