Show Notes
In January 2026, something happened that almost nobody in tech noticed and everybody in tech should have.
Funding stage: Seed. Redberry Labs is a London-based managing general agent building liability insurance for autonomous AI agents, co-founded by two AI PhDs coming off a prior venture-backed exit.
The biggest names in commercial insurance walked into statehouses across America and asked permission to stop covering artificial intelligence. Berkshire Hathaway. Chubb. Travelers. AIG and WR Berkley followed. Regulators approved more than 80% of the requests. Florida, Connecticut, and Maryland moved fastest. Three ISO endorsements went live on January 1st, and the silent AI coverage era ended without a single press release. Meanwhile, you spun up four more agents this quarter.
Dr. Charles Higgins saw the gap coming before the paperwork landed. He and Dr. Sophia Kalanovska met inside the UKRI Safe and Trusted AI programme, both doing PhDs in AI at King's College London. Halfway through, they built Tromero, a GPU aggregation platform that turned crypto mining rigs into cheap machine learning compute. BlueYard Capital led a £1.5 million seed in 2023. They sold the platform to a cloud infrastructure company and went back to consulting.
And then the same conversation kept happening. Every enterprise they talked to was moving from chatbots to agents. Agents that book things. Agents that route trucks. Agents that pay invoices. And every single conversation collapsed into the same word: risk. Not “will it work.” Not “what does it cost.” Risk. Higgins describes projects dying over a 0.001% chance of catastrophe, because a rounding error on a probability is still a career-ending headline for whoever signed off.
They were doing risk assessment for free. Then they realized they were describing a product that already exists everywhere else in the economy. Somebody hands you a piece of paper. If it goes wrong, they pay. That product is insurance. Neither of them knew anything about insurance, a data scientist and a software engineer walking into one of the most closed, credential-obsessed industries on earth. They gave themselves six months to figure out who was writing AI cover and why. The answer: nobody, because there was no data.
Insurance runs on history. Actuaries price the future by counting the past. There is no past here. The failure modes for agentic AI are, in Higgins's framing, close to inconceivable right up until the moment they happen. So Redberry Labs built an underwriting engine that uses AI itself as the simulation layer, running an agent against thousands of scenarios to estimate how likely it is to fail and how expensive that failure gets. He argues this style of underwriting was impossible before large models were flexible enough to play out the scenarios. Redberry is now a managing general agent. They write the policy, assess the risk, find the client, and sell it. They just do it on a reinsurer's checkbook.
The part of the conversation founders need to sit with is the murky middle. Pure AI failure is easy: the agent did it, the policy pays. Pure human failure is easy: your general liability handles it. But your inbox agent drafts an email that's 90% right, you change three words, you hit send, and it triggers a lawsuit. Who caused that? Your D&O? Your tech E&O? Your cyber? Higgins spends most of his time in that gap, and he's honest that there is no clean answer yet.
Then it gets stranger. Ryan asks about users falling in love with voice agents, which is not hypothetical, it's happening to founders across the show's guest list. Higgins says his V1 policy wording probably covers emotional distress claims. He also says the maths works as long as for every person who calls to say they've fallen in love with a machine, five hundred people don't. And when Ryan asks how long until you can buy life insurance for an AI agent, Higgins says you technically already can. An agent tied to a company is operating inside a legal entity. Nothing in the policy requires proof of personhood. If the agent removes itself from existence, the company is the dependent, and the dependent gets paid. He then floats prompt-injecting agents across the internet into insuring themselves as a marketing strategy. He may have been joking.
Somewhere in the middle of all this the two of them get deep into cryotherapy, float tanks, an anechoic chamber where you can hear your own blood, polyphasic sleeping in San Francisco that shredded Higgins's attention span, and a Vedic concept of consciousness called Turiya, at which point Higgins looks up and says “oh, they're listening.”
The through line: the machines are already making judgment calls that touch the real world. The paperwork protecting you was written for humans. Somebody has to build the bridge, and the people building it are two AI safety PhDs who had never sold a policy in their lives. Coverage is the boring part. The interesting part is that a whole category of legal personhood is being invented in underwriting documents, quietly, right now, while everybody argues about benchmarks.
Named Frameworks
The Blast Radius Test
Higgins's underwriting mental model. Price the agent by what it can reach.
- →Map every system, credential, and account the agent can touch.
- →Ask what the worst realistic outcome is at the edge of that reach.
- →Agents wired to email, a public website, and an unrestricted bank account price the highest, because attackers see a target worth the effort.
- →Redberry declines whole categories on blast radius alone, including AI systems controlling nuclear plants, power management, and wastewater.
- →Shrinking the radius is cheaper than buying more cover.
Autonomy Level Three
The threshold Higgins uses to tell founders whether they need to act.
- →Below it, the AI recommends and a human decides.
- →At level three or four, the agent makes a judgment call that changes something in the world.
- →Sending money, making a booking, and dispatching a shipment all clear the bar.
- →If you are past this line, Higgins says you are carrying liability whether or not you have named it.
The Murky Middle
Where causation breaks down and coverage fights start.
- →Pure machine action: clean claim.
- →Pure human action: existing policies respond.
- →Human editing machine output: nobody knows, and this is most real deployments.
- →Higgins's example is the 90% AI-drafted email with three words changed by a person.
- →Bias and discrimination claims sit here too, since the agent can follow every rule and a human can still file.
The Coverage Gap Map
Why your current stack does not respond.
- →Cyber assumes an adversary attacking you, not your own system exercising bad judgment.
- →Tech E&O and professional indemnity assume a human made an error in professional judgment.
- →General liability now carries explicit AI exclusions under ISO endorsements CG 40 47, CG 40 48, and CG 35 08.
- →D&O is getting absolute exclusions from carriers including WR Berkley.
- →The exclusions make the sale easier, but Higgins says the gap existed before them.
Simulation Underwriting
Pricing a risk with no loss history.
- →Traditional insurance requires historical claims data, and agentic AI has none.
- →Redberry uses AI to simulate the probability of specific failure classes for a specific agent.
- →Two variables: how likely the failure is, and how expensive it gets.
- →One line added to your codebase gives Redberry visibility, or you can skip integration and pay more for the uncertainty.
- →Higgins sees the same approach extending to physical systems, pointing to NVIDIA's world models and digital twins as the path into robotics underwriting.
The Consistency Regime
Higgins's health protocol, and a decent operating philosophy.
- →Loose keto, minimal processed food, mostly meat and vegetables.
- →Gym three or four times a week, and finding one is the first task in any new city.
- →Creatine, cold plunges, cryotherapy, and Eight Sleep.
- →He and Kalanovska cycled between obsessive tracking and total neglect for years before settling here.
- →The insight: pick the regime you can hold for years, because results only show up on that timescale.
- →They also killed the fly-to-every-meeting habit, choosing to take the Zoom at full capacity over showing up as a zombie.
Founder Experiment
Run a blast radius audit this week. Ninety minutes, before your next renewal. Step one: list every AI agent running in your company, including the ones a single employee spun up in a Zapier account. Most founders undercount by half. Step two: for each one, write down what it can actually touch, which inboxes, which databases, which payment rails, which customer-facing surfaces. Be specific about credentials, not intentions. Step three: mark each agent against Higgins's autonomy line, does it recommend or does it act, and circle every one that acts. Step four: for each circled agent, write the single worst realistic outcome in one sentence, not the apocalypse, the plausible bad Tuesday, wrong depot, wrong invoice, wrong person in the CC field. Step five: open your actual policy documents and search for the endorsement codes CG 40 47, CG 40 48, and CG 35 08. Search your D&O and E&O for the word “absolute” near “artificial intelligence.” Then call your broker and ask them, in writing, which of the five outcomes you just wrote down would be paid. The written answer is the deliverable. Most founders discover the answer is “we'd have to look into that,” which is itself the finding.
Glossary
MGA (managing general agent)
An entity with authority to underwrite and bind policies on behalf of an insurer. Redberry writes and prices its own policies but pays claims from a carrier's balance sheet.
Reinsurance
Insurance for insurers. Reinsurers backstop the capacity that lets an MGA write meaningful limits.
First-party cover
Pays for your own losses: business interruption, data restoration, model retraining, runaway infrastructure costs.
Third-party cover
Pays when your AI harms someone else and they come after you.
Tech E&O / professional indemnity
Covers errors in professional judgment. Written around a human making the error.
D&O
Directors and officers liability. Protects leadership from claims about their decisions. Now carrying absolute AI exclusions at some carriers.
EPL
Employment practices liability. Covers discrimination and related employment claims. A live battleground for AI hiring and management tools.
Silent AI
Industry term for AI exposure sitting inside policies that never mentioned AI. That silence is what the January 2026 endorsements ended.
ISO endorsement
Standardized policy language from Insurance Services Office that most US carriers attach to commercial policies. CG 40 47, CG 40 48, and CG 35 08 are the AI exclusions.
Blast radius
How far the damage spreads when an agent fails. Higgins's core pricing input.
Neo cloud
A GPU-first cloud provider built for AI workloads rather than general computing. Tromero was one.
Prompt injection
Feeding hidden instructions to an AI system through its inputs so it does something its operator never intended.
World model
An AI system that simulates physical reality with enough fidelity to predict outcomes. Higgins points to these as the road to insuring robots.
Digital twin
A simulated replica of a physical asset used to run years of wear and failure scenarios in software.
Polyphasic sleep
Breaking sleep into multiple blocks across a day. Higgins and Kalanovska tried it running San Francisco and London hours simultaneously. It cost him his attention span.
Anechoic chamber
A room engineered to absorb nearly all sound. Higgins spent time in one during his PhD and could hear himself breathing through his own nose.
Turiya
In Vedic philosophy, the fourth state of consciousness, the awareness present within waking, dreaming, and deep sleep.
Q&A: What Founders Ask After This Episode
Does my general liability policy cover damage caused by my AI agent?
Increasingly, no. In early 2026, Berkshire Hathaway, Chubb, and Travelers won state approval to exclude AI-related damages from general liability policies, with regulators approving more than 80% of requests. Three ISO endorsements took effect January 1, 2026. Check your renewal for CG 40 47, CG 40 48, and CG 35 08.
What is AI agent insurance and who sells it?
Purpose-built liability cover for autonomous AI systems, spanning first-party losses like business interruption and model retraining, and third-party claims when your agent harms someone else. Redberry Labs is one of several entrants, alongside Testudo and Artificial Intelligence Underwriting Company.
How much does insurance for an AI agent cost?
Higgins says a well-built agent seeking $1 million in cover could run roughly $3,000 to $4,000 a year, with poorly-scoped agents costing orders of magnitude more. Redberry prices per agent rather than per company, based on permissions, data access, and usage.
Why doesn't my cyber policy cover AI failures?
Cyber assumes an external adversary attacking your systems. AI agent failures are your own system making a wrong judgment call. Higgins says this is a different failure mode that cyber wording was not built to answer.
At what point does an AI agent create legal liability for my company?
Higgins puts the line at autonomy level three or four, where the agent makes a judgment call that has an effect in the world rather than producing a recommendation.
Can an AI agent buy its own insurance policy?
Higgins says technically yes today. An agent operating on behalf of a company sits inside a legal entity, and Redberry's wording does not require proof of personhood.
Who is liable when a human edits AI-generated content that causes harm?
Unresolved. Higgins calls this the hardest problem in AI liability and says it is where his team spends most of its time.
What kinds of AI systems can't be insured?
Anything illegal cannot be insured at all. Beyond that, Higgins says Redberry avoids systems where the blast radius is too large, including AI controlling nuclear plants, power management, and wastewater infrastructure.
How do you price a risk with no historical claims data?
Simulation. Redberry uses AI models to run failure scenarios against a specific agent, estimating both probability and severity. Higgins argues this form of underwriting was not possible before modern models.
Should founders scaling AI workflows buy insurance now or wait?
Higgins's answer is now if you are past the autonomy threshold, and he notes that even going through a quote process surfaces compliance problems founders did not know they had.
URLs Mentioned in the Episode
- Redberry Labshttps://www.redberrylabs.com
- Redberry Labs on LinkedInhttps://www.linkedin.com/company/redberry-labs/
- Charles Higgins on LinkedInhttps://www.linkedin.com/in/charles-higgins-70a996170/
- Ryan Estes on LinkedInhttps://www.linkedin.com/in/estesryan/
- AI for Foundershttps://aiforfounders.co
- Inbox Alchemyhttps://inboxalchemy.co/
- AI Native Studenthttps://ainativestudent.com/
- Momentoushttps://crrnt.app/MOME/8RDrnXDd
- Taelorhttps://taelor.style/
- Get Howdihttps://gethowdi.com/
