In April I wrote two pieces about where insurance distribution was heading. The first argued that the interface was disappearing: insurers were quoting inside ChatGPT, and the first explanation of value was moving from the insurer to the AI. The second went further. The next buyer of your insurance product, I suggested, might not be a person at all. It might be an agent acting on one's behalf, and an agent doesn't care about your UX. It cares about your data structure.
Five months later, that buyer has turned up. What I didn't anticipate was how quickly the industry's first response would be to shut the door.
What happened in September
On 8 September, Meta launched Muse, a personal AI agent that can open a browser, fill in forms and check out on behalf of the person using it. It isn't a chatbot that describes products. It runs the errand. Within a fortnight Meta had partnered with Shopify to enable agentic checkout across every Shopify store, and Muse had expanded from the US into Canada.
On 29 September Meta pointed it squarely at the customers commercial insurers care most about. Muse for Small Business connects the agent to a business's Facebook Pages, Instagram analytics and ad accounts, and to tools like QuickBooks, Shopify, Slack and Stripe. Meta's pitch is that you give Muse a goal and it gets it done. According to Axios, roughly a third of early Muse users were already connecting it to a business account before the small business version existed.
Think about what that means for an SME owner whose business pack renews next month. The agent already knows what the business sells, what it turns over and who its customers are. Asking it to "find me cheaper insurance" is a natural next request. The agent will go and do it, on websites built for a human to fill in.
The twist: the company I held up as the model blocked it
In April I cited Insurify as one of the first insurers to distribute inside ChatGPT. In September, Insurify blocked Muse from its comparison platform.
Its reasoning is worth reading closely, because it isn't anti-AI. Insurify's argument is that an insurance quote is an offer of cover, not a price. When an agent scrapes a quote it tends to keep the number and drop the context: limits, deductibles, discounts applied, eligibility conditions, required disclosures. Its co-CEO put it in a single line: "A quote without its context is not a fair comparison. It is a number."
The second objection is economic. Each quote request can trigger paid third-party data checks for the carrier, whether or not a real shopper is behind it. An agent shopping in bulk creates acquisition cost with no corresponding customer, and someone ends up paying for that. Speaking to Digital Insurance, Insurify's co-CEO said that when he tested Muse it behaved more like a scraper than a shopper, and that the industry needs quote-to-bind ratios that reflect real consumers rather than bots.
Insurify still supports AI agents. It just supports them through APIs and MCP integrations it controls, not through an agent driving its consumer website. Others are drawing the line differently. Digital Insurance reports that Kin, the direct home insurer, is among the companies weighing whether agent reach is worth the distortion, and it has been reported to be adapting its quotes so agents like Muse can retrieve them. Employers, the US workers' compensation specialist, doesn't let third-party agents into its systems at all. Outside insurance, Amazon blocked Muse too, arguing the agent doesn't identify itself when it browses.
What I got right, and what I missed
The core prediction held. Machines are now acting as the buyer's representative in insurance, and they are doing it at consumer scale rather than in a lab.
What I underplayed was the shape of the arrival. In April I framed the challenge as readiness: could an agent reach your product if it wanted to? That turns out to be the second question. The first is permission: which agents do you let in, through which door, and on whose terms?
There are now two quite different channels, and they need different answers.
| Invited agents | Uninvited agents | |
|---|---|---|
| How they arrive | An app, API or MCP integration you built or approved | A general-purpose agent driving your public website |
| Who controls the context | You. Limits, excesses and exclusions travel with the price | The agent. It decides what to keep and what to summarise away |
| Cost per quote | Known, and you can meter it | Unknown, and invisible until you look for it |
| Identity | Authenticated | Often unidentified, sometimes indistinguishable from a person |
| Examples | Insurers' ChatGPT apps; Insurify's own API and MCP access | Muse filling in a consumer quote form |
My April advice, make your product machine-readable, only ever addressed the first column. The second column is where the volume is about to come from.
What I'm seeing in Australia
Muse isn't available here yet. As of the end of September it's live in the US and Canada only, and Meta hasn't given a date for Australia.
That doesn't mean Australian brokers are agent-free.
Last week, a broking business I'm involved with received three enquiries from startups shopping for insurance. All three had clearly been written by AI. They shared an identical structure, despite coming from three different businesses, and they answered questions the enquiry form never asked. When the team replied, each reply was met with an automated response. No founder came back.
Three enquiries is an anecdote, not a trend, and I can't tell you which tool sent them. It wasn't Muse, which can't be used here yet. It may have been a founder's own assistant with an auto-responder rather than a fully autonomous agent. But the shape is exactly the one this piece is about. A machine put the submission together. A person at the broker read it and did the work. The reply went back into a machine. Nobody on the buyer's side was present at the point where a broker would normally start asking the questions that matter.
It's also no surprise that startups were first. Founders are the earliest adopters of AI tools in any market, and they're the buyers most likely to treat insurance as an errand to delegate rather than a decision to make.
Why commercial lines are more exposed than motor
The debate in the US is mostly about personal lines, where a bot-generated quote costs the carrier a data call or two. In Australian SME commercial insurance, the cost structure is different and, I'd argue, worse.
A great deal of small business insurance here is still quoted with a human in the loop. A broker reviews the submission. An underwriter refers it. Someone checks the occupation, the turnover, the claims history, and comes back the next business day. That work is paid for by the policies that bind.
Now imagine an agent that has been asked to find the cheapest cover. It doesn't fill in one broker's form; it fills in ten. Each one lands on a person's desk looking exactly like a genuine enquiry. Nine of those people do the work for nothing, and every one of them sees their quote-to-bind ratio fall without knowing why. Those three startup enquiries were a small version of this: real broker time spent on submissions whose only response was an autoreply. The more manual your quote process, the more an uninvited agent costs you, which inverts the comfortable assumption that specialists and brokers are insulated by complexity.
There's a regulatory wrinkle here too. Under the design and distribution obligations, distributors have to take reasonable steps so that products reach customers within their target market, which is why so many quote forms ask eligibility questions. When an agent answers those questions on someone's behalf, drawing on what it has inferred from their accounts rather than what they told it, the reasonable-steps question gets harder to answer. So does the question of whether the customer took reasonable care not to make a misrepresentation. I don't think anyone has a settled view on either yet. Australian insurers and brokers would be wise to form one before the agents arrive in volume rather than after.
What to do now
The April checklist still stands: make your products discoverable, get your API layer real, own your specialism. September adds four things.
Decide your agent policy, per agent. Partner, allow or block. Not a blanket position, and not a decision made by default because your bot filter happens to catch some agents and not others. Insurify and Employers have made two different, defensible choices, and Kin appears to be heading towards a third. Making none is the only indefensible one.
Build the front door before they find the window. If you want agent traffic, give agents a proper route in: an authenticated quote endpoint that returns the price together with its context. The quote should carry its own limits, excesses and key exclusions, so that whatever summarises it downstream has the information to summarise accurately. Uninvited agents use your consumer form because there's nothing better available.
Instrument before you decide. Most Australian insurers and brokers can't currently tell how much of their quote traffic is human. Start measuring: completion speed, field-fill behaviour, declared user agents, quote-to-bind by source. You can't set a sensible policy for traffic you can't see.
Protect the people in your quote process. If a human reviews every commercial enquiry, an agent-originated enquiry needs a different path: triage, a holding response, or a request that the human principal confirm before an underwriter spends time on it.
The question has changed
In April I asked: if a customer asked an AI assistant for the insurance you sell, and that AI could quote and bind in real time, would your product be available?
That question still matters, but September showed it was incomplete. The agents aren't waiting to be invited. They're arriving through the public website, filling in forms meant for people, and taking the price without the policy.
The better question for this week's leadership meeting is: when an agent turns up on behalf of a customer, do we know it's there, and have we decided what it's allowed to do?
Sources: Forbes on Muse's launch and Amazon's block (Sep 2026), PYMNTS on Shopify and Muse (Sep 2026), TechCrunch on Muse for Small Business (Sep 2026), Axios on Muse business usage (Sep 2026), Life Insurance International on Insurify's block (Sep 2026), Digital Insurance on insurers' block-or-allow decisions (Oct 2026), Agency Checklists on Insurify's agent access policy (Sep 2026), Trend Hunter on Kin (Oct 2026), ASIC RG 274, product design and distribution obligations