An indemnity is a promise. Insurance is money.
Your agreement with every AI vendor you use almost certainly contains an indemnity. It is worth exactly what the vendor can pay.
The AI company that recommended the wrong thing, took the wrong action, or drove into the wrong shelf is frequently a company with two years of runway and no balance sheet to speak of. When the loss is larger than their remaining cash — and the losses in this category are unusually good at being larger than the vendor — the indemnity does not transfer the risk to them. It leaves the risk with you, plus the cost of suing someone who is about to be insolvent.
This is not a new problem and your organisation already solved it everywhere else. You require general liability from contractors, professional indemnity from advisers, cyber cover from anyone touching your data, workers' compensation from anyone on your site. In each case the requirement exists because a signature alone proved uncollectable. Insurance is the mechanism that converts a counterparty's promise into a balance sheet that can actually pay.
Software that makes decisions is the one category where that requirement has not yet been written into the standard agreement — which is why it is currently the largest unhedged exposure most companies are carrying through their supply chain.
What to require, by what the system actually does
The requirement should scale with what the vendor's system is allowed to do unsupervised, not with how the vendor describes itself. A company calling itself an "AI platform" whose product only drafts text carries a different exposure from one whose agent has your payment credentials. Ask what it is permitted to do without a human, and require accordingly.
Advisory
Produces output a person acts on — recommendations, summaries, diagnoses, eligibility answers, pricing, drafted content.
- Fails like
- Confidently wrong advice, acted on at scale before anyone notices.
- Typical limit
- $1M–$5M each claim
- Ask for
- Tech E&O / professional liability, media & IP, no AI exclusion
Agentic
Takes action without contemporaneous approval — transacts, sends, writes to systems, changes records, calls tools with real-world effect.
- Fails like
- Correct instruction, wrong intent, executed thousands of times before a human sees it.
- Typical limit
- $5M–$25M each claim
- Ask for
- The above plus cover for unauthorised transactions and data alteration
Embodied
Moves, lifts, or operates in physical space — robots, humanoids, autonomous vehicles and equipment on your site or among your customers.
- Fails like
- Bodily injury. The only tier where the worst case is a person.
- Typical limit
- $10M+ each occurrence, higher in public space
- Ask for
- The above plus general liability for BI/PD, products liability, and confirmation the autonomous operation is not excluded
The clause
Drop this into the insurance section of your MSA, DPA, or vendor terms. It names no carrier and no broker, and it is satisfied by any A-rated insurer — that is deliberate, because a requirement that steers business to one insurer is a requirement your vendors are right to resist.
- Coverage. Throughout the Term and for twenty-four (24) months following its expiry or termination, Vendor shall maintain at its own expense insurance covering liability arising from the performance, output, decisions or actions of any artificial intelligence, machine learning, autonomous or automated system that Vendor provides, operates, or incorporates into the Services (an “AI System”), with limits of not less than [AMOUNT] each claim and [AMOUNT] in the aggregate.
- Scope. The insurance required by Section 1 shall respond to third-party claims arising from: (a) erroneous, misleading or harmful output of an AI System which any person or system acted upon; (b) actions taken by an AI System without contemporaneous human authorisation, including transactions, communications, and changes to data or systems; (c) infringement, misappropriation or breach of confidence arising from AI System output; and (d) where the AI System is embodied, bodily injury and property damage arising from its physical operation.
- No artificial intelligence exclusion. The policies required by this Section shall not contain any exclusion, sub-limit, or condition which excludes or materially limits cover for loss arising from the use, output or operation of artificial intelligence, machine learning, or autonomous systems. On request, Vendor shall provide the complete policy wording, or written confirmation from the insurer that no such exclusion or sub-limit applies.
- Insurers. All policies shall be placed with insurers rated not lower than A− (Excellent) by AM Best, or a recognised equivalent.
- Additional insured. Customer and its affiliates shall be named as additional insureds in respect of liability arising from the Services. Vendor’s insurance shall be primary and non-contributory to any insurance maintained by Customer, and shall include a waiver of subrogation in Customer’s favour.
- Claims-made policies. Where any required policy is written on a claims-made basis, the retroactive date shall be no later than the Effective Date, and Vendor shall either maintain the policy continuously or purchase an extended reporting period of not less than twenty-four (24) months.
- Evidence and notice. Vendor shall deliver certificates of insurance evidencing the cover required by this Section on execution, on each renewal, and on request. Vendor shall give Customer not less than thirty (30) days’ prior written notice of cancellation, non-renewal, or any material reduction in limits.
- No limitation. The insurance required by this Section shall not be construed to limit Vendor’s liability or its indemnification obligations under this Agreement, and Vendor’s obligations shall not be excused by any failure or refusal of an insurer to pay.
Section 3 is the one that matters most, and it is the provision your vendors are least likely to be able to satisfy today. This is no longer a drift — it has form numbers. In January 2026, ISO published optional generative-AI exclusion endorsements for commercial general liability (CG 40 47, CG 40 48, CG 35 08), more than sixty US insurance groups have filed with state regulators around adopting them, and W. R. Berkley has introduced “absolute” AI exclusions reaching D&O and E&O. According to reporting, regulators have approved the large majority of these filings. The endorsements attach only where a carrier adopts them — which is exactly why no certificate will tell you whether your vendor's renewal picked one up. And because the ISO wording targets generative AI specifically, whether an autonomous or robotic system falls inside it is genuinely unsettled. A vendor can hold a perfectly valid $10M technology E&O policy and still have no cover — or unknowable cover — for the exact thing you are worried about. Asking for it in writing, before a claim, is free.
Checking the certificate
A certificate of insurance is a summary written by a broker, not the policy. It is evidence that something was bought; it is not evidence that you are covered. Five things are worth checking, and the last two are where requirements usually fail quietly.
Must be the entity that signed your agreement — not the parent, not a sibling company, not the trading name. Group structures are where this breaks.
Look for the endorsement number, not a tick in a box. A certificate can say "additional insured" while no endorsement was ever issued.
If claims-made, check the retroactive date. A retro date later than your contract start means the first year of your relationship is uninsured.
The per-claim number is the one everyone reads. The aggregate is the one that runs out. A $5M / $5M policy that has already paid a $5M claim to somebody else has nothing left for you, and no certificate will tell you that.
The certificate will not show it. Exclusions never appear on a COI. This is the single check that requires asking for the wording or an insurer's written confirmation — and the reason Section 3 of the clause exists.
Telling your vendors
Requirements propagate by email, not by legislation. This is the note that does it — send it ahead of renewal so the vendor has time to place cover rather than a reason to argue about it at signature.
Where this is heading
Contractual requirements almost always run ahead of statutory ones, and that is the normal order of things — motor insurance was a condition of commercial contracts and municipal permits before it was compulsory by statute. It is worth being clear about what is actually in force today versus what is merely moving.
Autonomous vehicle permits
Several US states condition AV testing and deployment permits on minimum liability cover — California, for example, requires $5M for its programme. This is the clearest existing precedent for compelling insurance on an autonomous system.
Motor and workers' compensation
Both are statutory almost everywhere, and both became the largest lines in their markets because they were required rather than chosen. Compulsion, not danger, is what makes a line enormous.
EU Product Liability Directive (2024/2853)
Brings software and AI expressly within the meaning of a “product” for strict liability, including defects arising after supply through updates or learning. Member states are due to transpose it by December 2026. It expands who can be liable; it does not mandate insurance.
EU AI Act (2024/1689)
Obligations for high-risk and general-purpose systems phase in through 2026–2027. It is a safety and governance regime rather than a liability one, and it does not require deployers to carry liability insurance.
EU AI Liability Directive
Proposed in 2022 to ease the burden of proving AI-caused harm, and withdrawn by the Commission in 2025. Worth knowing because it is still cited as forthcoming. The liability pressure it represented did not disappear — it moved into the revised Product Liability Directive and into national courts.
Compulsory AI liability insurance
No jurisdiction currently requires a general AI vendor to carry liability insurance. Anyone telling you otherwise is selling something. Until that changes, the requirement lives in your contracts — which is precisely why the contract is the lever.
The honest read: statutory compulsion for embodied systems in public space is plausible within a few years, because that is how every other machine allowed near the public has been handled. For software that decides, it is further off. Neither timeline helps you with the vendor you are signing with this quarter.
Our interest in this
Roboline sells the insurance this clause requires. We publish the clause because more companies requiring cover means more companies buying it, and we would like to be among the sellers. That is the whole motive and there is no second one.
So the clause is written to be usable against us. It names no carrier, specifies no broker, and is satisfied by any insurer rated A− or better. If your vendor already holds compliant cover elsewhere, the requirement is met and we get nothing, which is the correct outcome — a requirement that only one company can satisfy is a sales document, and your vendors will read it as one.
Use it, fork it, strike the parts you disagree with. It is CC BY 4.0.