Three weeks into my job at Newfront, I noticed a constant line at the commercial printer. We had spent tens of millions of dollars building automation software in-house, so I asked the CTO: what's the deal with the printer? Brokers, he told me, often had to print and mail documents to the carrier. Scan-to-fax was the upgrade.
Follow that fax. It's an ACORD form — eight to twelve pages, usually filled out by hand, full of questions the broker doesn't know the answers to, so the answers are often made up. It lands in a carrier inbox. An underwriting assistant opens it, swivels to a second screen — a green-screen application from the nineties, built on COBOL, running on a mainframe that crashes constantly — and transcribes the form by hand. A junior underwriter reviews it for completeness, finds ten things missing, and sends it back to the broker.
Two weeks, on average, from application to bindable quote. For a one-to-five-thousand-dollar policy.
Nobody was winning
The easy read on that story is "quoting is slow." That's not the problem. The problem is that nobody in the chain makes money.
You don't open a bakery to buy insurance — you want your policy so you can start baking. The broker has to service that bakery and loses money on every small policy they place. The carrier's expense ratio on a $3,000 policy is the same machinery they'd run on a $100,000 account, so most of them lose money too. History is littered with insurers that got into small commercial and got back out — not because it's bad business (it's low-volatility, reasonably low-loss), but because the expenses eat it alive. There are thirty-five million small businesses in America, and the industry's honest answer to most of them has been: you're not worth quoting properly.
That's the opportunity. Not "make the form faster." Take the expense out of the system entirely.
Stop asking for the roof's age
When we started rebuilding quoting, we went back to first principles: what are we actually solving for? Reliable, accurate quotes, quickly, that we'd stand behind.
State it that way and you notice something. Most of what the industry treats as deterministic requirements were only ever proxies. Roof age is a proxy for roof quality. People fixate on whether the roof is nine years and 364 days old or ten years — but that extra day doesn't deteriorate anything. Today, machine vision can tell you whether the roof is actually in good shape. You never needed the age. You needed the condition, and the age was the best available stand-in forty years ago.
The whole application is like this. An eight-page form isn't underwriting — it's a fact-gathering exercise imposed on the person least equipped to do it. Square footage, year built, construction type: these aren't judgments, they're data. They're out there. They should be gettable without asking anyone.
Underneath it all: insurance is a probabilistic problem, and it got force-fit into deterministic software because deterministic software was all that existed. AI is a probabilistic solution. For the first time, the shape of the tool matches the shape of the problem. Software couldn't do what AI can do for this industry.
I know the distribution side of this cold. The architecture is Eric's — he built it, so he should explain it.
What we built instead
Eric here. When Graham first described the manual pipeline to me, the engineer in me couldn't get past it. Emails printed out, transcribed onto other printed forms, scanned back into electronic systems. Why are humans doing translation work between machines?
Aimee was our first answer. She started narrow: our customer-service logs showed one question dominating call volume — "do you cover this?" Answering it used to mean a rep frantically pinging the underwriting team while a broker held. Appetite for hundreds of business classes, with subclasses and coverage limitations varying by state and county, is more logic than any human keeps in their head. So we pointed Aimee at our appetite guides and let her answer directly.
Our design philosophy mattered more than the tech. We didn't start with a human approving every response — we don't have the headcount for that, and it defeats the purpose. We made Aimee smart enough to handle the job, gave her scope and context, and monitored her decisions, tightening her logic when she missed. Guardrails exist; Graham still stress-tests her about once a week to make sure she hasn't gone rogue. But the posture is trust-and-verify, not gate-everything.
Two things happened that we didn't fully plan for.
First, within a couple of weeks, our own employees started using her. Aimee lives in Slack, and she became the internal gut-check for questions we thought we already knew the answers to.
Second — and this is the one in the title — brokers started thanking her. If you read the chat logs, they say "that's so great, thank you so much, how do I get this going?" They ask her to start the quote. They talk to her the way you talk to a sharp, fast colleague. Reading those transcripts, we genuinely don't think most of them realize she's an AI.
What we're doing about that
But notice what the broker behavior actually is: it's demand. Nobody pushes a conversation toward "how do I bind this?" with a system they distrust. Aimee currently stops there — she says "reach out to our team," which is honestly a little deflating after how well the conversation went. The market has already told us what to build next.
Two questions, three thousand data points
The next layer is Magic Dust — our data layer. The idea we call two-question quoting: give us a business name and an address, and we pull everything else. Some of it bought via API, some gathered from public sources, some inferred. A data-science agent, Prometheus, decides which data best fits and fills the application with no broker input beyond those two answers.
We're connecting the dots so now you can start a conversation with Aimee while Magic Dust pulls the data, so by the time you ask for a quote, we've done all the work.
Here's why the data layer matters more than the convenience. We've tried to measure broker-supplied data accuracy — our estimate is around sixty percent. That's not a knock on brokers; their job is representing the policyholder and moving fast, not data hygiene. Sophisticated carriers do better by pulling more data, but manually — an underwriter checking maybe forty data points, logging into a different system by hand for twenty-five of them. We look at roughly three thousand data points per risk, instantly. If you asked Graham to review three thousand data points, he'd tell you himself: he couldn't remember the name of the business by the end. AI doesn't lose focus at row 2,900.
The incumbents' response is usually "but we have decades of data." Decades of data is decades of compounding the sixty percent. Businesses change, so even accurate data drifts — and a lot of it wasn't accurate to begin with. If the data is flawed, the moat is a liability, and building AI on top of it makes the problem worse, not better.
The part that only works because of MCP
One agent trying to do everything doesn't work — context bloats, consistency drifts, hallucination risk climbs. So we run specialized agents: Aimee facing the broker, Magic Dust and Prometheus on data, Athena reasoning about the risk (she gets her own post), Quill on the back end. The problem that creates is connection: how do agents talk to each other, and to systems still sitting behind traditional APIs?
MCP is the translation layer — the language for talking between things. The contrast with a traditional API is the whole point. When something major changes behind an API, you update the API, and then everyone downstream updates too — a big, brittle chain. This is exactly the failure mode Graham watched at Newfront, where a single carrier field change would shatter the pipeline. MCP was built for a world where the things behind the tools evolve constantly. A tool called "get quote" can be a database today and an agent next quarter; the caller never knows or cares. We can swap implementations, retrain agents, add data sources — and the orchestra keeps playing.
And it points outward, not just inward. The same layer that lets our agents talk to each other can expose our entire company as a tool. Someone chatting with Claude or ChatGPT about starting a business mentions they'll need insurance — and through MCP, that conversation can reach us directly. We're not building point solutions for pieces of a quoting flow. We're building an insurance company that runs as a set of orchestrated agents, reachable from wherever the conversation is already happening.
The forms were never the product
Here's the part I find funny about all of this. The end state — call us, text us, describe your business in a sentence, get a live quote — isn't futuristic. It's how buying insurance worked before deterministic software forced everyone into forms. The forms were never the product. They were a compromise with the tools of the time, and the tools have changed.
We're going back to a world where people could do it any way they wanted. The forms just don't come with us.
