Introduction
Enterprise underwriting still relies, in large part, on a file assembled at a point in time: questionnaires, architecture descriptions, test results, and a compliance dossier. That method is coherent for a deterministic system whose executable behavior is expected to remain the same between two versions.
It is a weak method for a generative or agentic system in production. The object declared at binding can change without a declared change of activity. An audit at T0 can therefore remain formally complete while becoming operationally obsolete.
This article distinguishes static audit from continuous behavioral evidence and states what the second can, and cannot, contribute to insurability.
Defining Static Audit
A static audit is a documented assessment of an AI system at a defined moment. It typically records:
- the intended use and the declared policy of use
- the technical architecture available at that date
- test or evaluation results obtained on a chosen sample
- organizational controls described by the insured
The output is a snapshot. It may be accurate at T0. It does not, by itself, describe the system at T+n.
Why a Point-in-Time Audit Ages
After deployment, several changes can alter the risk profile without rewriting the insurance file:
- remote updates of a foundation model by the provider
- changes to prompts, tools, or routing
- modification of retrieval sources or reference corpora
- variation in volume and in the financial value attached to each output
- weakening or bypass of human review
A questionnaire updated once a year does not capture that tempo. Silent AI is one market consequence of the gap: the contract remains unchanged while the operational object moves. That mechanism is defined in Silent AI: When Generative Risk Hides in Existing Policies.
Defining Continuous Behavioral Evidence
Continuous behavioral evidence is a set of observations produced over time under a documented measurement protocol. It is not a second audit. It is a temporal record of observable outputs and of the conditions under which those outputs were measured. A static audit establishes a state. Continuous behavioral evidence establishes a trajectory in a documented measurement frame.
In this sense, evidence improves when it can show:
- what protocol was used
- which system was on the measurement path
- how selected behavioral signals changed relative to a documented baseline
- whether changes in the documented operational context accompanied the observed behavioral variation
The relevant property for insurability is not a single score. It is the possibility of reconstructing change after the fact. Measurement does not determine insurability. It makes part of the evidence required for insurability observable. That requirement is already part of the definition of insurability in What Is AI Insurability?
What Continuous Evidence Is Not
Continuous behavioral evidence is not:
- a permanent recording of every prompt and response
- an automatic detection of undeclared AI uses outside the measurement path
- a prediction of a future loss
- a finding of fault, liability, or cover
- a substitute for the insurance contract
A change observed under protocol is a signal. It does not decide whether a policy responds.
Implications for Insurability
Static audit and continuous evidence are complementary. The first describes the object at inception. The second provides evidence of whether measured behavioral properties remained comparable over time.
For an underwriter or a claims handler, the useful distinction is:
| Static audit | Continuous behavioral evidence |
|---|---|
| Describes the system at T0 | Describes observed variation after T0 |
| Relies on declaration and a sample | Relies on a repeated, documented protocol |
| Ages when the system changes | Can show that a change occurred |
| Does not reconstruct later behavior | Does not, by itself, interpret that change as a claim |
Insurability improves when both exist. It does not appear automatically when a dashboard exists.
Relationship to Measurement
Behavioral measurement can support continuous evidence only when the system is integrated into the measurement path and the protocol is documented. NeoMundi produces runtime behavioral measurement signals from observed model outputs under documented conditions. It does not underwrite, price insurance, or adjudicate claims. The operational value is not a calculated return. It is a reduction of factual uncertainty: a later reviewer can see what was observed, under which protocol, and how selected signals differed from a documented baseline.
The measurement methods that make a temporal comparison interpretable are set out in the governance-metrology corpus, in particular From Controlled Measurement Campaigns to Continuous Production Monitoring and Longitudinal Monitoring and Behavioral Drift Detection in Production AI.
Detecting a change identifies a difference in observed outputs or signals. It does not, by itself, establish the cause of that difference, nor convert the observation into an insured event.
Limits
This article does not:
- reject static audit as useless
- treat continuous measurement as a completed market standard
- use a single campaign result as a general law of AI failure
- present any unfinished passport or claims tooling as an available product
- equate drift with damage
Conclusion
A static audit answers a necessary question: what was declared and observed at inception. It does not answer the question that appears at claims time: what observable behavioral state was recorded around the event, and how did it differ from the documented baseline.
Continuous behavioral evidence does not replace the contract. It makes the second question addressable. The next article examines what kind of loss, once evidenced, can enter the insurable core.
This reference article was developed from the NeoMundi working paper “L’illusion du contrôle statique” by Frédéric Dumollard (September 2026). It is a condensed definitional version for the AI Insurability Reference series, not a full republication.
Source: https://doi.org/10.5281/zenodo.22297814