“When AI should be used” Framework
Many executives now are bewildered by the word “AI”. They think that everything should be made into AI. Before we have the term “There is an app for that”, nowadays, it seems that the new term is evolve into “there is an AI for that”.
Do you have a sales lead problem? There is an AI for that.
Do you have a recruitment problem? There is an AI for that.
Do you have a child discipline problem? There is an AI for that.
Do you have a problem driving your car aka too lazy/drunk to drive? There is an AI for that.
Do we have an issue differentiating cats from dogs? There is an AI for that.
The reality is, not all problems can be solved using AI. In fact most of the problems we are facing doesn’t need any digital intervention at all. (I would argue, digital is the problem, but that’s for another story).
I know it is tempting for Product Manager to shove AI into anything, but we need to understand that, AI is expensive, complex, and potentially carries too many risks if not done properly.
So these are rule of thumb of a good task to be augmented or replaced by AI:
- When human expert could perform a task in few seconds. For example: QC staff inspecting defect on product, Customer Support staff forwarding right ticket support, Merchandise specialist checking correct product placement on the rack.
AI, at least for now is not yet suitable for a task which needs a long analysis, planning, and thinking, for example: Writing down your product roadmap, preparing reports to your boss about why your sales did not make it last month or analyzing the alternative history of what would happen to San Francisco if Nazi wins the WWII.
2. Difficult or impossible to write down rules. We can easily write down simple rules to grading.
If the score is above 90 then the grade is A
if the score below 55 then the grade is D
But how can we write down rules on how to recognises vegetables? Even on simple image recognition task such as recognising cucumber versus zucchini, we would have a thousands of rules.
If we only have a few rules we can simply just use tradition algorithm. But if you have several thousands rules, just build it using Machine Learning.
3. Easy to get examples aka labeled data. A good Machine Learning model needs a GPU, Data Scientist to build the model, Data Engineer to launch the model into production and Data Analyst to sort and clean data. But we need a good and lots of data in the first place. So if we don’t have this, then forget about AI.
4. A good company culture which value the experiments are an investment. Building an AI is by nature, a research, with so many uncertainties. So many AI initiatives failed because the company does not have the mentality of experimenting and taking the risk.
So now, think again. Is AI really needed to solve that particular problem ?
