Introduction
Insurance can transfer a risk only when that risk can be described, bounded, and evidenced. Many AI systems in production fail this test. They change after the contract is signed. Their failures are often distributed across a model provider, an integrator, a prompt, a knowledge base, and a human workflow. The resulting loss may be real without being easy to classify.
AI insurability is therefore not the same thing as the existence of an “AI insurance” product. It is the set of conditions under which a loss linked to an AI system can enter an insurance contract without collapsing into an unallocated commercial disappointment or an unprovable allegation.
This article defines AI insurability, distinguishes it from adjacent notions, and situates it relative to behavioral measurement. Measurement can support insurability. It does not create cover, price a premium, or decide a claim.
Defining AI Insurability
AI insurability is the capacity of a loss associated with the development, deployment, or use of an AI system to be:
- identified as a damage rather than a mere business variance
- attributed to a defined activity, system, period, and contractual role
- evidenced under documented conditions
- bounded in severity, frequency, and excluded causes
- transferred, in whole or in part, by an insurance contract
An AI risk may be operationally serious and still poorly insurable if those conditions are not met.
What AI Insurability Is Not
AI insurability should not be confused with the following:
- The marketing availability of a dedicated AI policy
- A vendor indemnity or intellectual-property warranty
- Regulatory compliance under the EU AI Act or any other statute
- A reliability score, benchmark result, or one-off audit
- A measurement signal produced at runtime
A signal can inform a later underwriting or claims assessment. It is not itself an insurance verdict, a finding of liability, or a pricing decision.
Conditions That Make an AI Risk Insurable
In practice, insurability depends on several jointly necessary conditions.
A defined object of risk
The insured activity, system, and use case must be described with enough precision for a contract to attach. “We use AI” is not a risk object.
A separable damage
The loss must be distinguishable from ordinary commercial underperformance: a missed sale target, a weak marketing text, or a slow process improvement are usually not insurable events.
An evidence path
After an incident, it must be possible to reconstruct, at least in part, what system was in use, under which protocol, with which versioning, and with which human controls active. Absence of evidence does not prove absence of damage. It does limit what a contract can settle.
A temporal frame
AI systems drift. Cover that assumes a frozen system at inception is structurally weak. Insurability improves when change over time can be observed rather than merely declared. This point is developed in Static Audit vs Continuous Behavioral Evidence.
Allocated roles
Provider, deployer, integrator, and user do not carry the same duties. A contract that cannot locate the insured’s role will struggle to locate the insured’s loss.
What AI Insurability Is Not designed to absorb
Administrative fines, deliberate misuse, and losses outside the declared use case typically remain outside the insurable core. Those limits are treated in later articles of this series.
Why Traditional Underwriting Assumptions Break
Classic enterprise underwriting often assumes that:
- the risk declared at inception remains stable
- a periodic questionnaire updates that declaration
- a technical audit at T0 is a sufficient proxy for later behavior
Generative and agentic systems violate those assumptions. Model providers update systems remotely. Retrieval sources change. Prompts and tools are revised. Volume and financial exposure fluctuate. A file that was accurate at binding can become incomplete without any formal change of activity.
Silent AI is the market expression of that gap: AI-related losses absorbed by contracts that were not designed or priced for them. That mechanism is defined in Silent AI: When Generative Risk Hides in Existing Policies.
Relationship to Behavioral Metrology
Behavioral metrology does not make a system insurable by itself. It can, under documented conditions, make three things more tractable:
- observation of behavioral change over time under a documented protocol
- comparison against a baseline
- reconstruction of how a signal was produced
Those functions belong to the governance-metrology corpus, in particular “What Is AI Governance Metrology?” and “Signal vs Verdict: Core Principle of Responsible AI Evaluation“.
The boundary remains strict:
- metrology produces observations
- insurance interprets those observations inside a contract
- human decision-makers retain authority over cover, exclusion, and settlement
NeoMundi produces runtime behavioral measurement signals from observed model outputs under documented conditions. It does not underwrite, tariff, or adjudicate claims.
Limits of This Definition
This definition does not:
- claim that all AI losses are insurable
- treat current market products as complete or equivalent
- convert a runtime signal into proof of fault
- require permanent central storage of prompts and responses
- present any unfinished instrumentation project as an available product
Insurability is a property of a risk inside a contract. It is not a property of a model in isolation.
Conclusion
AI insurability is the point at which an AI-related loss becomes contractually usable: identifiable, evidenced, bounded, and transferable. Many production systems are used at scale before that point is reached.
The rest of this series examines the market form of that problem (Silent AI), the evidence problem over time, financial materiality, the limited support provided by the EU AI Act, the structure of existing covers, and the conditions under which claims can be investigated without turning measurement into a verdict.