When not to use Artificial Intelligence

 Artificial intelligence is at the peak of the Gartner hype cycle[1]. It can pressure teams to take up unreasonable projects. One needs to perform his/her due diligence before embarking on a machine learning use case. Alternative solutions can work out to be cheaper and more efficient.

 


Projects with no timely action

A Machine Learning(ML) project may be brilliant in itself. But unless the business can act upon the results, the model is of no use. Not only should we be able to take action we should also be able to take this action promptly. Here are a couple of examples.

Employee attrition: Consider a team working on a model to forecast employee turnover. The data science team goes to great lengths to collect data from several corners, including social media, browsing habits, and so on. They develop a robust black-box model using this hard-won data to determine whether an employee would quit in the next 30 days. When the HR gets these results, he/she does not know what to do. Unless the HR knows why someone is attriting, they don’t have a means of stoping it. Since this is a black-box model, the data scientists have no idea why anyone would attrite. Instead of being awed by the marvel of their prediction, HR asks them, “Why are they leaving?” and without having a good answer to the question, the data scientists think, “What we give is never sufficient”

Telecom churn:– In one of our projects, we predicted churning internet users three days ahead of the event. It was already discussed and decided with the marketing team that we’ll be sending out non-transferable discount coupons to these customers at risk. These discounts would be scattered over the next few months, incentivizing them to stay with the provider. We also did a cost-benefit analysis, and the benefit of retaining them overweighed the costs of the discounts. However, there was one thing we had overlooked. These discounts were to be sent by physical mail (not e-mail). It could take up to 3 days for these discount coupons to reach them. The ETL (Extract Transform Load) and data preparation were already taking two days. As a result, our coupons arrived a little too late to avoid churn. Although we fixed this with a lot of efforts, this served as a good lesson for me. If you cannot take timely action on your model predictions, it may cause more harm than good. In some cases, it may not be practical to improve on the time-performance of the models.

Projects with better alternatives

Every company wants to use Artificial Intelligence, and business is eager to come up with more use-cases. Some use-cases, particularly those developed/thought of without feedback from those who might need it, can end up being overkill.

Video analytics to detect industrial fire: I was briefly pulled-in for some advice on a project similar to this. I don’t have the full details of the project. I am just using this as an example of what not to do. From what I understood, the client/prospect wanted to launch drones, capture videos and identify when there is a fire accident. To do so, they needed to find fire in the image/video.

First, real-time video analytics is expensive. Every day, drones will send terabytes, if not petabytes, of data. This data must be labelled, stored, processed on the cloud. It’s also costly to train deep neural networks models with large datasets. The cost will vary depending on the facility’s size, picture resolution, and other factors, but it will cost at least 25,000 USD per year to keep this project going. At a one-time cost of a few thousand USD, you can instead buy thermal cameras or IR pyrometers, mount them on the same drones, and trigger an alert when the temperature detected is above a given threshold. No AI is required. Images are ineffective at detecting flames. Some fires are orange, while others are blue, and some aren’t visible. How do you label such data? Let’s say you train your model on orange flame, and your fire accident starts with a blue flame or a green flame, then your model will never be able to catch it. After burning all the cash on AI, your facility could be in ashes too.

Video analytics to detect if passengers are using a seat belt:  A friend I was speaking with some time ago came up with the Idea that flights should install cameras and use image recognition to identify if they are wearing their seat belts. This is yet another instance in which AI is overkill. You only need a 15 cent tamper-proof sensor on your seat belt buckle.

There are many such cases where we pick a bazooka to kill a fly. Mostly these scenarios are inadvertent or come to light after the fact. Engaging consulting teams with the right expertise can help overcome such scenarios.

 

Comments

Popular posts from this blog

Virus: Discussing macroscopic possibilities

Could Generative AI pose an existential threat to us?

Advantage; Homeloan