Last week, Allstate told analysts it will launch its own personal agentic assistant early in 2027 — an assistant that will work with other assistants, like Muse, and quote competitors too. Tom Wilson's line on the call was that Allstate is "now marketing to machines and people, not just marketing to people."
Two weeks ago in this series, the quote left the building: a carrier put BOP quoting inside ChatGPT, and the quote started happening in an assistant the carrier doesn't own. Now the buyer is leaving the building too. A distribution executive put the next step plainly this week: a customer's voice AI agent calls an insurer's sales center, gives the driver and vehicle details, gets a price, and binds the policy — and the rep on the other end may never know it wasn't a person.
What it means for carriers
First, disclosure is the wrong question. The laws people reach for were written for a business's bot talking to a consumer, not a consumer's bot calling a business. And even where a law says the AI has to answer honestly when asked, somebody has to ask — which means the question has to be on the sales script, in every call center, on every shift. Anyone who has added one question to a sales process across a large operation knows what that takes: training, QA, and a few quarters before it actually happens on every call. Disclosure is a coordination problem wearing a legal costume, and the coordination is the hard part.
Second, authority is the real question, and it is older than AI. Agency law has always known three kinds of authority, and only one of them binds a policy: express authority, granted orally or in writing, which permits an agent to countersign, issue, deliver, and bind. Implied authority covers what is not formally communicated — it does not cover binding a policy. A voice agent that answers the underwriting questions, hears the recorded disclosures, and authorizes the payment is exercising express authority on the customer's behalf. The question is whether the customer actually granted it, in a form the carrier can produce. Today the whole call is built around the idea that the voice on the line is the person applying. There is no login and no click trail — just a caller who sounds like a person. The first real test will come at a claim, when a carrier finds a wrong answer on an application and has to decide whether something the customer's software said on a recorded call counts as the customer's own answer. That is not a disclosure question. It is an authority question, and the industry has no clean answer to it yet.
Third, the ledger cuts both ways. The recorded call is the evidence in that claim — and it is also the training data for the other side's next call, which is what this series argued yesterday. The carrier that keeps a clean, attributable record of who said what, under what authority, wins the claim. The carrier that can't produce the record pays. The same discipline that protects the carrier in the dispute is what makes the dispute winnable.
From the operator's chair
One: put one question on the script, and log the answer. "Are you the applicant, or an agent acting for the applicant?" Every call, every shift, with the answer in the record. It is the cheapest control in this whole problem, and it is the one that makes every other control enforceable.
Two: demand a delegation record before the bind. A scoped letter of authority from the customer — what the agent may do, for which policy, for how long — pinned in the ledger next to the recording. No delegation, no bind. That is the Track 3 device, applied to the customer's side of the desk: the same discipline carriers already demand of their own agents, extended to the customer's agent.
Three: price the authority, not the call. Know what a bound-by-agent policy costs you — in claims, in disputes, in the reserve you hold for the answers you can't attribute. The cost per decision includes the cost of a wrong answer given by a voice you couldn't identify. If you can't state it, you aren't ready to take the call, whatever your straight-through-processing rate says.
Four: put the customer's agent on your risk register, next to your own models. Your AI governance program tracks your models' bias, your data handling, your examiner exposure. Add one line: the adversary's models, trained on your outcomes — and the customer's models, acting with your customer's authority. It changes what you settle, what you document, and what you bind.
The quote left the building. The buyer left the building. The question is no longer whether the machine can buy the policy. It is whether anyone can prove who bought it.
— Krishna