AlphaGo and the shift in what we expect from AI

Last week Google DeepMind's AlphaGo won its match against Lee Sedol - one of the strongest Go players in the world. The final score across five games: 4-1 in favour of the program. Go has long been considered a game beyond machines - the space of possible positions is too vast, the play requires too much of what people call intuition, something thought impossible to formalise.

Apparently it is not.

I want to talk not about the technical details of AlphaGo, but about what this result means for how we think about AI in a business context.

The boundary has moved

Until recently the practical AI conversation was mostly about recognition: images, speech, text. These are classification tasks - the system looks at something and says which category it belongs to. Useful, applicable, but conceptually clear.

AlphaGo is different. It is not a system that classifies a board position - it is a system that makes a sequence of decisions with long-term consequences. It plays several moves ahead, evaluates positions that do not yet exist, and acts under conditions where the outcome of each move depends on many future moves by the opponent.

That is closer to strategic planning than to pattern recognition.

What this means beyond games

The approach that allowed AlphaGo to win at Go is not limited to games. The same architecture - a combination of deep learning and reinforcement search - applies to problems where you need to make sequential decisions in a complex space of options.

Supply chain management with dynamic constraints. Production scheduling with equipment failures. Pricing under competitive conditions. All of these problems share a structure similar to a game: there is a state, there are permissible actions, and there are deferred consequences of decisions.

This does not mean AlphaGo will replace a production scheduler tomorrow. But it does mean the class of problems where AI can be useful has expanded beyond what many expected.

The window of expectations shifts

Every technology has a "window of expectations" - an implicit sense of what it can and cannot do. For AI that window has been fairly narrow for a long time: good with clearly defined patterns, poor with uncertainty and strategy.

AlphaGo's result shifts that window. Not in the sense of "AI can now do everything" - that is not true and will not be true for a long time. But in the sense of "problems we considered out of reach for automation are worth revisiting".

For a leader, this is a practical question: are there tasks in your business that you have not considered as AI candidates precisely because they require something more than classification?

How to think about this now

A few questions worth putting to your team:

1. Which of our decisions are made sequentially and have long-term consequences that are hard to foresee?
2. Do we have historical data on such decisions - what was chosen, what resulted?
3. Where do we lose the most from suboptimal decisions made under time pressure?
4. Which experts in the company spend most of their time on work that could be described formally?

These questions will not produce a ready project. But they will help identify more precisely where AI's new capabilities might apply - before it becomes obvious to everyone.

The boundary has moved. This is a good moment to revisit the map.