Insights
AdaptationPerformAI, Cynefin and the Danger of Making Complex Problems Look Simple
28 August 2026
A few years ago, Steve McCrone and Doug Maarschalk at AGLX introduced me to something that has stuck with me ever since: the Cynefin framework.
Cynefin, developed by Dave Snowden, is essentially a sense-making framework. Its value isn’t that it gives you the answer to a problem. It helps you work out what kind of problem you’re dealing with before deciding how to respond to it.
That distinction matters.
In the Clear domain, cause and effect are obvious. There are established answers and repeatable practices. Sense what is happening, categorise it and respond appropriately.
In the Complicated domain, cause and effect still exist, but expertise or analysis may be needed to find them. Think diagnosing an engineering fault or designing an integration architecture. There may be several good answers, but investigation can identify them.
The Complex domain is fundamentally different. Cause and effect only really become clear in retrospect. The system involves people, behaviours and interactions that change as you intervene in it. You can’t simply analyse your way to the right answer in advance. Instead, you probe, observe what happens, learn and respond.
And then there is Chaos, where the immediate priority is to act, create enough stability, and then work out what to do next.
It’s a deceptively simple idea: understand the nature of the situation before selecting the management response.
I think AI is quietly making that discipline harder.
AI makes almost everything look Complicated
Give a modern AI model a difficult business question and it will usually do something remarkable.
It will structure the problem, identify the drivers, find relevant research, develop and assess alternatives, then hand you an implementation plan and, if you ask nicely, a beautifully formatted board paper explaining why it all makes sense.
That is enormously valuable.
For genuinely Complicated problems, AI is becoming extraordinarily capable. It can bring together expertise from different disciplines, analyse quantities of information no individual could reasonably digest, challenge assumptions and explore alternatives incredibly quickly.
The problem is that it responds to Complex questions in almost exactly the same way.
Ask:
How should we restructure this organisation?
Why has growth stalled?
How should this consulting business respond to AI?
What should our new commercial model be?
How do we change the behaviour of our sales team?
And AI will answer.
It doesn’t put its hands in its pockets, stare at the ceiling and say, “Interesting. I’m not sure anybody can know that yet.” It produces an answer, usually a very good-looking one.
That’s the trap.
Some answers don’t exist yet
I’m encountering this increasingly in the advisory work I do.
A business has a problem, and we can give an AI system financial information, customer data, market research, competitors, organisational structure, meeting notes and whatever else we can find. It can synthesise all of that and produce an extremely plausible strategy.
But sometimes the thing we’re trying to discover isn’t sitting inside the data waiting to be found.
Take a technology services business trying to move away from selling people by the hour.
AI can analyse the market, research outcome-based pricing, managed services and productisation, model margins, develop propositions, identify what competitors are doing, and even draft the new commercial offers.
Useful? Absolutely.
But it cannot know in advance how customers will react to those offers.
Neither can I.
Will a customer pay $15,000 a month for an ongoing capability rather than $1,800 a day for a consultant? Will the sales team know how to sell it? Will delivery people behave differently when hours stop being the primary measure? Will customers accept shared outcome risk? What unexpected behaviour will new incentives create?
Those aren’t analytical questions, they’re Complex ones.
We can develop hypotheses about them, and improve those hypotheses enormously using AI, but eventually somebody has to put something into the real world and see what happens.
Fluency can disguise uncertainty
This is where AI introduces a subtle management risk.
Before AI, genuinely difficult strategic questions tended to look difficult.
You commissioned research, people debated the problem, experts disagreed, information stayed incomplete. The uncertainty was visible.
Now I can ask an AI system the same question and receive twenty pages of coherent analysis before lunch.
The uncertainty hasn’t necessarily changed.
Its appearance has.
A Complex problem can arrive back looking remarkably like a solved Complicated one.
That’s potentially more dangerous than a hallucinated fact because a hallucination can be checked. The bigger mistake is selecting the wrong way of thinking about the entire problem.
If leadership believes there is a correct answer available through sufficient analysis, the natural response is to keep analysing until somebody produces it.
And AI is exceptionally good at producing one.
Perhaps we’re asking AI the wrong question
The answer isn’t to use AI less.
It’s to change its role.
In a Complex environment, AI shouldn’t necessarily be the oracle. It can be the probing instrument.
Instead of asking:
What should our new commercial model be?
Ask:
What are five materially different commercial models we could safely test with customers over the next 30 days?
Instead of:
How should we reorganise the company?
Try:
What small changes could we make to decision rights or team structure that would give us evidence about where the real constraint lies?
Instead of:
What will customers want?
Ask:
Design three inexpensive experiments that would allow us to observe what customers actually value rather than asking them hypothetically.
That plays directly to AI’s strengths.
AI makes experimentation dramatically cheaper. It can develop multiple propositions, create prototypes, analyse feedback, identify weak signals, compare results across experiments and help us decide what to probe next.
Work that might previously have made experimentation prohibitively expensive can increasingly be done in hours.
That could make AI extraordinarily powerful in the Complex domain.
Just not in the way we usually describe it.
The answer may be more experiments, not better answers
This is one of the reasons Cynefin has become more relevant to me, not less, as AI has improved.
Steve and Doug introduced me to it before generative AI had become part of everyday business life. At the time, I saw it primarily as a useful way of thinking about strategy and uncertainty.
I now think it provides another useful guardrail.
Before asking AI to solve something, ask: what kind of problem is this?
If it’s Clear, automate aggressively. If it’s Complicated, AI-assisted analysis and expertise can be extraordinarily powerful. If it’s Complex, resist the temptation to mistake a convincing recommendation for knowledge that simply doesn’t exist yet.
Form a hypothesis. Design a safe-to-fail probe. Put it into the real world. Observe what actually happens.
Then use AI again to help make sense of what you’ve learned and decide what to try next.
There is an irony here.
AI gives us access to more analysis, expertise and apparent certainty than we’ve ever had before. Used badly, that may tempt us into treating more of the world as predictable.
Used well, it could do almost the opposite. It could make organisations much better at experimenting, sensing and adapting because the cost of doing so has collapsed.
The important thing is knowing which job you’re asking it to do.
Sometimes you need AI to help find the answer.
Sometimes you need it to help design the experiment that tells you whether an answer exists at all.