GLG Frontier
>
Research

Insurers Can't Price AI Risk Until Someone Can Show How the System Behaves

As carriers rethink how AI losses fit into existing policies, the evidence behind AI deployments is coming under new scrutiny. Four practitioners – from frontier-model evaluation, insurance claims, AI policy, and enterprise AI delivery – explain what today’s testing proves, what it misses, and what an underwriter would need to see.
Insurance
Frontier Briefing
GLG Frontier
November 2026
CONTRIBUTORS
GLG Frontier Research Team

AI evaluation grew up around one question: should a lab ship this model? But as the technology moves from drafting text to taking actions, insurers are demanding evaluations answer more pointed questions – namely, how often will the system fail, and what will it cost when it does?

We spoke to four practitioners in the insurance and risk space – including an independent advisor who evaluates frontier models, a former chief claims officer, an AI company CEO and policy advisor, and a former VP of AI platforms at a global IT firm – to understand how carriers, evaluators, and buyers are meeting such questions.

All four say insurers are approaching AI as they would any new, hard-to-price risk: pull back and wait for evidence to price on. Where they split is on what closes the gap – pressure from insurers and buyers, or public rules on transparency and independent testing.

“Insurers can’t price what they can’t measure,” the evaluation specialist said.

Insurers Step Back First, Then Price on Evidence

Carriers’ approach to AI resembles a familiar precedent. They start by narrowing what existing policies cover, which pushes losses that boards assumed were insured back onto the company’s balance sheet. When the CFO notices, the internal question shifts from whether the company uses AI responsibly to whether it can prove that to an investor or insurer.

In the evaluation specialist’s view, carriers that write affirmative AI coverage – policies written specifically for AI failures – will need something to underwrite against, such as independent test results, documented controls, and monitoring data. He expects insurers to turn assurance best practice into the price of cover, similar to how cyber insurers helped spread basic security controls. That would also change who pays for testing. Today it’s often the CIO or head of AI. Once capital is at stake, it will become the CFO, the general counsel, and the risk committee, with bigger budgets and less patience for paperwork that won’t hold up.

What Gets Tested Today Isn’t What Gets Insured

Most frontier testing is funded, scoped, and published by the labs themselves, with government institutes and a small group of specialist testers alongside them. Outsiders see the results mainly through system cards, which have gotten more transparent in the past few years, but still don’t quite fulfill the independence insurers demand. Each lab reports against its own framework, the effort spent drawing out risky capabilities is rarely visible, and outside testers’ findings arrive as the lab’s summary.

“A system card is a disclosure document, not an audit opinion,” the evaluation specialist said.

Even with reliable tests, scope represents a wide gap for insurance. More than just covering a model, an underwriter has to cover a company that has wrapped a model in its own prompts, tools, permissions, and data, often on a later version than the one tested. An insurer’s exposure sits at that deployment layer with the agent built on top of the model, and so does most enterprise audit work.

A Certificate Proves Process, Not Behavior

By now, enterprise procurement teams regularly look for ISO 42001 certification, the international standard for AI management systems. While it tells a buyer that a company has an AI policy, clear ownership, risk and impact assessments, and incident handling, it’s not quite sufficient – it says little about whether a specific system is accurate or safe in a specific use. Certified companies can still be breached, the evaluation specialist noted. He asks for a second layer of evidence – like use-case evaluation results, red-team findings and how they were fixed, production error and drift metrics, and what changes when the vendor swaps the underlying model.

The former AI platforms executive described a similar imbalance inside enterprises. Most firms he works with already keep AI use policies, use-case inventories, and due-diligence records, driven more by customer RFPs than by regulation. Such governance is strong on inputs – e.g., which model, which data, who approved it – but weak on the outcomes an underwriter would want to see, like whether the model is drifting, whether error rates differ across customer segments, or how many agent actions were reversed last month, for example.

Validate the System, Not the Weights

Banking-style model risk management offers a template, in the platforms executive’s view, but its validation step breaks down on generative AI. Deployers can’t inspect a closed model’s weights or training data, and the same input won’t always produce the same output. He described four practices that hold up:

  • Tier systems by decision impact and risk, not by technology.
  • Validate what the deployer controls: prompts, retrieval, guardrails, and evaluation harnesses.
  • Treat a suite of curated behavioral test cases, rerun against thresholds after every change, as the new validation report.
  • Keep generative AI out of the regulated decision itself. Let it prepare the case, and leave the decision to an explainable model.

Validation thresholds need a realistic baseline, too. “The mature practice is to benchmark the model’s output against the human error rate,” the platforms executive said. Tolerance can be looser where a person reviews the output and tighter where the system completes a transaction.

For long-running agents, the evaluation specialist adds that a pass rate doesn’t reveal enough. Reviewers need to read what the agent actually did – whether it took shortcuts, tampered with a test, exceeded its permissions, or redefined the goal. “For an agent, the score is a claim, the trajectory is the evidence,” he said. Because models can also recognize when they’re being tested, he expects weight to shift from one-time pre-release testing toward monitoring in production.

Independence Is Still Unresolved

If insurers are to rely on test results, they need to trust who produced the test. Achieving that trust is a tall order today, the evaluation specialist noted, because the developer usually chooses the model version, the testing window, and what gets published, and often pays the tester. It took decades for financial auditing to build independence rules against exactly such an arrangement.

The policy advisor ranks independent testing among the hardest pieces of any safety floor. “Lastly, independent evaluation, this is the hardest to build, but it’s the most important,” he said.

While they agree on the importance of independence, the two experts see different solutions. The evaluation specialist expects commercial pressure from insurers, procurement teams, and boards to raise the bar, with credible testers earning a premium once insurers start to rely on their reports. The policy advisor emphasized transparency: a public process for reporting AI incidents, rather than leaving victims to come forward, and independent evaluation governed jointly by government, industry, and the public.

Nobody Has Settled Who Is Liable

Better evidence won’t help if policy can’t articulate what it covers. When an AI system fails, the former claims chief asks what counts as a failure, what triggers the policy, and whether liability sits with the deployer or the AI provider. A single carrier can rewrite its wording quickly, but the market needs shared definitions that are precise and durable. “Getting those definitions right is as important as working on this quickly,” he said.

For now, contracts fill in the gaps where insurance pulls back. The evaluation specialist expects enterprises to push liability onto vendors through warranties, indemnities, and audit rights, and vendors to commission independent assurance to win those deals. Whoever collects structured data on AI failures, incidents, and test results may end up holding what insurers lack most: something like a loss history.

Our Take: What This Means for Teams Building AI Systems

If insurers follow the path these practitioners describe, the most valuable evaluation work will be the kind an underwriter can price. For teams deploying models into regulated or high-stakes settings, a few questions are worth testing now:

  • Can you produce evidence about your deployed system – its tools, permissions, and data – rather than pointing to the base model’s system card?
  • Does your eval suite rerun on every model swap and prompt change, with thresholds and a diff against the last run?
  • Are your error thresholds anchored to a measured human baseline for the same task?
  • Do you review agent trajectories, not just outcomes, and log reversed actions and incidents in a form an outsider could audit?
  • Can you state what your testing didn’t cover, including how hard you tried to elicit risky behavior?
  • If many deployments share your foundation model, have you tested for failures that would hit all of them at once? That’s the accumulation question an insurer will ask.

Such questions aren’t purely technical. The right thresholds, failure scenarios, and evidence standards come from people who have priced risk, handled claims, and validated models inside regulated institutions – the kind of judgment that turns an eval suite into evidence leaders are confident enough to put capital behind.

Go Deeper

This article draws on four GLG Library interviews: “AI Evaluation, Assurance and Third-Party Audit Market” (October 14, 2026), a discussion of specialty cyber insurance strategy (October 1, 2026), “The AI Policy Race: Regulation, Risk and U.S.-China Competition” (September 29, 2026), and “Enterprise AI Policy, Data Governance, and Business Impact” (September 12, 2026). GLG clients can read the full transcripts in myGLG. To pressure-test your own evaluation or assurance approach, GLG can connect you with practitioners in underwriting, AI evaluation, and model risk through a one-on-one expert call. Teams building benchmarks for high-stakes deployments can work with GLG Frontier to design evals and red-team scenarios grounded in real practitioner judgment.

About this piece: This Frontier Briefing draws on four GLG Library interviews recorded in September and October 2026. The experts are an independent advisor who evaluates frontier reasoning models in high-consequence settings, with more than 20 years in enterprise strategy and emerging technology; a former chief claims officer at a major U.K. insurer with nearly three decades in the industry; an AI company CEO and policy advisor who has briefed U.S. officials on AI governance; and a former vice president of AI platforms at a global IT services firm. The views on testing, underwriting, and governance are theirs. The takeaways for builders are GLG Frontier’s.

Get connected to the world's leading platform for trusted human expertise

Enter your information below to get in touch with our team.