Markets are behaving as though AI is about to deliver a step change in corporate earnings and productivity. There is only a handful of years of data available to test that assumption, and most of it comes from a period when funding was abundant. If this feels like déjà vu, that's because it is: we have already seen a version of it with climate change, a problem that has not been solved yet.

Under IFRS 9, banks provision use probability-weighted, forward-looking scenarios to estimate expected credit losses (ECL) rather than waiting until borrowers fall behind on payments. It is a sound approach and a clear improvement on the crisis-era practice of recognising losses only after they had occurred. Yet recent conversations with banks, investors, auditors and regulators suggest that even where climate risk is being taken seriously, it is proving strikingly hard to see in the numbers.

Why climate risk remains hard to quantify

Banks are not ignoring climate risks. Modelling approaches are still maturing and, understandably, differ between institutions. Climate risk is being considered in credit loss calculations, but it is often judged immaterial, partly because the assets may be sold before the effects are expected to become visible. There is also a real worry about double-counting climate factors already captured implicitly in existing economic scenarios. And practitioners are largely in agreement that the common constraint is data, not interest.

Investors want a clearer link between the assumptions used in these models and the figures reported in financial statements. This would help them assess banks’ climate exposure rather than simply take it on trust. A growing number of jurisdictions are implementing ISSB-based reporting requirements mandating in scope companies to disclose current and anticipated financial effects from the climate risks, where material.

When climate risk fails to show up in the numbers

The climate case illustrates a more general problem. IFRS 9 works by weighting scenarios which are based in some way on past experience and gradual change. Climate risk doesn't necessarily behave that way. It can arrive through a policy shock, a sudden repricing of an asset class, or a physical event that changes the risk profile of a portfolio in one single step rather than a gradual drift. A model built to detect gradual change is poorly suited to a risk that moves suddenly.

Some will argue that the model should already be factoring in all counterparty risk, whatever its source. While that should work in an ideal world, in practice, the industry doesn't yet have a clean way to get non-linear, low-precedent risks into numbers that are, by design, built from precedent.

This is not a case against IFRS 9, which does what it was designed to do reasonably well. It is arguing for the industry to be clearer about what the standard was never built to catch, and about who is doing the work of catching it instead.

Could AI be the next blind spot?

A large share of the equity rally on both sides of the Atlantic over the past few years, rests on a bet that AI will eventually show up in the numbers: in margins, productivity and, in time, in the credit quality of the companies borrowing heavily to build it out. That bet is being priced with real conviction but remarkably little history to back it. Climate risk threatens banks mainly through physical assets and long-dated exposures. A downward turn in AI valuations would be more likely to affect equity valuations and funding for costly infrastructure. The risk would be concentrated in a small number of very large companies.

If AI causes a major earnings shock in either direction, it is fair to ask whether banks' credit models, insurers' reserving and market risk frameworks are any better at spotting it than climate risk. That is before we get to the shadow banking system and the structured debt funding large parts of the AI boom. A system built on past scenarios struggle with risks that have little historical precedent.

None of this argues for tearing up existing frameworks, neither for climate nor for AI. It calls for honesty about which forward-looking models genuinely look ahead and which still rely on the past, despite using scenario language. Potentially, supervisory stress tests may be better suited than accounting standards to explore the paradigm-shift scenarios ECL was never designed to cover.

Can AI help solve the problem it creates?

There may be a twist. AI is already used, with varying accuracy, to forecast the financial health of companies, and if the frontier models are anything to go by, that accuracy will improve rapidly. The caveat is that these models are also trained on historical data, so they may share the same blind spot for low-precedent shocks. Even so, the problem around AI that needs to be addressed may also hold part of the solution by using AI to solve it.

What about non-financial risk?

There is a wider question: how do we capture the non-financial costs of these developments? The potential uninsurable and unbankable, created by AI hyper-personalisation, are second-order effects that sit outside counterparty risk and, I would argue, outside any current measure. If recent events show these risks can threaten society itself, we need to start measuring them, and that means looking beyond the balance sheet of any single entity.

To support members in assessing and quantifying climate risks, ICAEW has entered into a partnership with the Met Office to embed authoritative climate science into members’ professional competence and judgement.