The Death of the General-Purpose Model in Enterprise Risk

For the last two years, the prevailing narrative in Silicon Valley has been the rise of the general-purpose model—the idea that a single, massive LLM could eventually handle every corporate task from writing emails to managing complex risk. But in the world of high-stakes enterprise risk management, that hype is hitting a wall of reality.

As first reported by the Financial Post, the insurance industry is signaling a definitive pivot away from general-purpose AI. While there is a massive appetite for automation, there is an equally massive distrust of ungoverned models.

**Opinion:** From a practitioner's perspective, this isn't just corporate caution; it is a necessary correction. In a sector where a single hallucination or a misapplied policy wording can lead to catastrophic financial exposure, the 'black box' nature of general-purpose AI is a non-starter. The industry is realizing that efficiency without governance is simply a faster way to make a mistake.

Data from a report published by ISG and commissioned by mea Platform underscores this divide. While 83% of the global insurance market supports using AI to execute repeatable work, the conditions for that deployment are strict. A significant majority—75%—would only permit AI execution if the model is specifically built for insurance or governed within their own internal rules.

When asked what they would trust for high-consequence claims and underwriting decisions with limited oversight, the results were telling: 75% opted for a governed hybrid or an insurance-specific model. Just 6% of those surveyed indicated they would trust a general-purpose model acting alone.

This rejection of general AI is rooted in the definition of 'governed' systems. As detailed by the Financial Post, these systems are not left to their own devices; they operate using controlled underwriting appetite, claims guidance, endorsements, and policy wording. Crucially, these systems separate reading permissions from decision-making permissions and allow a named person to stop or override the AI.

The motivation for this shift is operational capacity. Carriers estimate that one in nine broker submissions is currently left unquoted or declined—even on risks they have the appetite for—simply because operations cannot keep pace. By automating repeatable tasks like submission intake, triage, quote generation, and compliance screening, insurers aim to reclaim this lost business.

Despite the push for automation, the human element remains non-negotiable. The ISG report found that 86% of senior leaders across technology, operations, claims, and underwriting believe that consequential decisions—those that define a company's distinction—must remain with people. The goal is not to replace the underwriter, but to use AI to sharpen the evidence available to them, allowing for faster, more informed risk decisions.

Currently, the transition is in its infancy, with less than 1% of the market running fully AI-native operations. However, for those already deploying these tools, the results are tangible: 61% report productivity gains and 51% report faster cycle times. Operating costs are projected to drop by 16% over the next two years.

Ultimately, the insurance industry's stance serves as a blueprint for other highly regulated sectors. The era of the general-purpose AI experiment is ending; the era of the governed, domain-specific tool has begun.