When the Caller Is a Bot: Liberate's Amrish Singh on AI Shoppers

Consumer AI assistants have started picking up the phone and calling insurers for quotes. Meta's Muse agent launched in the U.S. in September and Insurify blocked it within weeks. Carriers and agencies can't block their phone lines that way, so the question becomes what you do when the "customer" on the line is software.

Liberate just launched AI Intercept, which detects when an inbound caller is an AI agent and routes it accordingly. I asked Amrish Singh, Liberate's CEO, how the detection works, what happens when it gets it wrong, who owns the mistake when a bot buys the wrong coverage, and whether a small independent agency needs to care yet. His answers are below, lightly formatted.

How do you actually tell an AI agent from a person on a phone call? What signals are you picking up, and how often do you get it wrong?

We use a specialized AI model we developed called Argus-1. It looks at how the caller sounds, the language they use, and how the conversation unfolds to assess whether we are speaking with a person or an AI agent.

If it cannot make a confident determination, it can engage the caller in conversation to gather additional signals. No detection system is perfect.

We set a high bar before treating a caller as an AI. If there is uncertainty, we err on the side of treating them as a person.

What happens when you misclassify a real policyholder as a bot, especially one calling about a claim?

We are incredibly sensitive to this scenario. We would rather mistake an AI agent for a person than mistake a person for an AI agent, especially when someone is calling about a claim and may be in a vulnerable state. We configure our model with that priority in mind.

The carrier or agency sets the rules for how calls are handled. In the approach I described, we only route a caller through the AI-specific path when we have a very high degree of confidence. Otherwise, a human takes the call. The goal is to protect access to help while managing the additional traffic from consumer AI agents.

Insurify blocked Meta's Muse over concerns that automated quoting strips out limits, deductibles, and eligibility details. How does AI Intercept make sure a quote given to a consumer's AI agent doesn't lose those details?

On our side, the carrier or agency’s requirements apply regardless of whether the caller is a person or an AI. Our AI collects the required information, checks it, and confirms it with the caller. If essential information is missing, the process does not advance to a licensed agent for a quote. Liberate’s AI does not itself quote coverage.

But collecting complete information and ensuring that a consumer’s AI faithfully explains the resulting quote are two different problems. We can control our part of the interaction. We cannot guarantee that another company’s AI will preserve every qualification when it reports back to the consumer. That remains an issue the industry needs to address.

If an AI agent requests or binds coverage on a consumer's behalf and it turns out to be the wrong coverage, who owns that? The consumer, the agent's maker, the carrier, or the agency?

There is no settled answer that covers every scenario. My view is that a business cannot simply say, “The AI did it,” and consider its responsibility over. But when a consumer’s AI, an agency, and a carrier all participate in a transaction, determining who made the mistake and who bears responsibility becomes more complicated.

Our role is to make our part of that process controlled and traceable. Liberate’s AI collects information, checks it, confirms it with the caller, and sends it to the client’s system. We retain the transcript and the information submitted so there is a record of what was said and what was passed along. We also review those records after the call for compliance issues and escalate discrepancies. A licensed agent remains responsible for verifying the information before binding a policy.

These controls help prevent mistakes and establish what happened if something goes wrong. They do not, by themselves, resolve the broader liability question.

How does this sit with licensing and E&O? At what point does the conversation have to go to a licensed human?

In our workflow, the AI collects and verifies information. It does not act as a licensed producer or independently quote or bind coverage. We hand off to a licensed agent for the quote, and that agent must verify the information before binding.

When we identify the caller as an AI, we tell the licensed agent so they can take that into account. We also hand off earlier when the carrier or agency’s rules require it, or when the AI cannot reliably complete a task. For example, repeated difficulty capturing and confirming a VIN can trigger a handoff rather than allowing the conversation to continue with uncertain information.

On E&O, my view is that using AI does not make responsibility for errors disappear. The record of the conversation and the information submitted helps establish what happened. Whether a particular loss falls under an E&O policy depends on the circumstances and the policy.

You say AI-placed calls rarely convert. What are the early numbers on volume and conversion you're seeing from carriers?

We have seen roughly a threefold increase in the number of inbound calls from consumer AI agents, although the total volume is still low. The calls also come in spikes. My suspicion is that some consumer applications are still testing how their agents interact with businesses.

Most of what we see today is shopping for quotes. Service, and cancellation requests remain rare. We do not have a conversion figure to share here, and I would be cautious about drawing broad conclusions from these early volumes. What we are watching is the rate of change. A person has limited time to call around for insurance. An AI can contact many agencies and carriers on their behalf. That can increase the work required to handle quote inquiries without a corresponding increase in the number of consumers ready to buy.

For an independent agency without a big contact center, is this relevant today or a few years out?

It is relevant today as something to prepare for, even though volumes are still low. A small agency has fewer people available to absorb an increase in shopping calls. If consumer AI agents start calling multiple agencies for every quote request, that can consume staff time quickly. This will have major workforce management ramifications. The first step is to decide how you want to handle those calls: what information you require, when a licensed agent should get involved, and how you protect staff time while keeping the agency accessible to customers. You do not need a large contact center to face this issue. Over time, I expect an AI front door to become as normal for a business as a website.

You've said voice is a bridge and agents will eventually talk machine-to-machine. What does that look like for how carriers take quote requests three years from now?

Voice is the bridge because insurers and agencies already answer phones. A consumer’s AI can use that existing channel today without waiting for every carrier to build a new way to accept requests.

Over the next year, not three years, I expect more of those requests to move directly between AI systems. Instead of an AI reading out a driver’s details over the phone, it could send that information in a format the carrier’s AI system can process and receive a response that keeps the premium, limits, deductibles, and eligibility conditions together. That would make the exchange more efficient, but it would not remove the need to verify who the AI represents, what authority the consumer has given it, and when a licensed person must participate. Carriers will still set the rules for access and transactions. The technology may change faster than the licensing and liability framework, so I expect voice and direct system connections to coexist.

Amrish Singh is CEO of Liberate. Connect with him on LinkedIn.