As I was putting Anthropic's latest model, Fable, through its paces yesterday, I asked it who will stand on the podium when the Tour de France reaches Paris in late July. It gave me a fast answer: the favourite, the challenger, and a confident choice for third (let's call him "Remco").

Then I asked a follow-up question - what would need to be true, stage by stage, for Remco to be on the podium in three weeks' time. The AI laid out the conditions across the 18 remaining stages. As a casual watcher of cycling, it seemed like a less confident bet than I'd originally been given, so I asked it to weigh the rider's chances against his nearest rivals, and then asked it to look at its own working and answer again. The confident choice became a coin flip. Remco's chances turned on a single big condition (smoking everyone in a time trial), whereas his rival's hopes rested on many smaller ones (limit losses throughout the race). If I was going to make a decision based on the AI's analysis, I could now make that call based on the set of conditions I thought were more likely - not just on the initial answer that had smoothed these details away.

Now back to business. What does a prediction about a bike race tell us about how to use AI for the decisions that actually matter?

I nearly made this newsletter a simple note about technique: ask the AI for its answer, make it show you what its answer depends on, and have it reassess. But with a sporting analogy this good, it would be remiss of me not to go deeper.

For those who don't follow Le Tour, it is a three-week bike race across France. There is one big winner at the end, and a different winner each day. The dynamics of this race are closer to the questions that cross your desk than it might seem. The race runs in stages that compound. Nobody wins it in the first week, but plenty of riders lose it there, and the effort spent early gets paid for late, often in ways that are hard to predict - not unlike how the first phase of any business transformation shapes what comes next. The riders around the eventual winner aren't all racing the same race, either: most can't win overall, so they chase daily wins and other prizes on the same roads. Your market works the same way - the companies around you are mostly fighting for different prizes than you are, and reading every one of them as a direct rival misallocates your attention.

All this to say that the approach I used to get a better understanding of who takes third on the podium is one you can use to build a deeper understanding of the dynamics at play in the decisions you make. Take an acquisition as an example - it runs in stages that compound, the other players are running different races on the same road, and one unseen liability can end it, in the same way a crash ends a Tour. Instead of just asking for a recommendation, ask the AI what would need to be true, stage by stage, for that recommendation to hold. Push both it and yourself on the conditions you read differently - you know your business and market in a way it does not. Then make it look at its own working and see if the recommendation changes. The technique transfers whole, with the benefit coming from the disentangling of the conditions more than from the recommendation itself.

The conditions that don't transfer from a stage race are equally instructive. Unlike most business decisions, the race resolves - there is a podium in Paris, and we get a definitive answer on whether the prediction - and the conditions behind it - held. The structure of the race is fixed too: 21 stages known in advance, named riders, and no one entering the competition mid-race. I also had no stake in the answer (although Remco wouldn't have been my pick). Real life has none of this. There are real consequences, we rarely know for certain whether we made the "right" call, and the rules and the players can shift underneath us.

This tells us something real about the role AI plays in high-stakes, judgement-heavy decisions. Handed a decision with a dozen moving parts and several players, the useful thing it gives you is not one fused, confident answer - it is the list of conditions underneath it, some carrying far more weight than others. What it cannot, and should not, do is tell you which of those conditions to bet on. A bike race is a passing interest with nothing real riding on it; your decision is not, and there is no podium in Paris to check your call against. The weighing, and the decision itself, stays yours.

So next time a decision or a recommendation lands on your desk with several stages and several players in play - an acquisition, a restructure, a market entry - ask your AI what would need to be true for it to succeed, however you define success. Push on the conditions you know better than the model does, then decide. The most valuable thing you get back usually is not the AI's read on the decision itself - it is the list of conditions. Unlike the bike race, where the analysis was the whole point and stopped there, those conditions do not stop working once you have decided. They tell you what to watch for, what to move to mitigate, and where you might still influence the outcome.

Use AI to disentangle complex decisions, not make them for you. For what it's worth, I'm not convinced Remco will be on the podium in Paris - but now I know which stage to watch to tell me if the AI was right or wrong.