Here's a bet I see businesses make without realising they're making it. They look at what AI can't do today, they build their plans around those limits, and they treat those limits as fixed. It's a quiet bet, and it's a losing one.
It loses because the thing you're betting against keeps moving. Two forces are pulling in the same direction at once, and together they make "wait and see" more expensive than it looks.
The first is capability. Work that was considered out of reach for AI a few months ago becomes ordinary surprisingly fast. The edge of what these tools can do keeps shifting, and it shifts faster than most plans assume.
The second is cost, and the numbers here are stark. Stanford's AI Index for 2025 tracked the price of running a model at the level of GPT-3.5 - a capable, recent-generation system. In November 2022 it cost about $20 to process a million words of text. By October 2024 the same work cost about seven cents. That is a fall of more than 280 times in under two years. Put plainly, the price of using capable AI dropped by more than 99% in that window, and it has kept falling since.
Put those two together - capability climbing, cost collapsing - and the maths of "wait and see" turns ugly. The business that sits back isn't holding steady. It's falling behind a line that keeps rising, and the gap compounds quietly while it waits.
Ethan Mollick, who studies how people actually work with these tools, makes the leadership point well. The job is to keep updating your beliefs about what AI can do and how well it can do it. The common mistake is to decide once that AI is only good for low-level tasks and then leave that judgement untouched for a year. The more honest move is to keep testing the edge, and to treat the tool as a real thinking partner rather than a novelty.
In my own risk work I call this Future Capability Mismatch. You design a system around today's limits, ship it, and within a surprisingly short window it's less competitive than it was on day one - sometimes simply obsolete - because the ground moved underneath it. The danger isn't that you adopt AI badly. It's that you adopt it rigidly, around assumptions with a short shelf life.
So what survives this? Not a single big bet on where AI stops, because nobody can place that bet well. What survives is a posture.
- Build flexible, modular systems you can upgrade a piece at a time, rather than monoliths you'd have to rip out.
- Roll out in small, adaptable steps, so you're never far from a point where you can change direction.
- Keep your team's core AI knowledge current, because the judgement of where to use these tools ages faster than any system you build.
- Plan for more than one AI future. Hold a couple of scenarios in mind rather than betting the strategy on a single guess about where capability stops.
One caveat, the mirror image of the one I made in the companion piece to this. If you've read me arguing that you should prefer boring, rule-based software and distrust the AI's guess, you might think this contradicts it. It doesn't. That argument is about the task in front of you this month - reliability now. This one is about the trajectory - where AI is heading over the next few years. Prefer rules for today's job. Don't build your three-year strategy on AI staying as limited as it is today. Both are true, because they answer different questions on different clocks.
None of this means you should chase every new model or rebuild your business every quarter. That's its own kind of mistake, and an expensive one. It means holding your plans a little more loosely, keeping your knowledge current, and refusing to treat today's limits as permanent. The businesses that do well from here won't be the ones that guessed the future correctly. They'll be the ones that stayed ready to change their minds.
If you're setting direction for your business and you're not sure which of your assumptions about AI have a short shelf life, that's worth a conversation. Hit reply and tell me where you've drawn your lines, and I'll tell you honestly which ones I'd hold loosely.
-- Alastair
P.S. Pressure-testing the AI assumptions baked into a business plan, and designing for more than one future rather than betting on a single one, is the kind of strategic work I do with leadership teams. If that's what's on your desk, my recommendation for the next step is a short conversation about where your plan is most exposed.