AI won’t scale without trust: What boards can learn from cyber insurance

QBE CEO Andrew Horton delivered a timely warning recently that artificial intelligence will not deliver the productivity gains business and government are counting on unless the risks that come with it are properly managed. His call for cooperation between business, government, insurers and big technology providers is, on its face, uncontroversial. Everyone agrees collaboration is needed.

Insurance is not merely protection against AI going wrong. Done well, underwriting becomes part of the governance infrastructure that gives boards, regulators and the market the confidence to adopt AI at scale. And for the genuinely systemic risks, no single sector, business, insurer or government, can carry that exposure alone.

I made a version of this argument in 2024, in a speech calling for government to establish a “catastrophe” cyber fund to support organisations facing the reality of systemic cyber risk. My case then was that the private insurance market is built to spread risk that is largely independent: for example: one factory fire does not cause the next. Cyber risk does not behave this way.

A single vulnerability in a widely used platform can crystallise losses across thousands of organisations simultaneously. That is precisely the kind of tail risk that overwhelms private capital and requires government to stand behind the market as an insurer or reinsurer of last resort, in the same way it does for terrorism and some natural catastrophe risks.

AI has the potential to create the same structural problem, correlated losses arising from common models, cloud infrastructure and technology providers, but at greater scale and potentially much greater speed. QBE’s own research gives an early indication of how quickly this is materialising: 29% of businesses globally reported experiencing a cyber incident in the past year in which AI was believed to have been used in the attack.

Insurance as risk transfer and as strategy

The value of insurance extends well beyond the payout after an event. At its best, it forces an organisation to identify its exposures, understand which risks it can retain and which it should transfer, and demonstrate that it has controls an underwriter is prepared to back with capital. Insurance is, properly understood, two things at once: a means of transferring risk an organisation cannot or should not hold on its own balance sheet, and a strategic discipline that shapes how an organisation manages that risk day to day, because the underwriting process asks the questions a board should already be asking itself.

If AI adoption is to scale responsibly, insurance needs to sit inside that strategy from the start, not be bolted on afterwards once something has gone wrong. That means insurers need to be genuine partners in how organisations deploy AI, not simply the party asked to price an unfamiliar risk after the technology is already embedded in core systems. It also means business cannot treat insurance as a substitute for governance.

A policy is not a control. It is a risk analysis and transfer mechanism, a backstop for the controls that should already exist, and the best boards use the process of securing and renewing cover as a genuine strategic input, not a compliance exercise handed to the risk team once a year.

The board’s role has not caught up with the risk

This is where board governance has to catch up. Cyber risk, and now AI risk, is not a technology issue that can be delegated to management and monitored from a distance. It is an enterprise risk that sits squarely within a board’s oversight duties, and the standard expected of directors has moved well past a general awareness that the risk exists.

For boards, that means knowing where AI is being used, which applications could create material harm, what risk appetite applies, who is accountable, how systems are tested and monitored, where critical third-party dependencies sit, and whether the organisation’s insurance program actually responds to the resulting exposures.

A board that lacks sufficient collective capability to understand and challenge management on material cyber and AI risks now has a governance gap that needs to be addressed. Boards that engage properly with their insurers, not just at renewal but as an ongoing source of intelligence on how the threat and the market are shifting, put themselves in a materially stronger position, both to manage the risk and to demonstrate they have discharged their duties if something goes wrong.

Five questions every board should be asking

Boards do not need to become AI or insurance experts, but they do need to ask the right questions and be satisfied with the answers. Five questions are particularly important:

1. Where are we using AI, and what could go seriously wrong? Do we understand our most significant AI applications, their potential consequences and our dependence on third-party providers?
2. What is our appetite for AI risk? Have we established clear boundaries for AI adoption, particularly where decisions could materially affect customers, employees, financial outcomes or our reputation?
3. Are our governance and controls keeping pace? Is accountability clear, are our most significant AI risks being adequately tested and monitored, and are we confident in our ability to respond when something goes wrong?
4. What risks are we retaining, transferring or potentially leaving uninsured? Do we understand how our existing insurance policies respond to AI-related losses, including exclusions, limitations and gaps in cover?
5. Are we learning from our insurers and the wider market? Are we using insights from insurers and other external experts to strengthen our governance, anticipate emerging risks and improve our resilience?

These questions should form part of regular board oversight, not simply an annual insurance renewal discussion. The objective is not to eliminate AI risk, but to ensure the organisation understands the risks it is taking, manages them deliberately and has the resilience and financial protection to pursue AI opportunities with confidence.

The standalone AI insurance debate

There is a live and legitimate debate about whether AI risk warrants its own standalone insurance product, separate from cyber, professional indemnity, product liability and D&O cover, or whether it is better addressed by extending and clarifying those existing lines. I do not think that question can be properly answered yet, and I would treat anyone who claims certainty on it with some scepticism.

Existing Australian laws already apply to many AI-related harms, but significant uncertainty remains about how liability will be allocated as AI systems become more autonomous, complex and embedded in business decisions, particularly across developers, technology providers, deploying organisations and human decision-makers. Insurers cannot price a risk with confidence while that allocation is still unsettled.

A small number of specialised underwriters are now offering dedicated AI liability policies, and in my view, strong, comprehensive standalone AI products will only emerge once the risk is distinct enough and the data is good enough to support them. In the meantime, we are likely to see continued patching of existing policies with AI-specific endorsements, exclusions and sub-limits, in much the same way silent cyber exposure was addressed within traditional professional indemnity, property and casualty policies before a dedicated cyber market matured around it.

The lesson we should not repeat

That last comparison is the one that matters most, and it is the one I would urge boards, insurers and government to consider honestly. The cyber insurance market did not mature in an orderly way. For years, cyber exposure sat unpriced and unrecognised inside other lines of cover.

As the scale and concentration of cyber exposure became clearer, the market corrected hard and fast: premiums rose sharply, capacity was withdrawn, and cover was made conditional on risk management protocols, such as multi-factor authentication, endpoint detection and tested incident response plans, that many organisations were then forced to retrofit under pressure, rather than having built in from the outset. The market got there in the end, but the correction was disruptive, and it left a protection gap that persists today, particularly for small and mid-sized organisations that can least absorb an uninsured loss.

We have the chance to do better with AI, precisely because we can see the cyber experience so clearly in the rear-view mirror. That means insurers, government, business and technology providers building the risk management protocols and the data-sharing arrangements now, ahead of the losses, rather than in response to them. It means government and insurers considering now whether existing public-private risk-sharing mechanisms could cope with genuinely systemic AI-related losses, and, if not, what form of backstop might ultimately be required.

It also means boards treating AI governance with the same rigour they are now, belatedly, applying to cyber governance: actively tracking emerging trends, incorporating AI risk thinking into strategic decisions, testing their organisation’s resilience, and using insurance as a genuine strategic tool, rather than waiting for a regulator or an insurer to make it mandatory.

Horton is right that AI will not deliver its full productivity potential unless its risks are managed properly. I would go further. Risk management that ignores risk transfer can leave organisations carrying exposures they do not fully understand. Insurance without effective board oversight can create false comfort. And an insurance market that develops without genuine collaboration between government, business, insurers and technology providers risks repeating the painful evolution of cyber insurance.

We have already paid the learning fee once. The opportunity with AI is to apply those lessons before the losses, rather than after them.