Introduction
For several years, losses linked to AI systems have been absorbed by enterprise policies that were neither designed nor priced for that exposure. The mechanism is not new. It reproduces, in another technological cycle, what the market already observed with Silent Cyber.
Silent AI names that situation: a generative or agentic risk that sits inside an existing contract without being identified as a distinct peril. The contract may respond. It may also respond late, partially, or not at all, because the wording, the questionnaire, and the premium never treated AI as the object of cover.
This article defines Silent AI, locates it in common policy families, and separates it from dedicated AI insurance.
Defining Silent AI
Silent AI is the implicit absorption of AI-related loss by insurance contracts that do not name AI as a covered, excluded, or specifically underwritten peril.
Three elements are required:
- an AI system is used in an insured activity
- a loss occurs in connection with that use
- the applicable contract was not constructed around that use as a distinct risk object
Silent AI is a market condition. It is not a product, a guarantee, or a measurement result.
How Silent AI Enters Existing Policies
AI systems enter ordinary business processes without a corresponding change in the insurance file. Typical entry points include:
- a customer-facing chatbot that produces contractual information
- a drafting or coding assistant whose output is published or shipped
- an internal agent that ranks suppliers, prices, or claims
- a retrieval-augmented system that answers from evolving document stores
The insured activity appears unchanged on paper. The operational exposure does not.
The analogy with Silent Cyber is useful only up to a point. Cyber risk was often treated as an extension of existing liability or property wordings until incidents forced explicit grants and exclusions. AI risk follows a similar path, with one added difficulty: the system can change after binding without a declared change of activity.
Why Silent AI Is Not Dedicated Cover
Silent AI and dedicated AI insurance must be kept distinct.
| Silent AI | Dedicated AI cover |
|---|---|
| Cover is inferred from a general wording | Cover is granted for a defined AI use |
| Premium does not price the AI exposure | Premium, at least in principle, targets that exposure |
| Questionnaire rarely isolates AI systems | Underwriting asks for systems, uses, and controls |
| Claims handling reconstructs the risk after the event | Claims handling starts from a declared AI object |
A dedicated product can still be narrow, conditional, or poorly adapted. Its existence does not abolish Silent AI in the rest of the portfolio.
Typical Policy Families Concerned
Silent AI is most often discussed in relation to:
- professional indemnity / professional liability
- cyber
- directors and officers
- business interruption
These families can respond to some AI-related losses. They were not built to describe model updates, prompt changes, retrieval drift, or the division of roles between provider and deployer.
Whether a given loss falls inside one of these contracts is a matter of wording, fact, and applicable law. Silent AI does not mean automatic cover. It means the question is asked too late.
What Silent AI Conceals
Silent AI hides several operational facts that matter at claims time:
- which model version was in use
- which instructions, tools, and reference sources were active
- whether human review was required and actually performed
- whether the use exceeded the activity declared at inception
- whether the loss is a damage or a commercial variance
Those facts are not supplied by the mere presence of an insurance policy. They depend on evidence assembled over time. That evidence problem is treated in Static Audit vs Continuous Behavioral Evidence.
A public case may illustrate the contractual surface of the problem without exhausting it. In Moffatt v. Air Canada (2024 BCCRT 149), a chatbot communication was treated as attributable to the company. The case is an illustration of responsibility for published output. It is not a template for every AI claim and not a measure of insurability.
Relationship to Measurement
Behavioral measurement does not convert Silent AI into dedicated cover. It can make the hidden exposure more visible by documenting:
- the presence of an AI system on the measurement path, when NeoMundi is integrated into the production flow
- change in observed behavior after inception
- the conditions under which a signal was produced
Those observations remain signals. They do not decide whether a policy responds.
This distinction is the same as in Signal vs Verdict: Core Principle of Responsible AI Evaluation. The definition of the underlying property is set out in What Is AI Insurability?
NeoMundi does not determine whether a silent exposure is covered.
Limits
This article does not:
- map every AI loss to a named policy
- treat Silent AI as equivalent to an exclusion or to an affirmative grant
- present current market products as a complete substitute
- use a measurement campaign as proof of an insured event
- rely on any unfinished claims or passport tooling
Conclusion
Silent AI is the lag between the operational use of generative systems and the contractual description of that use. Existing policies may absorb part of the loss. They do not, by themselves, make the risk identifiable, priced, or easy to investigate.
The next question is not only which wording applies. It is whether the risk can still be evidenced after the system has changed. That is the subject of the following article.