Top 5 non-practical applications of AI

There is an overwhelming proof that some problems can now be effectively solved with machines: face recognition (on images), voice recognition (without accent or group speaking), fraud detection (certain kinds). In fact you will notice that these use cases often come with small (and sometimes not so small) qualifiers.

So it is not that hard to understand why it is difficult for companies to get started with AI or Machine Learning. Because when technology solves some problems, it's all to easy to try to apply it too broadly.

Here are the top list of things to avoid:

1. "It already works in X, let's quickly expand it to Y". Expanding market or process coverage with technology is often an easy, linear thing to do. Make sure that it really is an expansion, not a brand new initiative. Courtesy of XKCD:
2. "Let's get 90% automation in our mature, already fully optimized operation". If there is no known room for improvement, Machine Learning likely is not the first thing to consider. Change management is.
3. "Can it to replace these 3 people we have in this fringe process"? Machine Learning is not about sweeping things that do not matter under the carpet. If it's not a large problem worth solving, likely it won't be worth the investment of time.
4. "This one report from last week was a big pain for me - can we automate that"? If the problem is not repetitive, it won't have enough data for Machine Learning to work. It probably is also a problem not worth solving with the new approach.
5. "I don't know how they do it , but can you do it with machines"? Using Machine Learning to dis-intermediate vendors or "automate what they do with all these known data" can be tricky. Try to avoid replicating process where vendors have known economies of scale and scope - they likely invested a lot of money and time in what is often an idiosyncratic process.

Originally published on LinkedIn.