Posted on 

 by 

 in , , ,

Strategy as simulation

It’s a pretty sure bet to say that AI simulation and modelling is going to play a much larger role in the future of strategy. You might even say that over the next few years, simulation will move from the edges of the strategy process to somewhere close to its centre. We’ll have the ability to test hundreds of variations of a plan, a proposition, or a campaign against modelled markets or synthetic audiences to help us understand the option or combination of options which will give us the optimal probability for success. The outputs of this modelling will be good enough that it will start to feel like a responsible choice to rely on them. Why take a gamble when you can run a scenario a thousand times over?

I’m a believer that AI-supported scenario modelling is an excellent way to test or open up our own thinking. But I think there is a big risk inherent in the scaling of simulation as strategy. If we’re not careful, the use of sophisticated AI modelling won’t just bring speed an efficiency, it may just get everyone to the same answer faster.

In his book Blink Malcolm Gladwell tells the story of how, in October 1994, Herman Miller launched a new office chair that was designed by Bill Stumpf and Don Chadwick. The chair cost $1,000, had no upholstery and looked like nothing else on the market, which was rather the point. Chadwick believed that most office chairs were boring because they were so predictable. In testing, the Aeron scored well for comfort but abysmally for looks. People said it looked like garden furniture, or the exoskeleton of a giant, prehistoric insect. Some even assumed that they were looking at an unfinished prototype.

But rather than go back to the drawing board Herman Miller launched it anyway. The sales were sluggish to begin with but by the end of the decade they were growing by 50 to 70 percent a year. The Aeron went on to become the biggest selling chair in the company’s history and on of the most imitated office chairs of all time. A chair that was once described as ugly became thought of as a beautiful as well as functional piece of design.

Gladwell argues in his book that our reactions to genuinely new things are unstable, which often means that new and different ideas tend to be the most exposed in market research. One of my favourite examples of this reaction is the BBC Comedy script editor who, in 1974, told his bosses that the pilot script he’d just read was ‘a collection of clichés and stock characters which I can’t see being anything but a disaster’. Fortunately his bosses ignored him and Fawlty Towers went on to become an icon of British comedy. Our methods for judging ideas, it seems, are structurally poor at telling the bad from the merely unfamiliar.

Synthetic audiences are a model of how people have behaved or responded in the past, built from data about the world as it already is. If you showed a modern day Aeron chair to a synthetic audience you’d get exactly the results that the Herman Miller team got in 1994 but mauch faster and in way that would look like a universal rejection. There’s also growing evidence that synthetic audiences flatten the spread of opinion. Researchers at Bocconi University compared five leading models with large human survey samples from the US and China and found that the variance of the model’s answers was sharply compressed (in some cases to less than half that of the human responses). They warned that such simulations may overstate consensus and understate trade-offs or contested areas. Other studies report the same finding that models reproduce human averages whilst underrepresenting the diversity of human responses. The pattern is that a simulated audience pulls towards the mean, and the mean is where new or uncoventional ideas go to die.

This presents a risk for individual organisations but it’s also one that potentially compounds across a market. In his 1996 essay What is Strategy? Michael Porter argued that operational effectiveness, performing similar activities better than rivals, isn’t strategy (and yet so many ‘strategies’ are just this). Best practice, he said, is easily copied so competing on it produces absolute improvement whilst also giving relative improvement to no-one. The more companies benchmark against each other the more they start to look like each other and the more indistinguisable they become.

Simulation risks doing this to strategy. Competitors using similar models, trained on similar data, will tend to converge on similar optimal answers. And an optimum that everyone shares soon becomes the new baseline. In a 2021 PNAS paper Jon Kleinberg and Manish Raghavan showed that when many decision-makers use the same algorithm the overall quality of their decisions can fall, even if that algorithm is more accurate for them in isolation. Researchers Anil Doshi and Oliver Hauser found that writers given story ideas by an AI produced work that was actually rated as more creative and better written but the stories also ended up more similar to one another. Individually AI can help us to augment our ideas but collectively the range of ideas actually narrows.

In the early nineties there was a particularly fascinating related study that predates AI by three decades. James March’s research into the relation between exploration of new possibilities and exploitation of old certainties in organisations was itself built around modelling – individuals in an organisation learn from a shared code its collective beliefs about the world, while the code in turn learns from whichever individuals happen to be right. Counterintuitively perhaps, the finding was that fast learning isn’t always good. When people adopt the code rapidly, the organisation converges and performs well in the short run, but it loses the diversity of belief that would allow the code itself to improve. Slow learners on the other hand, kept diversity alive long enough for the organisation to discover where it’s wrong.

In this sense simulation may of course be the fastest learner we’ve ever built. Its whole purpose is to converge rapidly on the best-performing option given what we already know. That can be useful, but its a process that is at risk of filtering out oddly shaped or unflamiliar ideas.

So what to do? Simulation can still be very useful for finding the centre, as long as we remember that this could well be where the crowded ground is. Deliberately protecting some slow learning, or ideas that didn’t score as well but which are championed by people with good judgement, will be important in maintaining differentiation. And being clear about the jobs that should stay with people, like deciding what the simulation should optimise for, recognising when an idea is unfamiliar rather than bad, and taking responsibility for when to overrule the model. The decisive moment in the Aeron story is when Herman Miller decided to launch it anyway, against the verdict of its own research. As simulation becomes more capable, more trusted and more ubiquitous there’s a distinct risk that this kind of decision will become harder and harder to make.

A version of this post appeared on my weekly Substack of AI and digital trends, and transformation insights. To join our community of over thirteen thousand subscribers you can sign up to that here.

To get posts like this delivered straight to your inbox, drop your email into the box below.

Image: By Brooklyn Museum – Creative Commons , CC BY 3.0

Leave a Reply