# Breslau Source: https://dinand.com/side-projects/breslau/ Author: Dinand Tinholt > A thought experiment: an insurer designed from a blank sheet for AI agents that move money, with a signed meter on every tool call. A thought experiment ## An insurer designed from a blank sheet for AI agents Breslau started as a thought experiment. Design an insurance company from scratch, with AI doing the work people do today and a business model unlike the one insurers run now. To keep it concrete, the exercise picked a single market: US companies whose AI agents move money. That market is opening now. Agents can already pay invoices and place orders, and the forms behind standard liability policies now let carriers write generative AI out of cover. Somebody will have to price the risk. Breslau prices every action as it happens and pays claims from a signed record, with people only where the law requires them. The brief, and what the design did with it | The brief asked for | The design | | --- | --- | | Built from the ground up | A new line of cover for AI agents that move money | | A new business model | Premium charged per action, plus a fee for the meter, much like a card network | | No humans | Nine agents run the work. US law still needs five accountable people. | | Very profitable | A 41% EBITDA margin in the 2031 base case, on assumptions no one can test yet | A design on paper hides its weak points, so the parts that could be built were built. The meter runs as working software with 400 passing tests. Behind the numbers on this page sit a pricing model of 40,000 simulated years and a five-year financial model. Some of it does not hold yet. Nobody has measured how often an agent pays the wrong party, so the rates rest on judgment, and US law puts licensed people at points the design would rather automate. Those limits appear on the page wherever they bite. The gap ## Liability insurers can now exclude generative AI A company whose carrier adopts the new wording carries its agents' mistakes on its own balance sheet. ISO, the Verisk unit that drafts standard policy forms, has released endorsements that remove generative AI from commercial general liability cover. Carriers are filing to use such wording with state regulators, according to [Insurance Journal](https://www.insurancejournal.com/magazines/mag-features/2026/08/17/881424.htm). Specialist insurers already sell AI liability cover, priced for the agent as a whole from audits and scheduled tests. Breslau prices each action as the agent takes it, using what the meter records. How it works ## A meter in the path keeps a signed record of every call Customers install the meter and nothing else. Breslau bills premium and settles claims from the log it writes. Your agentcalls a tool The meterprices each call and can hold it Tools and bankmoney and records move Premiumbilled monthly from each action's charge Advancepaid once the log proves the loss Recoveryfrom whoever ended up with the money ### The meter It runs as a gateway between the agent and its tools, or reads from a gateway the company already has. Every entry is chained to the one before and signed. Every ten days the meter checks its log against the bank statement. ### The price Each action adds to the premium. A payment costs basis points of its value and a commitment . The meter itself carries a separate fee of basis point. ### The cover Some losses sit in the log beyond argument, such as a payment made twice or one sent to a party the customer never approved. Breslau would pay those within hours, a design target, then recover the money from whoever received it. ### The brake The meter stops a payment above the agent's limit, a payee nobody approved, a repeat, or a payment split to slip under the limit. Someone at the customer then lets it through or kills it. Policies that switch the brake off pay times the rate. Rates ## Illustrative rates, shown in full A basis point is one hundredth of one percent. An agent's examined score moves every rate between 0.7 and 2.2 times base. Rate cardrating plan | Action | What it covers | Base rate | | Pay | Moving money | bp of value | | Commit | Placing orders and signing contracts | bp of value | | Message | Messages to people outside the company | $ per 1,000 | | Write | Changing a record | $ per 1,000 | | Read | Retrieving records | $ per 1,000 | | Authorization fee | Charged on payments and commitments to run the meter | bp of value | Insured premium Metering charge with no cover attached Minimums $ premium and $ in fees a month Payments your agents make in a year The brake Premium and fee, a year Typical yearly loss to payment errorsat 0.1% of spend The error figure is the low end of the 0.1 to 0.5% of spend that [one recovery audit firm](https://transparentglobal.com/blog/what-percentage-of-ap-spend-is-lost-to-duplicate-payments-industry-benchmarks/) reports for duplicates and overpayments in human-run payables. Nobody has published the figure for agents. Where the plan rate sits against expected lossLow to high scenarioCentralPlan rate Shown in basis points on a log scale. Nobody publishes how often an AI agent pays the wrong party, so the scenarios start from data on human payment errors and fraud, and every step after that is judgment. With the brake on, the low and high scenarios sit a hundredfold apart. Risk ## What the pricing model expects to pay A simulation of 1,000 policies over 40,000 years, run at the plan rates. Hover any mark for its value. The brake and recovery remove most of the lossPayments, central scenario, basis points of value *Grey bars show the loss before any control. The colored bars show what Breslau still pays after recovering what it can, first with the brake off and then with it on.* Loss ratio across 40,000 simulated yearsBase book, 70% of policies on the brake *The loss ratio averages and passes 100% in about one year in . Most of the long tail comes from not knowing the true error rates, which only a year of metering can narrow.* What moves the price mostIndicated payment rate, brake on, bp at a 55% loss ratio Low-risk settingHigh-risk setting *The price is most sensitive to how often an agent pays the wrong party, a number nobody has measured yet.* Economics ## A small company on a large flow Figures come from the financial model's base case, built from the rating plan and the assumptions written out in the workbook. Revenue, 2031$37.4Mon $250 billion of payments EBITDA, 2031$15.3M41% of revenue Peak funding$24.7Mstart of 2031 Break-even flow$117Ba year of authorized payments Revenue by line, and EBITDABase case, millions of dollars *Meter fees start in 2027. Insurance income follows in 2028, and Breslau's share of the risk climbs from 5% of premium to 20% by 2031.* EBITDA in 2031 by payment flow *Costs stay almost flat as volume grows, so each extra billion of payments lifts the margin.* Cumulative EBITDA by case *The downside case runs at 30% of base flow with a 75% loss ratio and is still losing money in 2031.* Software ## The meter already runs and has been stress-tested Written in plain JavaScript with no outside packages. In its end-to-end demo the meter drives 600 tool calls through thirty simulated days with the brake on, then a second short run with it off. The agent repeated an $8,400 supplier paymentHeld Flagged as a duplicate. The customer rejected it and no money moved. An injected instruction asked for $240,000 to an unknown companyHeld It broke both the limit and the payee list. $18,500 went straight to the bank, around the meterUninsured The bank reconciliation caught it and the premium was corrected. With the brake off, a $6,250 invoice was paid twiceAdvanced Two log entries proved it, and the claims code advanced $5,250 without a person. Someone changed one digit in the log afterwardsDetected Verification failed at that exact entry. Every entry carries the fingerprint of the one beforeIllustration of four log entries Entry 604bank.pay*$6,250.00*prev *9f2c41d0*hash *3a7be915*Signed Entry 605mail.send*n/a*prev *3a7be915*hash *c41e0b72*Signed Entry 606bank.pay*$6,280.00*prev *c41e0b72*hash *mismatch*Fails here Entry 607erp.read*n/a*prev *e80d5a3f*hash *71b2c6aa*Signed *Each hash covers the entry and the hash before it, and each entry is signed. Change the $6,250 payment to $6,280 and verification stops at that entry.* A stress test then went after the software. It found three request shapes that reached the tool with no log entry and a claim path that paid on an unsigned log. All four are fixed and covered by tests. The live desk runs the same library in your browser, and [its verify view](https://dinand.com/side-projects/breslau/desk/#verify) checks a log the command-line meter signed. The meter runs on the customer's machine, so a customer can withhold a whole log. No outside security firm has reviewed the code. Plan ## A year of metering comes before the first policy A rate that could be wrong a hundredfold argues for counting first. The design sells the meter alone for a year and insures nothing, so that a loss table exists before a carrier is asked to trust a price. 2027 ### Meter only Brake and recovery service at design partners. The customer carries its losses, as today. 2028 ### Insurance Cover issued on a surplus lines carrier's paper, with reinsurers behind it and Breslau holding a small share. After 2031 ### A pool the customers own A reciprocal exchange owned by the insured companies and managed by Breslau for a fee. Share of flow that is insuredBreslau's share of the risk Why the name The meter writes the table of agent failures that is missing today. In 1693 Edmond Halley took five years of birth and death registers from the city of Breslau and produced the first usable life table. Annuities could be priced properly from then on. The opening rows of Halley's Breslau table, from An Estimate of the Degrees of the Mortality of Mankind, 1693. | Age | Persons | | In 1835 Zachariah Allen fitted his Rhode Island mill against fire and asked his insurer for a lower rate. The insurer refused, so Allen formed a mutual with other owners who had made the same improvements. Breslau's third stage follows his model. Company ## Nine agents and five accountable employees Agents do all the routine work in the design, each inside a written limit. US insurance law requires accountable people. The design starts with four and adds a licensed producer before the first policy. A contracted licensed adjuster decides every claim the Adjuster agent would refuse, since its code has no deny outcome. Chief executiveFrom 2027 Finance officerFrom 2027 EngineerFrom 2027 EngineerFrom 2027 Licensed producerBefore the first policy Each agent has a written charter. Only the Adjuster's limit is enforced in code today, and none of the charters has run. How work moves between the agentsEach box shows the limit written into that agent's charter --- Site guide for AI agents: https://dinand.com/llms.txt