The Wayfinding Problem: Why Better AI Still Won't Tell You What to Do Next

There is a version of the AI story that goes like this: AI gets better, humans become less necessary, the gap closes until machines can do everything that matters. This version is popular in certain corners of the internet, occasionally useful as a provocation, and largely wrong as a description of what is actually happening.

The version that better fits the evidence is less dramatic and more interesting. AI is getting extraordinarily capable at a specific class of tasks: synthesis, pattern recognition, research, analysis, first-draft generation, anomaly detection, prediction within well-defined domains. These capabilities are genuine, they are improving faster than most people expected, and they are already changing what it means to do knowledge work.

What they are not doing, and what there is good reason to believe they cannot do in any near-term timeframe, is navigate.

Not navigate in the GPS sense. Navigate in the older, deeper sense: to find one's way through genuinely uncertain territory, where the map is incomplete, the destination is not fully specified, the conditions are changing, and the choices made will shape the terrain for those who come after.

That capacity is not a feature that will be added in the next model release. It is something different in kind from what current AI systems do, and understanding why matters enormously for anyone responsible for leading an organisation through the next decade.

What AI Is Actually Good At

To make the argument properly, it helps to be precise about what the current generation of AI systems does well.

Large language models are, at their core, extraordinarily sophisticated pattern-matching and synthesis engines. They have processed more text than any human could read in a thousand lifetimes, and they can retrieve, combine, and express the patterns in that text with a fluency that often exceeds what a human expert would produce in the same time. For tasks where the relevant knowledge exists in text form and the quality of output can be assessed against that corpus, they are formidable tools.

This covers more ground than most people realised two years ago. Legal research. Medical literature synthesis. Financial analysis. Code generation. Strategic scenario modelling, in the sense of identifying precedents and analogues for a given situation. Competitive intelligence. Most of what junior and mid-level knowledge workers spend most of their time on.

The economic implications of this are significant and are already being felt in hiring patterns, billing structures, and the distribution of value across knowledge work industries. The BCG and WRITER surveys cited in a recent post on this site reflect, among other things, the organisational turbulence that follows when a general-purpose capability removes the rationale for significant layers of a workforce.

So the capability is real. The question is where it ends.

The Problem With Synthesis

The answer to "what should we do?" is not, in most consequential situations, a synthesis problem. It is a navigation problem.

Synthesis produces an accurate, well-organised representation of what is known. It can surface the relevant options, model the likely outcomes of each, identify the considerations that bear on the choice, and express all of this with clarity and apparent authority. What it cannot do is tell you which option is right, because rightness in a navigation context depends on factors that cannot be fully encoded in any corpus.

It depends on what you value. Not in the abstract, but specifically: what your organisation is actually willing to sacrifice in pursuit of what outcome. These commitments are rarely fully articulated, often inconsistent in practice, and only become clear under the pressure of a real decision with real costs.

It depends on the relational context. Who else is navigating in this space, how will they respond to your choices, what does the history of your relationships with them make possible or foreclose. This is not information that exists in any retrievable form. It is knowledge that lives in the heads and relationships of the people involved.

It depends on what kind of future you are trying to build. Not the most probable future, which is what forecasting gives you, but the future that reflects what you believe is worth building toward. That is a values question dressed as a strategy question, and no model can answer it on your behalf.

And it depends on the wisdom to know which of your existing frameworks applies to this situation and which ones don't, because the situation is genuinely new in ways that matter. Pattern recognition is the identification of similarity to prior cases. Wisdom is knowing when the prior cases are misleading.

The Wayfinding Thesis

The framework I find most useful for thinking about this comes from the work of Richard Hames, futurist and strategic adviser, and his collaborator Marvin Oka. Their Wayfinding thesis, developed over years of work with governments and organisations navigating genuinely complex transitions, holds that as AI commoditises the bottom of the cognitive stack, the most consequential human capacity becomes not intelligence in the analytical sense but wayfinding: the ability to orient oneself in genuinely uncertain territory, maintain integrity under pressure, and make consequential choices in situations where the models offer no reliable answer.

The Wayfinding framing is useful precisely because it reframes the question. Most AI discourse asks: what can humans do that AI cannot? The Wayfinding thesis asks: what is the capacity that becomes most valuable as AI takes over everything it can take over?

The answer is not creative thinking, which AI is increasingly capable of in the generative sense. It is not analysis, which AI handles better than most humans on most well-defined problems. It is the navigational capacity: to hold a values orientation steady under pressure, to read relational and political terrain that exists nowhere in a database, to know when to break the rule rather than follow it, and to make decisions that you can stand behind when the outcome is not the one you anticipated.

This capacity cannot be built through information. It is built through experience, reflection, rigorous challenge, and the kind of thinking-alongside that good mentorship provides: someone who knows your context, pushes on your assumptions, and asks the question you were avoiding.

Why This Changes the Value Equation

If the Wayfinding thesis is right, then the strategic implication for leaders and organisations is more specific than "invest in human skills." It is: invest in the specific capacity for navigational judgment, because that is the layer that is becoming scarcer and more valuable as everything below it automates.

The organisations that will be in the best position in ten years are not necessarily the ones that adopted AI earliest or spent most on AI tools. They are the ones that understood which human capacities to protect, develop, and position as their genuine competitive advantage, and built the structures to do that deliberately rather than by default.

For individual leaders, the equivalent question is: am I developing the navigational capacity that will be most valuable in the environment I'm moving into, or am I spending my development time on capabilities that AI is making progressively less scarce?

These are not comfortable questions to sit with. They require a degree of honest self-assessment about where one's actual value comes from that is easier to defer than to confront. But the leaders who have confronted them, who have a clear-eyed view of which of their capacities are AI-substitutable and which are not, and who are actively building the latter, are operating with a strategic clarity that is relatively rare and increasingly valuable.

The Adviser Problem

One practical implication of this analysis is that the kind of support most valuable to a leader navigating genuinely complex terrain is not more information. It is access to thinking that challenges their framing, pushes on their assumptions, and helps them develop the navigational capacity rather than just providing better maps.

That is a specific and relatively scarce thing. It is not a research subscription or a strategy framework. It is a relationship with someone who understands the terrain, has navigated genuinely difficult situations themselves, brings a coherent and tested intellectual framework to the conversation, and is willing to be honest rather than agreeable.

The scarcity of this kind of thinking partner, relative to the number of leaders who need it, is one of the more interesting strategic problems created by the current moment. The leaders who have access to it are better equipped to navigate than those who don't, and the gap between them is widening as the terrain gets more complex.

The question of how to make that kind of thinking more accessible, at the quality level that actually makes a difference, is one I find genuinely interesting. The answer is not simply "AI can be your adviser," because that collapses the distinction this entire post has been trying to draw. But it is also not simply "wait for a scheduled engagement with a human expert," because the moments when navigation is most needed rarely wait for a calendar opening.

What sits between those two options is worth building toward. The capacity for on-demand access to genuine strategic thinking, grounded in a real intellectual tradition, available at the moments when it is actually needed, is one of the more interesting problems at the intersection of AI capability and human judgment right now.

The intersection of foresight, strategic navigation, and what AI genuinely cannot do is a thread I'm exploring in my work with Richard Hames. If this framing resonates and you want to continue the conversation, I'd welcome it.