We were an insurance company before we were an AI company

How Notch went from underwriting cyber policies as a specialty MGA to building the AI agents that now run insurance operations for carriers, MGAs, brokers, and reinsurers.

Most AI vendors selling into insurance have never bound a policy, paid a claim, or been the one a broker calls when a quote is late. We have. Before Notch was an AI platform, Notch was an insurance company. That order is the whole story, and it’s why the product looks the way it does.

Notch is an AI operating system for regulated industries, built first for insurance. We build and run AI agents that handle underwriting, claims, policy servicing, and customer support for carriers, MGAs, brokers, and reinsurers. We didn’t arrive at that from the outside. We got there as the customer, by living the work.

We started as a specialty MGA in 2021

We founded Notch in 2021 as a managing general agent, and our line was deliberately narrow: specialty insurance for digital assets, a niche inside cyber. It’s a hard, technical class of risk, the kind most carriers were still learning how to price. We underwrote it with HSB (Hartford Steam Boiler), part of Munich Re, as our carrier partner. Early backing came from Lightspeed, Jibe, and Munich Re Ventures - investors who knew insurance and knew software.

Then we did the real work of an insurance company. We got licensed and launched across 45 states. We bound several thousand policies. We ran submissions, quoting, endorsements, renewals, and every service call that comes attached to a book of business. For two years, insurance operations weren’t a market we studied. They were our job.

The work no one could automate away

Running the MGA taught us where insurance actually breaks. It’s not the pricing model, and it’s not the capital. It’s the operational grind in the gaps between systems.

A submission lands as an email with a PDF and a spreadsheet, and someone rekeys it by hand. A broker calls to check status, and a rep opens three systems to answer one question. A renewal needs prep, so an underwriter spends two hours reassembling history that already exists somewhere in the stack. A loss gets reported, and the first thing that happens is a human typing it into a form.

Every one of those minutes was our margin, so we tried to buy our way out of it. We hired people and we bought point tools. Neither held up. The tools automated a narrow step and handed the seams back to a person. The headcount just grew with the book.

So we started building our own automations to run the business: intake, classification, follow-ups, the repetitive middle of servicing. It was internal software, built by us, for us, because nothing on the market understood an insurance workflow the way an operator does.

GenAI arrived, and the automation turned out to be the product

When large language models got good in 2022 and 2023, the tools we’d built for ourselves changed character. Work that used to need brittle rules and constant human review - reading a messy broker file, pulling the right clause out of a policy, drafting an accurate response - could now be done by an agent that genuinely understood the document.

That’s when we saw it: the thing we’d built to survive as an MGA was worth more than the MGA. The operational layer was the product.

It’s an uncomfortable realization after two years spent building a carrier-facing insurance business. But we had three things almost no one else had. We’d felt the problem from the inside. We had real workflows and real compliance logic already baked into working software. And we could see that carriers were about to come looking for exactly this. So we productized it, and Notch became a platform for building, governing, and running AI agents on the actual work of insurance.

Why we sold to insurers, and why we ran it on ourselves first

We looked hard at adjacent markets. Financial services, telecom, and retail all have the same shape of problem: high-volume, document-heavy, regulated customer operations. The pull was real. But insurance is where we had earned the right to be trusted, and it’s where the workflows we’d already solved translated directly.

Insurance also has a long sales cycle, and we knew it. Rather than wait around, we kept running Notch on our own operation and treated ourselves as customer zero. Every agent we shipped had to survive contact with a live book of business before we put it in front of a carrier. That discipline is still how we build.

What Notch is today

Notch runs on AI agents grouped into three kinds of work. Customer-facing agents handle inbound contact over chat, voice, and email. AI coworker agents sit next to your team and answer anything in plain language, then take the next action. Back-office agents fire automatically on business events and move work through the system of record.

Those agents serve the people who actually run insurance: adjusters working claims, service and support reps, underwriters, and the operations managers accountable for throughput. We now have a set of solutions built around those personas rather than one generic bot pointed at everything.

Underneath sits the part carriers care about most: governance. Notch enforces deterministic controls over what every agent can say, do, access, and execute, with a live audit trail on every interaction and compliance rules applied by jurisdiction. Across our deployments, agents have handled more than 20 million conversations at resolution rates north of 70 percent. And because the platform already understands insurance, carriers go live in weeks, not the 12 to 18 months an in-house build usually takes.

We go deep across the insurance stack

The reason we can move quickly inside a carrier is that we know the work function by function. We usually start where the volume and the pain are highest, in customer experience and support. Once those agents are running and the governance has earned trust, we extend the same platform into claims and underwriting. Claims means FNOL intake, triage, indexing, and loss summaries. Underwriting means submission intake, risk triage, renewal prep, and broker follow-up. Each step is the same operating system reaching further into the business, not a new project stood up from scratch.

That reach is the difference between a point tool and an operating system. A point tool automates one step and stops at the seam. Notch carries the insurance workflow and the compliance logic across every function it touches, so the second use case is a configuration, not a rebuild. That pattern is also how our investors came to see us: we’ve raised $45M to date, including a $30M Series A led by Headline in March 2026, alongside Illuminate Financial and our earlier backers.

Where we’re going

We opened a new office at the end of 2025 to support the next stage, run out of New York with our engineering core in Tel Aviv, and a team now around 80 people. The direction is wider on two axes. On lines of business, we’re moving past specialty and commercial into personal lines and health insurance. On who we serve, we’re extending past carriers into brokers, MGAs, and reinsurers.

The through-line hasn’t changed since 2021. We build for the person doing the work, because for two years that person was us. Insurance runs on operations, operations run on repetitive judgment, and that is exactly what a well-governed AI agent is good at. We just happen to have learned it the hard way first.

See how Notch agents handle claims intake, underwriting prep, and broker servicing.

‍