Building a complex analytical model is akin to crafting a cunning plan. One which anticipates all the possible setbacks and has a contingency for one's contingency plan – then doing absolutely nothing with it. An analytics or AI model, like a plan, is only useful if you implement and effectively use it. The chasm between sculpting an elaborately accurate model and having that same model provide insight for your organisation is where intelligent decisioning adds value.
The last mile
Organisations are under enormous pressure to deliver informed and personalised decisions swiftly. Whether it is a bank detecting fraudulent activity on credit card transactions, an insurer finding the optimum policy premium or a retailer deciding the next best offer for customers, that requirement is here and growing.
Recent years have brought an explosion of AI and machine learning techniques to these organisations, both proprietary and open source. Whilst these vastly improve the accuracy of the models, the expected benefits from a superior model do not always translate in practice.
For example, a supermarket needs to make a next best offer decision about a mobile voucher for the customer whilst they are doing their shopping. It is easy to build an accurate AI model that predicts the likelihood a customer will respond positively.
The hiccup occurs when trying to take the model and apply it to real-life data to make decisions. Should the store offer Mr Smith a buy-one-get-one-free voucher on detergent? Or 25% off on a bottle of merlot whilst he’s doing his weekly shopping in real time? This problem, or chasm, is also known as the last mile. According to Gartner, only 50% of such clever AI models ever get deployed, put into practice and used to make decisions about Mr Smith getting the right offer. Furthermore, of those 50%, most – 90% – take over three months to yield decisions.
What’s the problem
Why is there a problem getting those decisions made? Why do organizations deploy only 50% of their models? And why does it take months to get going? We have established that building a model ascertaining the likelihood of a positive response is not a problem. In fact, this is getting increasingly sophisticated. So what’s the obstacle? Broadly speaking I think there are three issues.
1. More decisions
Previously fewer decisions were needed, hence the method of deployment was appropriate. Alas, times have changed. And as with much of modern life, there is a need not just for more decisions, but answers needed right now, whether it is in real time or in stream.
2. More intricate AI models
The availability of more complex models has been a double-edged sword. Whilst they increase the accuracy of predictions to make decisions, the downside is the difficulties when deploying models of a more intricate nature.
3. More complex decisions
It is not just a case of more decisions are now needed to be made in real time, such as the case of Mr Smith as he does his weekly shopping. Or even that the model to ascertain his likelihood to respond to a mobile voucher has been upgraded from good old-fashioned regression models to their shinier AI counterpart. The decision process itself is more intricate in nature. Take our case with Mr Smith. It is not sufficient to know he is likely to respond to an offer – we need to know which offer to put forward. Furthermore, we may wish to ensure the product we are promoting to him is in fact in stock. So rules and logic also feature in our decision-making process.
We’re working with organisations across Europe to help them operationalise analytics and address exactly these types of challenges with intelligent decisioning. In my next blog, I’ll illustrate how intelligent decisioning can help organisations overcome the challenges listed above using real-life examples.The chasm between sculpting an elaborately accurate model and having that same model provide insight for your organisation is where SAS Intelligent Decisioning adds value. Click To Tweet