A claim begins with something going wrong. A pipe bursts at two in the morning. A driver hears the hard little percussion of hail against a roof. Someone stands beside a damaged car and calls the number on an insurance card. The person on the other end has to listen, collect facts, check a policy, enter data, follow rules, and decide where the matter goes next. The voice is only the front door. Most of the work is behind it.
Amrish Singh has spent his career moving toward that back room. He trained as a software engineer, studied information systems at Carnegie Mellon, added an MBA at NYU Stern, and worked through enterprise software, consulting, and product leadership. In 2018 he joined Metromile, the pay-per-mile auto insurer, where he helped build its enterprise software business and saw both halves of insurance at once: the customer-facing operation and the systems underneath it.
By 2022, he had a thesis. Artificial intelligence could take routine work off insurance teams, but only if it understood the industry's limits and could operate the machinery where policies, claims, and customer records live. Singh co-founded Liberate with Ryan Eldridge and Jason St. Pierre. They did not start with a general assistant. They chose property and casualty insurance, then chose three lanes inside it - sales, service, and claims.
“The advantage of servicing only one industry, and within that servicing only three specific use cases, is that you can put a lot more guardrails in place.”Amrish Singh, 2025
01 / The seamThe work after hello
The company calls its customer-facing agent Nicole. She can answer a phone call, email, or text; take first notice of loss, commonly shortened to FNOL; gather information for a quote; send an ID card; or help update a policy. The name makes the product easy to picture. The harder part is intentionally less visible. Nicole must verify a caller, read the right record, structure what she hears, write an update into a core system, and hand the conversation to a licensed person when the request crosses a line.
This distinction sounds modest until one considers the enterprise AI demo. The model speaks fluently. Everyone claps. Afterward, an employee copies its output into another window and finishes the job. Singh's product argument begins where the applause stops. If AI only handles the conversation, the old labor remains. The software has to connect with systems such as Guidewire, Duck Creek, and agency management platforms, then complete the permitted action.
Marcus Ryu, the co-founder and former CEO of Guidewire, recognized the same seam. Battery Ventures led Liberate's $50 million Series B in October 2025, and Ryu joined the board. The round valued the three-year-old company at a reported $300 million post-money and brought total funding to $72 million. It also connected Singh with a builder who had spent decades making software for the same core insurance environment Liberate now wants its agents to navigate.
02 / The boundaryA workforce with a supervisor
Singh describes the product as an AI worker rather than a bot. The phrase is useful because it raises managerial questions. What is the job? What tools are available? How will performance be reviewed? What requires escalation? In a recent essay, he argued that a new AI agent should be treated like a new hire: give it defined responsibilities, access to the systems it needs, training, feedback, and measurement.
The analogy has limits, and Liberate tries to make those limits explicit. Its agents do not perform licensed insurance activities. They handle unlicensed tasks such as intake and data entry, while claims decisions and other judgment stay with professionals. The company also built a Supervisor tool to inspect interactions, flag anomalies, and send uncertain cases to people. Singh says the agent discloses that it is virtual. In one 2025 interview, he put the share of callers who initially ask for a person at roughly 18 to 22 percent and said comfort rises with experience.
This is where Singh's methodical streak shows. He talks about customer obsession and collaboration, but he returns repeatedly to process. Map the workflow. Connect the systems. Test the path. Measure the interaction. Preserve accountability. In insurance, a smooth voice cannot redeem a missed injury classification, a fabricated claim number, or a skipped compliance step. The industry's constraints force the AI product to become less theatrical and more operational.
That discipline has a commercial purpose. Liberate says a deployment can go live in six to eight weeks. Singh reported average sales gains of 15 percent and operating-cost reductions of 23 percent across customers in 2025. The company said one insurer cut hurricane claim-response time from 30 hours to 30 seconds. Those are company-reported outcomes rather than universal promises, but they reveal the scoreboard he prefers: completed work, response time, cost, and revenue.
03 / The operatorSerious systems, small tells
The public version of Singh is heavy on systems and light on mythology. His early research covered personalized ontologies for organizing search results and software for processing wearable-sensor data. Patent applications from his enterprise-software years dealt with turning business-system data into audio and executing transactions across heterogeneous sources. Long before today's agentic vocabulary, he was circling the question of how software finds, translates, and acts on scattered information.
There are warmer details around the edges. He speaks German professionally. He hosts the Insurance Experience podcast and often interviews colleagues and industry operators rather than placing himself at the center. During a winter podcast recorded in San Francisco, he noted the city's dependable 55-to-65-degree band and laughed that sweaters come out at 55. At an insurance conference, asked to predict the next Cricket World Cup winner, he chose India, pointing to its investment and facilities.
His recruiting style is similarly direct. In a 2025 post seeking people across customer success, engineering, product, and data science, Singh called hiring exceptional talent the hardest part of running a fast-growing company. He offered to personally shepherd applicants through the interview process. It is a small anecdote, but it fits the larger pattern: systems matter, and so does the handoff between them.
Singh also appears to enjoy the human density of the insurance business. He calls it the “full-contact sport of humans interacting with humans,” and says commercial relationships in the field frequently become friendships and mentorships. That affection complicates the simple automation story. His stated ambition is not an insurance company emptied of people. It is an operation where machines absorb repetitive transactions and people retain the moments that depend on trust, persuasion, empathy, or licensed judgment.
“I foresee a world five years out from now where the next $10 billion insurance agency will exist, which will have 20 employees and armies of AI agents working along with human agents.”Amrish Singh, 2026
04 / The betBuild for an ordinary Tuesday
Catastrophe season offers the obvious sales pitch for voice AI. A storm lands. Call volume multiplies. Hold times grow while frightened policyholders wait. Software that answers every call seems made for the surge. Singh argues that the better case is less dramatic: build for the other 350 days. A burst pipe at 2 a.m. is not a catastrophe event. It is an ordinary Tuesday, and immediate guidance may prevent a smaller loss from becoming a larger one.
The operational logic follows. If an AI agent is already handling everyday intake and service, surge capacity is not a separate system waiting in a box. It is a property of the system already working. This is the sort of observation that comes from crossing software and operations. Reliability is earned on routine volume; resilience appears when the routine suddenly becomes enormous.
By March 2026, Liberate said its platform was processing operations representing more than $100 billion in premium volume. In July, it announced that it had become an OpenAI Select Partner, with plans to apply newer frontier models across insurance sales, service, and claims. The company has widened from its original voice emphasis into email, SMS, digital interactions, and what it now calls a System of Action for Insurance.
The new label is grander, but the enduring idea is plain. Systems of record store the truth of the business. Singh wants a layer that can safely do something with it. A customer speaks in an ordinary sentence; the software turns it into structured work; the underlying system changes; a person enters when authority or judgment is needed. Each piece is familiar. The composition is the company.
Singh's career makes the bet feel less like a sudden response to an AI boom than the convergence of old interests. Software engineering supplied the mechanics. Enterprise product work supplied the patience for integration. Metromile supplied the insurance problem. The current models supplied a more natural interface and better reasoning. He put them together at the seam where a conversation becomes a transaction.
The caller at two in the morning will never see most of that architecture. Good operations tend to disappear in precisely this way. The phone is answered. The right questions arrive in the right order. The policy is found. The claim begins. If the situation needs a professional, the professional receives the context rather than a blank screen. For Singh, that quiet completion is the point: technology earning its place by finishing the routine work and preserving the human moment for when it is actually needed.