Quick answerSet a hard cap, typically two clear restatements of the same request within one conversation, after which the policy requires an immediate handoff regardless of whether the underlying issue could technically be resolved by the agent. Do not keep restating the disclosure or the agent's capability in response to repeated insistence; a customer who has already heard "you're speaking with an AI" and asks again is not asking for information anymore, they are asking to leave the channel. Build a scripted, single-sentence acknowledgment that stops arguing and starts routing, and log the trigger so the pattern shows up in your escalation metrics rather than getting buried in transcript review. ---
Why this is a different moment than the first ask
Our post on what to do when a customer asks if they're talking to a human or an AI covers the single-ask disclosure moment: answer plainly, once, and continue. That advice assumes the customer accepts the answer and moves on. This post covers what happens when they do not, when the same customer restates the same demand for a person a second or third time inside one conversation despite a correct and clear disclosure already given. Treating the second ask the same way you treat the first, with another polite explanation of what the agent is and is not, reads as stalling even when the agent means it sincerely.
Set a numeric trigger, not a judgment call
Leaving the decision to "use good judgment" produces wildly inconsistent behavior across conversations, since a model has no reliable sense of how many times is too many unless you give it one. A concrete trigger, for example two repeated variations of "get me a person" within the same session, removes the ambiguity and gives you something you can audit later. Keep the count scoped to the current conversation only; do not carry a running insistence count across sessions, since that is a different, standing-preference question covered separately.
What the agent should say at the trigger point
The response at the trigger point should do three things in one short turn: acknowledge the request plainly, state that a human is being brought in now, and give an honest expectation for how long that will take. It should not re-explain that the agent is an AI, re-ask what the customer needs, or offer one more attempt to help, since any of those reads as a delay tactic to someone who has already asked twice. Whatever the receiving human sees needs to include the full context so the customer is not asked to repeat themselves, the same requirement covered in what context a human should see the instant an AI agent escalates.
What this is not
This policy is not a judgment about whether the customer's underlying issue was something the agent could have solved. Insistence on a human is itself the signal to escalate, independent of whether the request was simple or complex, or whether the agent's answer so far was correct. Confusing this with a standing account-level preference, where a customer states they always want a human going forward, is also a mistake; that is a separate, durable setting covered in whether an always-route-me-to-a-human preference should persist across sessions, not something a single conversation's insistence count should silently create.
FAQ
Does the trigger count reset if the customer asks about a different topic in the same conversation?
No. Count insistence within the whole conversation, not per topic, since a customer switching topics while still frustrated about not reaching a person is still signaling the same thing.
What if no human is available when the trigger fires?
Say so honestly, with a real wait time if you have one, and queue the conversation rather than quietly reverting to the agent. A customer who already asked twice will notice immediately if the handoff turns out to be another agent turn in disguise.