Insure Your Agent Operator Edition

Future Proof Intelligence · Published 29 May 2026 · Rewritten 17 August 2026 · 12 minute read

What AI insurance costs a small business. What is actually published, and what is not.

This page used to carry euro ranges for AI cover: two to eight thousand for a cyber endorsement, five to twenty for technology errors and omissions, fifteen to thirty for a standalone policy. Those numbers were estimates, and we could not source a single one of them to a carrier. On 17 August 2026 we took them out. What follows is the honest version: the small number of figures anyone in this market actually publishes, the five things that determine what you will be quoted, and how to turn that into a real number in about three weeks.

Key takeaways

  • No carrier or MGA writing AI liability publishes a premium. Every placement is priced on an individual submission. Any euro-per-year figure you find, on this site before today or anywhere else, is somebody's estimate wearing the clothes of a fact.
  • What is published is limits. HSB: standard limits of USD 25,000 or USD 50,000, USD 500 deductible. Armilla: up to USD 25 million per organisation. Testudo: USD 1 million to USD 10 million. Munich Re with Mosaic, for aiSure: EUR, USD or CAD 15 million of initial capacity. That spread is the real story about this market.
  • The cheapest route into AI cover is almost never a standalone AI policy. It is an endorsement on a policy you already buy. That is where the affirmative wordings are appearing.
  • Sector determines severity and severity determines price. The same agent doing the same job costs a different amount to insure in a legal practice than in a logistics firm, because the claim it eventually produces is a different size.
  • The one lever genuinely in your hands is the scope of autonomous action. What your agent may do without a human, written down, with a threshold, is the document an underwriter reads first.
  • We do not publish a percentage for how much governance saves you. Nobody publishes one. What we can say is what specialist underwriters ask for, and that is in the submission section below.

Why you cannot look this up

Commercial insurance pricing is not usually secret. Employers' liability, commercial property and motor fleet all have published rate guides, comparison tools and decades of claims data behind them. You can get a defensible estimate before you speak to anyone.

AI liability has none of that. The oldest dedicated products in the market are barely two years old. Claims experience is thin enough that underwriters are pricing from adjacent lines and from judgement rather than from a triangle. The products themselves are still moving: a policy called AI liability in 2026 covers a materially different set of things from a policy with the same name in 2024. And the market is small enough that every carrier treats its pricing as competitive information.

The result is that there is no published rate, and there is no honest way to synthesise one. What there is instead is a set of published limits, a clear picture of what underwriters ask about, and a process that will get you a real number faster than searching for a fake one.

The figures that are actually published

These are limits and capacity, not prices. Each is read from the provider's own material. Where a provider publishes nothing, that is said.

Provider What it publishes Premium published? Who it is for
HSB (Munich Re) Standard limits USD 25,000 or USD 50,000, USD 500 deductible, higher limits available No Small and mid-sized businesses in the US, added to a partner carrier's policy, not sold direct
Armilla Standalone AI liability up to USD 25 million per organisation No Enterprises with complex AI deployments; Lloyd's coverholder
Testudo Published limits from USD 1 million up to USD 10 million No Middle-market to large enterprises; Lloyd's coverholder
Munich Re aiSure, via Mosaic EUR, USD or CAD 15 million initial capacity No AI developers and vendors, not AI users; settles on measurable performance data
Coalition Affirmative AI endorsement on cyber, two limbs only No Existing cyber buyers; does not respond to loss from a wrong AI output
Embroker AI coverage endorsement, auto-attached to eligible tech E&O and cyber quotes No Technology businesses already buying tech E&O or cyber

Read the first row against the second. HSB's product, the one built for a business of your size, carries a standard limit of twenty five or fifty thousand dollars. Armilla's, built for enterprises, goes to twenty five million. That is a factor of five hundred to a thousand between the SME product and the enterprise product, and it is the most honest single fact about where this market currently is. Cover for small operators exists, it is real, and it is small.

The practical consequence: if your realistic worst case is a customer complaint, a refund, a takedown and some legal time, an SME-scale AI endorsement is proportionate. If your realistic worst case is a professional negligence claim from a client who relied on your agent's output, a USD 50,000 limit is not the instrument, and you should be looking at your professional indemnity wording rather than at a new AI product.

The five things that determine your quote

You cannot look up the price. You can work out, fairly accurately, which direction each of these pushes you, and that is most of what a first broker conversation establishes anyway.

1. What a mistake costs in your sector. This is the largest single driver, and it is about consequences rather than technology. The same drafting assistant is a different risk in a law firm than in a landscaping business, because the claim it eventually produces is a different size. Healthcare, legal, financial advice and recruitment sit at the severe end: the error reaches a person's health, money, case or job. Retail, logistics and hospitality sit at the modest end, where an error usually produces a refund and an apology. Nothing about your governance changes which end you are at.

2. What the agent may do without a human. This is the factor most within your control and the one underwriters probe hardest. An agent that drafts and a human sends is a different risk from an agent that sends. An agent that recommends a refund is a different risk from an agent that issues one. The question is not whether you have a human in the loop in principle, it is what the agent is authorised to complete on its own, and at what value.

3. How many decisions it makes. Volume matters not because more interactions means more small errors, but because a systematic error propagates across every interaction before anyone notices. Fifty queries a day gives you time to catch a bad pattern. Fifty thousand does not. Underwriters look at throughput alongside the consequence per interaction, which is why a high-volume scheduling assistant and a low-volume advisory agent can price similarly.

4. Whether you can describe the system at all. An underwriter who cannot characterise a risk prices it conservatively, because conservatism is the only defence against an unknown. This is the difference between a submission that says "we use AI in customer service" and one that says what the agent does, what it is connected to, what it is allowed to do, what accuracy you monitor, and how you would know if it started going wrong. The second is not a better risk than the first. It is the same risk, priced instead of guessed at.

5. The limit you ask for. Premium does not scale linearly with limit, because the probability of a claim reaching the top layer falls. But in a market this young, capacity rather than appetite sets the ceiling, and the published capacity figures above are the constraint. Asking for a limit near the top of what the market currently supports costs disproportionately more than asking for one comfortably inside it.

What your existing policies do, before you buy anything new

Most operators who ask what AI insurance costs have not yet established what they already have. That is the cheaper question and it should come first.

Commercial general liability covers bodily injury and property damage. Most AI mistakes cause economic loss: a customer acted on wrong information and lost money. CGL was never the policy for that, with or without an AI exclusion. Separately, ISO has published generative AI exclusion endorsements that a CGL carrier may attach, which makes explicit what was already largely true.

Cyber is moving toward AI rather than away from it, which is the opposite of what most coverage summaries suggest. Coalition added an affirmative AI endorsement on 26 March 2024, and it has exactly two limbs: it extends the security failure and data breach cover to an AI security event, and it extends funds transfer fraud to a fraudulent instruction delivered by deepfake or other AI. Neither limb reaches a loss caused by your AI simply producing an incorrect answer. Embroker attaches an AI coverage endorsement to eligible technology E&O and cyber quotes from 5 August 2025. Vouch publishes affirmative AI cover naming hallucinations, algorithmic bias and certain intellectual property disputes. Read which of those shapes your endorsement is before you assume it reaches your exposure.

Professional indemnity or errors and omissions is the policy most likely to respond to the claim you are actually worried about. The live question is whether its professional services definition reaches an output that no named professional reviewed before it went to the client. That is worth a written answer from your broker rather than an assumption. See the professional indemnity guide for the specific wording to check.

How to get a real number in three weeks

This is the part that replaces the ranges. It produces an actual quote rather than an estimate, and it takes less effort than most operators expect.

  1. Week one: write the two documents. A one-page description of each AI system in production, covering what it does, what data and systems it touches, what it produces and who receives it. And a one-page authority note: what the agent may do without human approval, at what value a human must approve first, and who that human is. These are the two documents every specialist underwriter asks for, and neither needs a lawyer.
  2. Week one, same day: ask your broker the coverage question in writing. Name the agent, describe what it does, and ask how each of your current policies would respond to a claim arising from its output, with any exclusion identified by endorsement. Written answers only. This costs nothing and frequently ends the exercise, either by revealing cover you did not know you had or by revealing a gap precisely enough to price.
  3. Week two: ask for indications on the endorsement route first. An AI endorsement on your existing cyber or technology E&O programme is faster to place and cheaper at the margin than a standalone policy. Ask what the sublimit is, in money, and what categories of AI loss it names. A sublimit you cannot get in writing is not a sublimit.
  4. Week two: get one standalone indication as a benchmark. Even if you do not intend to buy it, one specialist indication tells you what the market thinks your risk is worth, which is the only pricing information that is actually about you. Your broker submits the two documents from week one.
  5. Week three: compare on what responds, not on price. Put the endorsement sublimit and the standalone limit next to your own worst realistic case. If neither reaches it, the honest answer may be that the market does not yet sell what you need at your size, and the right response is to reduce the agent's autonomous authority rather than to buy a limit that will not reach.

Start at least 90 days before renewal if you want the standalone market to be a genuine option. Specialist placement is slower than standard commercial lines, the underwriters ask for more, and there are fewer of them.

Frequently asked questions

How much does AI insurance cost for a small business?

Nobody publishes a rate card, and we will not invent one. Every carrier and MGA writing AI liability today prices by individual submission. What is published is limits, not premiums: HSB gives standard limits of USD 25,000 or USD 50,000 with a USD 500 deductible; Armilla writes up to USD 25 million per organisation; Testudo writes USD 1 million to USD 10 million; Munich Re and Mosaic launched aiSure with EUR, USD or CAD 15 million of initial capacity. None of the four publishes a premium. Any euro range quoted for AI insurance, including the ranges this page carried before 17 August 2026, is an estimate presented as a fact.

What actually determines what I will be quoted?

Five things, in roughly this order of weight: what a mistake costs in your sector; what your agent is allowed to do without a human; how many decisions it makes; whether you can describe the system precisely enough for it to be assessed rather than guessed at; and the limit you ask for. The second is the one most within your control.

Does my existing business insurance cover AI mistakes?

Probably not, but be careful about why. Commercial general liability covers bodily injury and property damage, and most AI mistakes cause economic loss, so it was never the answer. Cyber is moving toward AI rather than away from it: Coalition's affirmative AI endorsement of 26 March 2024 has exactly two limbs: an AI security event, and funds transfer fraud by deepfake or other AI. Neither limb reaches a loss caused by your AI simply being incorrect. Professional indemnity is the policy most likely to respond, and the question there is whether its professional services definition reaches an output no human reviewed.

What is the cheapest way to get some AI coverage?

Almost always an endorsement on a policy you already buy rather than a standalone AI policy. The endorsement must name what it covers and state a sublimit. An endorsement that says AI losses are covered, without categories and without a number, is worth less than it reads.

Will better AI governance reduce my premium?

It will change your submission, and in a market pricing by individual submission that is the mechanism that matters. We do not publish a percentage. This page previously claimed a 15 to 30 percent reduction for documented governance; no carrier publishes such a figure and it was removed on 17 August 2026. What is checkable is that specialist underwriters ask for a defined scope of autonomous action, an approval threshold, and performance monitoring evidence, and that an operator who cannot produce those is asking to be priced for an unknown.

Sources

Previous: does your business insurance cover AI errors? Next: how to choose an AI insurance product