Cyber underwriting is the wedge.

Why we started here.

We start in insurance underwriting because it is the regulated decision function on earth with the highest density of tacit expertise, the most acute retirement exposure, the heaviest active regulatory pressure on AI explainability, and the most mature data infrastructure for behavioural capture. The combination is not accidental. It is the reason the category is going to be born in insurance, and it is the reason the first company to build a defensible position in insurance underwriting will own a meaningful share of the broader regulated AI infrastructure market over the next decade.

Within insurance, we start specifically in cyber underwriting. Cyber is the segment in which the gap between the foundation model's native capabilities and the senior underwriter's judgment is widest, because cyber risk is the part of the property and casualty market where the historical loss data is thinnest and the value of unwritten judgment is correspondingly highest. It is also the segment with the youngest decision frameworks, which means there is no thirty-year-old actuarial table to compete against, which means the methodology we are bringing has the cleanest available path to differentiated performance. It is the segment where the specialty MGAs are moving fastest on AI infrastructure decisions, which compresses the procurement cycle relative to a tier-one admitted carrier from twelve-to-eighteen months to three-to-six. And it is the segment where our advisor network is densest, which means the first three customer conversations begin on warm introductions rather than cold outbound.

What the work actually looks like.

A typical engagement opens with a two-week archaeology run against the carrier's existing decision exhaust. The run produces a first pass at the heuristic inventory before we have asked any expert to do anything new. The output is reviewed by the carrier's senior underwriting leadership, who almost invariably recognise the patterns the system is surfacing and frequently point out the ones the system missed in the first pass.

The second phase introduces contrastive cohort analysis across the underwriting team and brings the Structured Interview Agent into rotation with the senior cohort. This is when the methodology starts to surface the patterns that historical data alone could not show. We typically run twelve to twenty hours of interview sessions across five to eight senior underwriters over a four-to-six week window, which is paced deliberately so as not to disrupt the underwriting workload. The third phase deploys perturbation probing against the rare-case tail and switches on live escalation capture against whichever agent platform the carrier is running in production.

By the end of month three, the carrier has a populated cognitive graph, a measurable difference in agent override rates on instrumented cases, and a regulatory-grade audit trail that documents, for each agent recommendation, the named heuristics applied, the confidence scores, the provenance, and the residual disagreement among the experts. By the end of month six, the override rate on the instrumented cohorts has typically reduced by 30 to 50%, and the carrier has the beginning of a durable internal capability for capturing senior judgment as the senior cohort works through the next renewal cycle.

Where we go from here.

Insurance underwriting is the wedge. It is not the destination. The methodology applies, with vertical-specific adaptation, to every regulated decision function where the failure mode is the same: senior expertise carrying a disproportionate share of the decision quality, the senior cohort retiring faster than they can be replaced, and AI agents being deployed against decisions whose runtime has never been observed.

The expansion path runs through commercial credit underwriting (asset-based lending, equipment finance, specialty lending) in years two and three of the company, into the heavy regulated industries where tacit expertise concentrates most acutely (nuclear power operations, oil and gas, complex industrial process control, clinical decision support in tertiary care) in years three through five. These are the verticals in which the consequence of a senior expert retiring without their runtime captured is most immediate, the regulatory environment most clearly mandates explainable decision-making, and the willingness to pay for a defensible solution is highest. The combined addressable spend across the expansion verticals sits at roughly $25-30 billion by 2030 by our bottoms-up estimate.^{14TAM estimateInternal bottoms-up TAM across insurance UW, commercial credit, clinical decision support, AML/KYC. Triangulated with Goldman Sachs Global AI Capex Forecast (2024) and IDC AI Spending Guide (2025).} We are not trying to capture all of it. We are trying to build the layer that the next decade of regulated AI deployment depends on. The path to that layer runs through being the best in one vertical before expanding to the next, and the long-term bet is that vertical AGI is not buildable, at any meaningful level of operational reliability, without behavioural observation infrastructure of the kind we are building. Whoever owns the cognitive cartography of regulated decision-making owns the data substrate on which the next generation of operational AI is going to be trained.