When to use AI
Your AI product idea is probably a bad one. This shouldn’t come as a surprise. Bad product ideas greatly outnumber good ones. If you want to avoid bad AI product ideas, you should first look for what makes good product ideas good.
Good product ideas solve a problem. More than that, good product ideas solve a problem for specific people. If the problem is sufficiently important, or if these people want badly enough to have the problem solved, they’ll be willing to pay. If these things are true for your idea, you might have a good AI idea.
So let’s assume that you have a problem worth solving; how then can you tell if this is a good AI problem? Start by first figuring out if it’s a good problem for software. Why? Because AI is implemented in software. Almost always, AI is a bad first attempt to solve a problem. Software can do a lot of what AI attempts to do, and it does so with fewer constraints. If, however, this is a problem that is suited to software, but software can’t solve it, you very well might have a good AI problem.
Taxonomy of ideas
Expect Good AI ideas to be vastly outnumbered by bad ideas.
Software problems
The first rule of thumb to determine if you have a good software problem is to ask, does this problem happen frequently? Things that don’t happen often are very rarely worth the investment needed to develop software. If you have a problem that happens thousands of times, you might have a good AI idea.
The second way to figure out if you have a good software problem is to ask, does this require scale? Software is exceptionally good at scaling. Scalability is why software is eating the world. Scalability is what allowed Uber to disrupt taxi markets around the globe and allowed Instagram’s 6 engineers to build a space for 30 million humans to connect and share memories. If you have a problem that requires scale, you might have a good AI idea.
The third way to figure out if you have a good software problem is to ask, does this require speed? Software is fast. It is faster than physical machines. For example, email is infinitely faster than sending a letter. For many tasks, software is faster than the fastest humans; high-frequency trading of equities takes place at a pace imperceptible to human traders. If you have a problem that requires speed, you might have a good AI idea.
The fourth rule of thumb to figure out if you have a good software problem is to ask, is there a pattern here? Software is automation, and automation is about doing the same or similar thing, again and again. What makes software truly useful is the ’similar’ in that last sentence. Being able to tell the difference between situations and reacting accordingly is a core part of what software does. Software uses data to check if some set of conditions are present (e.g. GPS location says you’re in Tibet) and based on this data, the software 'does something’ (e.g. it tags your Instagram photo with ‘Tibet’). If you have a problem that has a data pattern, you might have a good AI idea.
Non-obvious patterns
Data is a digital representation of conditions in the real or digital world. But the word ‘Data' is misleading because it can mean so many things. Having data is not enough to justify the use of AI. Only some patterns in data lead to good AI problems. So if AI is appropriate for some patterns in data, but not for others, which patterns should you rule out? Easily explainable patterns. Why? Because easily explainable patterns can be written down, and things that can be written down can be written down in software. If it can be coded, it is faster and cheaper to forget about AI. So we’re interested in non-obvious patterns; patterns that are difficult to explain.
Patterns that are difficult to explain are more common than you’d think. Most people can easily tell the difference between a picture of a cat and a picture of a dog. In fact, you’ve probably never confused a cat for a dog, or a dog for a cat. But if you tried to explain that difference to your friendly neighbourhood alien you’d quickly run into trouble. This isn’t at all obvious. Definitions are incredibly slippery things and our brains are rarely confronted with the need to be highly precise with our definitions.
You may for example say, dogs are bigger than cats. That's true on average. But my friend's cat, Mr. Fancypants, is bigger than any chihuahua I've ever met. You might say cats are fluffier than dogs; Google ‘Keeshond’, ‘Pekingese’ or ‘Lhasa Apso’. You might say cats have pointy ears; so do Huskies. Or you might say cats have longer tails; take a peek at a Great Dane. So there are always edge-cases. And there are more edge-cases than you may initially suspect. The point is that these aren’t even really edge-cases - it is our definitions that make them so. Most of the world is not immediately translatable into simple, statable rules that are accurate even 90% of the time. If you can’t explain it, you can’t code it. If you have a problem with decisions that are difficult to explain, you might have a good AI idea.
So AI might be a good idea for making decisions on patterns that are difficult to explain, where those decisions require speed and scale. Let's now consider how to make those decisions. Data is required for any AI problem. Almost always, lots of data is required. Do you have, or could you easily get, 10,000 to 100,000 examples of the decision you’re trying to automate? Because if you have this data, or could easily get it, you may well be on your way to a good AI idea.
Data
This data comes in two types. The first is the data used to make the decision - pictures of cats and dogs in the example above. If you have data, this is likely the type of data you have. The second type of data are the labels that tell you what decision should be made. In the example, this would be labels telling us whether there’s a cat or a dog in our picture. This data is more rare. If you have these two types of data, you very well might have a good AI idea.
This may seem like a stupid amount of effort just to figure out if your AI idea is a good one. But the potential benefits of a good AI idea are huge. One way to see this, is to ask yourself if a human could solve your problem in a few seconds or less. If so, you very well might have a problem that is solvable with AI. If you can use AI to replace human decision-making you can replace those humans and reap the reward of their would-be labour. If your two-second task requires expert decision-making you may be on a better path yet. Experts tend to be expensive and by definition there aren’t very many of them. If you can match the decision-making accuracy of these experts you can make their service available at a fraction of the cost. In Silicon-Valley-speak, you can democratise access to it.
What makes good AI ideas insanely good is that they have certain characteristics that aren’t present in many forms of software. Democratising access to a service creates a new market, and if you’re smart, one that you can dominate.
— — —