The risk territory has moved

With this level of disruption, causal pathways are harder to establish and investors lean more heavily on shared narratives to make sense of markets. Narrative-driven markets concentrate risk in ways that traditional attribution frameworks are not designed to detect. The first casualty of AI-era uncertainty is traditional diversification. On paper, a typical portfolio can look well-diversified: equities across sectors and market capitalization, credit across ratings and geographies, with increasing allocations to private markets. However, in practice many of these positions likely share an underlying exposure to the same disruption pathway.

Consider the AI compute stack. Demand for data centers is driving energy infrastructure investment. It is reshaping semiconductor supply chains, affecting commercial real estate in ways that are, for now, difficult to model cleanly and introducing labor market pressures that feed into wage inflation in specific geographies and skill pools. A portfolio that holds utilities, logistics real estate, semiconductor equities and investment-grade credit linked to technology capex may look diversified by asset class label. And yet when you step back, these positions retain a shared exposure to a common AI build-out pathway.

The second casualty is the stability of the correlation matrix itself. The relationships between asset classes that underpinned decades of portfolio construction – the negative correlation between long-dated government bonds and equities, or the diversifying properties of commodities – were always conditional on a particular macro regime. Equally, energy transition and AI energy demand are in tension, labor displacement and consumer spending are linked in ways that differ by geography and income cohort. It’s not just new correlations that are appearing, but the possibility of strengthened, weakened or lost correlations altogether as the economic pathways beneath them evolve.

In the years before the GFC, the models invariably told a coherent story. Banks had stress tested their books against a range of historical scenarios such as the savings and loan crisis, Russia’s default or the dotcom bubble. But none of those scenarios contained a nationwide housing price decline as a base case. What happened in 2008 was unmeasurable because the historical stress tests were rigorous applications of the past. In effect, they asked the wrong question.

AI-era portfolio risk management faces an analogous challenge. Historical stress testing is a backward-looking exercise. The implicit assumption is that the future will rhyme with the past; that the distribution of outcomes is stable enough for historical conclusions to constitute useful evidence about what is coming. This assumption is weakest precisely at moments of macro-structural transition.

Many investors already complement historical stress tests with forward-looking scenario analysis, and such tools are now common in off-the-shelf risk systems. What AI changes is not the existence of these approaches but their fragility: pathways can unfold faster, evidence is thinner, and exposures cut across sectors that still look separate in traditional asset-class frameworks.

The relevant question for our AI-disrupted world is how the portfolio would fare across a range of plausible AI-driven pathways that have no historical precedent. What happens if the energy-compute constraint bites harder than anticipated? What if AI-driven productivity gains arrive unevenly, widening sovereign debt dynamics in ways that pressure bond markets? What if the disruption plays out at Citrini-speed – fast, sharp and concentrated – rather than diffusing gradually?

The diagnostic must shift from historical scenario replay to forward-looking pathway analysis. Not a single base case with upside and downside variants, but a new map of structurally distinct futures, and a portfolio assessment of how it performs across all of them.

There is a temptation, when confronted with a more uncertain environment, to respond with more sophisticated forecasting. Better models, more data and finer-grained scenario analysis. However, the appropriate response to genuine uncertainty is robustness, not precision. Rather than relying on one expected path and optimizing around it, investors can treat forecasts as conditional inputs, test their sensitivity and consider the consequences of being wrong. The goal is to build a portfolio that survives and performs across a wide range of plausible environments, while retaining the ability to adapt as evidence develops.

How can investors know if they have a narrative-exposed concentration? Classic correlation matrices can show that assets have moved together. They cannot, however, explain why they did so or whether the relationship will persist. In a period of structural change, investors need to examine the economic pathways connecting their holdings.

Investors could start by mapping exposures to underlying drivers using a causal, pathway-based view of correlation rather than one that treats it as a fixed statistical property. I propose three diagnostic questions to get this thinking started:

This diagnostic test is not intended to produce a perfect forecast of AI’s consequences. Rather, the purpose is to reveal where the portfolio may be dependent on a common belief about the AI future, and assess whether it can remain resilient if it proves incomplete.

This reframes three pillars of portfolio construction. Diversification ceases to be an exercise in filling asset class buckets and becomes an exercise in identifying true sources of return and ensuring they are substantially uncorrelated at the level of underlying risk drivers, and not mere labels. Optionality – the capacity to reposition as pathways clarify – becomes a portfolio asset in its own right, replacing the need for better forecasting. And outcome orientation replaces benchmark optimization as the organizing principle. 

Many investors will anticipate where this train of thought is headed: The total portfolio approach (TPA). Although its harshest critics write it off as marketing spin, it represents an operational and mindset shift. TPA is an investment framework that starts with desired outcomes and risk budgets and blends together all exposures. It was developed (and then codified by Willis Towers Watson) precisely because traditional asset-label diversification was understood to be insufficient. It’s well-suited to AI-era uncertainty as it reframes the question from 'where should we allocate?' to 'which risks are we paid to hold?'

In the AI era, full of unknown unknowns, this question becomes more urgent and more consequential. TPA considers the portfolio in risk units rather than allocation units. It can make hidden correlations (the kind that emerge when assets share AI-disruption pathway exposure) more structurally visible. And it provides a framework for dynamic, factor-aware allocation that can evolve as pathways clarify, rather than locking the portfolio into positions calibrated to a regime that may no longer hold.

Crucially, TPA is architecturally suited to pre-mortem thinking. Rather than constructing a portfolio and hoping it survives stress events, it begins by hypothesizing the ways a portfolio can fail and builds the response protocols in advance. TPA is, of course, not the sole possible response to the diagnostic test set out above, but it is one credible, and potentially more resilient, way to act on it.

But visibility is only as good as the judgement of those who use it. An investor can adopt TPA in full and still misread the AI shift entirely. The same was true in 2008: it was not stress-testing methodology that separated winners from losers, but who correctly judged that the housing bubble was ending and found the instruments to act on that view.

Andrew Ang and others have argued that the future of sophisticated portfolio management looks less like traditional asset management and more like a self-driving system: one that continuously recalibrates in response to incoming information, holds a model of its own uncertainty and is designed to navigate a range of conditions rather than to optimize for one.¹ TPA shares the same underlying logic of continuous recalibration under uncertainty. It is one institutional expression of that principle. 

The rise of AI does not invalidate the fundamental logic of taking risk in exchange for return. Instead, it changes the conditions under which that logic applies, and, in doing so, it raises the cost of intellectual rigidity.

Investors who treat AI as a sector theme will allocate to data centers and semiconductor makers and consider the job done. Those who treat it as a macro-structural shock are forced to ask harder questions about the stability of the correlation matrix, adequacy of historical stress tests and whether their apparent diversification is masking concentrated exposure to a single disruption narrative.

Korzybski’s crucial warning was that confusing the map with the territory is dangerous. The investors who will navigate the AI era most effectively will heed this caution. They will remain perpetually aware that the territory is constantly moving. And although the total portfolio approach does not claim to chart what cannot be charted, it does at least build a portfolio that can function when the old map becomes obsolete. In an age of fundamental structural uncertainty, this type of investment humility offers perhaps the most durable edge available: the wisdom to know what cannot be known, and to position accordingly.

¹ Ang, Andrew, Nazym Azimbayev, and Andrey Kim. “The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management.” Working paper, 2026.