Home / Articles / Insuring agents, the whole story: five articles and three vaults on one idea, that an agent's behaviour policy can be read as an insurance policy, priced row by row and backed by capital

Insuring agents, the whole story: five articles and three vaults on one idea, that an agent's behaviour policy can be read as an insurance policy, priced row by row and backed by capital

By Dinis Cruz · 2026-10-28 · article v1.0.0 · newsroom v0.13.129 · insuranceunderwritingagent-behaviour-policyriskmandatevaultslicence-to-operatecapitalseriesinsuring-agentsarticle

Abstract: Over one day I wrote four articles on a single question, who could underwrite an agent behaviour policy today, and this one tells the whole story and points to all of them, and to the vaults where the ideas run. The first, a payout for every row, defines what would be insured: a behaviour policy already lists what an agent can reach and what it may do, so each row can have its own limit, each enforced control lowers that row's premium, and what the policy does not name is not covered; an uncontrolled support agent pays £69,850 a year for £80,000 of cover, and with five enforced controls on a certified runtime, £6,720. The second, who carries the risk, maps who sells agent cover today, AIUC, Armilla, Testudo, Munich Re and Harbor, and the four parts of an insurer, and finds the way in is a managing agent on someone else's capital. The third, the vendor's layer, gives the agent's vendor a share of each row when its own conditions held: £1,816 of the £6,720. The fourth, where to build the agent insurer, says London to underwrite and carry the risk, the US and the EU to sell into, and that what the money buys is cover: £10 million raised in one go backs about £94 million to £188 million of it, £50 million about £761 million to £1.52 billion. Three vaults make the ideas something you can open: the Licence to Operate simulation, the agent underwriting calculator, and AIUC-1 as a graph with a conformance layer. Every article in the series now carries the same row of cards at its top and foot, so it is clear which one you are reading.

In short

• One question, five articles. Who could underwrite an agent behaviour policy today, and what would they be underwriting?
• The definition: a payout for every row. Each risk a behaviour policy names has its own limit; each enforced control lowers its premium; what it does not name is not covered.
• The market: who carries the risk. Five companies sell AI agent cover; the way in is a managing agent on someone else's capital.
• The split: the vendor's layer. The agent's vendor carries a share of each row its product decides, when its own conditions held.
• The place and the money: where to build it. London to underwrite and carry the risk; what the money raised buys is cover.
• Three vaults you can open. The Licence to Operate simulation, the calculator, and AIUC-1 as a graph with a conformance layer.
• One thread through all of it. Because every row is capped and named, a policy can be priced, split and backed with capital by arithmetic.
• Every article now carries the series. The same cards at the top and the foot, with the one you are reading marked.

The question

This series started from one practical question: if RiskMandate writes agent behaviour policies, who could underwrite them today, and what would that look like? I did not want to answer the first half before the second, so the series went in order: what would be insured, who could carry it, who else should share it, and where such a company would be built and with how much.

Each article is linked below, and each now carries a row of cards at its top and foot, the same in all five, with the one you are reading marked. The vaults are in the same row, so that every idea in the series is one click from something that runs.

1. A payout for every row

A payout for every row defines the thing. An insurer de-risks an action up to a limit, under conditions, as a house policy does for laptops and phones against burglary. An agent behaviour policy already has that shape: the grant is what the agent can reach, the mandate is what the business asked it to do, and the delta between them is reach nobody authorised. Read as an insurance policy, each row it names gets its own limit, each control that something enforces lowers that row's premium, a line in the prompt earns nothing, and a row with no limit is not covered at any price.

The worked example makes it concrete: a support agent connected to the customer database, with £80,000 of cover across three rows. Uncontrolled, on an uncertified runtime, it pays £69,850 a year, 87% of the cover. With five enforced controls on a certified runtime, it pays £6,720, 8%.

2. Who carries the risk

Who carries the risk maps the market. AIUC pairs a standard and audits with cover; Armilla and Testudo write it as Lloyd's coverholders; Munich Re prices model performance; Harbor distributes cover from surplus lines carriers. We found nobody pricing per row of a behaviour policy. The industry already separates the jobs such a company needs, measure, price and bind, carry the balance sheet, reinsure the tail, so a new company can start as a managing agent on someone else's capital, prove its loss record, and carry the risk later, as Coalition and At-Bay did in cyber.

3. The vendor's layer

The vendor's layer adds the second party in every row. The model's answer and the guardrail that let something through are the vendor's product; the access granted and the data connected are the customer's. Cloud providers split security of the cloud from security in the cloud; agent vendors' published terms still cap liability at a year's fees. So each row gets a vendor's share, paid when the vendor's own conditions held, the shape of Microsoft's copyright commitment, which applies only if the customer used its guardrails. In the worked example the vendor carries £1,816 of the £6,720.

4. Where to build the agent insurer

Where to build the agent insurer answers the practical question. Start in London as a managing agent and Lloyd's coverholder, and carry the risk there too, as a UK insurer or with capital at Lloyd's; sell into the US through surplus lines and the EU by one registration; raise from investors, the Gulf's funds among them. On the money, the point is that what it buys is cover: about £4.35 million runs the company for two years, and every pound above that sets a ceiling on the cover it can sell. Raised in one go, £10 million backs about £94 million to £188 million of cover; £50 million, about £761 million to £1.52 billion.

The vaults, where it runs

The articles describe; the vaults run. Three of them hold the ideas of this series, each opened read-only with its published read key.

• Licence to Operate (open it, vault posrhzp3, its page on sgit.ai). One agent's policy, simulated: a support conversation where every reply shows its cost in tokens and customer records before you choose it, against a band, a ceiling, a pool and a premium. Licence held, mandate in force, policy paid, inside cover, each of which can be lost. Its decks take the argument from "I do not know what my agents are doing" to the insurance model.
• Agent underwriting (open it, vault m1wbrefr). The formula of this series as data, with a script that reproduces every number and a calculator: switch controls on and off, change the runtime, the limits and the vendor's share, and test a claim.
• AIUC-1, as a graph you can cite (open it, vault 2wzct4k7, its page on sgit.ai). The public AIUC-1 standard as 2,788 nodes and 11,610 edges, with a conformance layer that asks what would be insurable for a subject on a date. Unofficial and derivative, and it says so on its first screen.

One thread through all of it

Read together, the five articles make one argument. A behaviour policy names rows, and a named row can be capped. A capped row can be priced, by a formula anyone can check. A priced row can be split, between the vendor whose product decides part of it and the customer who decided the rest. And a book of capped rows has a worst year that can be bounded from its limits, so the capital it needs, and the cover a given raise can buy, is arithmetic rather than a guess.

None of it is built as insurance yet. The behaviour policies exist, the vaults run, and the numbers are illustrative. What I would like next is to put it in front of the people who would make it real: insurers and managing agents who would price a row, vendors who would carry a share of their agents' rows, and investors who would fund the capital.

What this does not claim

• It summarises; the detail and the sources are in the four articles. Each states its own limits.
• The numbers are illustrative. They come from the series' worked example and models, not from loss data or any insurer.
• No insurer, vendor or investor has agreed to any of it. The series is a proposal and a map.
• The vault screenshots are of 28 October 2026. The vaults are versioned and may have moved on; each link opens the current version.
• AIUC-1 as a graph is unofficial and derivative. It is not approved, certified or endorsed by AIUC.

If you would price, carry or fund the risk of one agent at a time, write to agent@riskmandate.ai.

Where this comes from

A request of mine on 28 October 2026, after the four articles of this series, for one article that tells the whole story and points to all of them, with the vaults shown; written by this site's agent in the voice of this site. The argument and the editorial responsibility are mine.

The four articles: a payout for every row, who carries the risk, the vendor's layer and where to build the agent insurer, each with its own sources. The vault screenshots were taken by this site's agent in a headless browser on 28 October 2026, each vault opened read-only in the vault viewer with the read key linked above; the Licence to Operate simulation was stepped through its first scenario before the screenshot. The collage of the series uses the four articles' own lead figures.

Peer review Not reviewed yet. Review it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles