I have previously written about how to think about AI, and it is a theme I hear again and again talking to CEOs, boards and employees alike. What can AI do? How do I think about AI?
Most people associate AI with chatGPT, Claude, <insert favourite LLM> and think in a world of +AI. How do I do what I do today with AI, not how will tomorrow look in an AI+ world. To get into the latter mindset it is useful to think about AI as pattern matching / pattern recognition at scale, across all modalities of text, voice, and vision.
What do I mean by pattern recognition? Well, LLMs are powerful next word predictors, generating “meaningful” texts simply by predicting the most likely next word given the context. Vision neural networks as somewhat similar, given this particular arrangement of pixels this is more likely to be a dog than a cat. Voice, this particular wave-pattern is associated with this particular phonetic pronunciation of this word, which gives you transcription.
Now let us take this notion, that everything that is a recurring pattern is something an AI would be able to do given enough relevant training material and compute. What then would be everyday tasks AI is suited for?
Regulatory compliance, matching specific regulations to specific work processes is largely about pattern matching.
Coding software, smaller atomic pieces of operations (boiler plate) is stitched together to create a more complex set of operations
Data cleaning, there is a finite set of types of data (string, numbers, etc) across a limited set of likely entities. That means I have a finite set of combinations and operations to go from messy data into clean data.
Just look at your average day and see how much is pattern recognition. Setting the breakfast table, we are 4 people that means 4 plates, we have milk, cereal, bread and lasagne in the fridge what is the most typical things to eat in the morning. Most of what we do are pattern recognition.
Now where this type of recognition becomes even more impressive is in AI agents. Now think about combining pattern recognition with agency, meaning I can take action to learn about the world. From the breakfast setting example. I have a “function” to get the time of day, I have another function for getting information about the number of seats at the table, the contents of the fridge, maybe even about any dietary restrictions of the people about to eat the breakfast. No there is no “intelligence” involved, it is all a set of most likely actions to string together, pattern matching against time of day, most likely items to eat for breakfast, filtering out anything that might go against the dietary restrictions.
You can now expand on this agent system, and think that each agent only has a narrow set of pattern matching capabilities and functions. So you need to string them together into a workflow or a team. Building on our example, we might have the pack your lunchbox agent (with some of the same functions as the breakfast agent), the clothing agent with access to weather forecast and wardrobe, the route planning agent, the find my shoes agent with inquisitive functions for asking where did you put them last?
All of a sudden you see how entire complex actions consists of narrow instructions with very specific forms of pattern recognition, and a limited set of operations or functions to gain more information about the world.
This is what AI can do – anything that is a well known process of relatively high statistical certainty and a limited set of instructions and actions can be done in an agentic AI world.
So what is left? The emotional connection to patiently explain the importance of getting to school on time even if you are tired, the ability to jump the tracks and break out of the set routine if the car won’t start, the humorous explanation of why lasagne is actually a decent option for breakfast when you are all out of milk and cereal.
In a pragmatic sense, AI and agents can deal with any pattern that is routine and does not leave a lot to ambiguity. If we can leave subjective judgements, value-based weighting of pros and cons, and emotional connection out of our application of the pattern recognition I would say AI could do it.