Call us at 650-400-3029 (PST)

Operational ML/AI

We’ve talked about different kinds of ML/AI in other posts and covered both Interface AI and Research AI. Today we’re talking about operational AI – the most powerful approach to applying ML and AI. If you want to MAKE money from applying ML and AI rather than simply SPEND money on ML/AI, this is where you should focus! 

Operational AI is about applying ML/AI to make probabilistic predictions or classifications of a transaction based on historical data you have available. Depending on the situation and the accuracy/false positive rate/precision of the prediction, you will act differently – probabilistically. 

Operational AI generally makes a high-volume transaction or interaction a little more precise, a little more automated. It shaves some fraud loss or tightens the risk assessment or focuses the marketing offer just a little better. And because you do this a lot, the pennies add up fast. 

If you are already focusing on operational decision making – building decision models or implementing business rules or both – then you can immediately identify ML/AI opportunities. Building decision models is a best practice and creates a visual blueprint of your decision-making, structuring the decision-making and decomposing it into a network of shared elements. With a decision model in hand, you can ask business users to do a couple of things 

  • Look at the sub decisions in the model and identify those based on rules or thumb, experience or best practices. It may well be possible to validate these with ML/AI and even replace them with an ML/AI model that can analyze more cases, more systematically than the experts did originally. 
  • Find sub decisions that involve human analysis of historical data. These can almost always be replaced with ML/AI models. Often the analysis is simplistic to make it possible to do in the time available and a more automated ML/AI approach can be more sophisticated. 
  • Identify sub-decisions where the available data doesn’t allow for “good” decisioning – too many maybes or manual referrals, limited segmentation or clustering etc. In these areas, ML/AI might be able to create probabilistic attributes – the likelihood of something being true – that can be mixed with “real” data to make a more complete or more specific decision. 

In each case the design of the decision – the model – is guiding the ML/AI development and provides an operationalization context for the model once it is developed (assuming it can be – not all will work after all). The visual blueprint of a decision model also allows the business users to feel engaged in the ML/AI work and to own the result. 

To wrap up back in our claims example: 

  • A model that predicts how likely an insured is to have failed to disclose something when they applied is combined with rules-based red flags to improve fraud detection and send the right claims for fraud review. 
  • Analysis of past hospital stays is used to flag those stays that are excessively long or expensive for a waste review. 
  • A prediction of the kinds of documents most likely to be needed before a specific claim can be processed is used as part of determining if all the documentation needed for processing has been received. 
  • Predict how likely additional claims are to be received so that situations with a potential for many repeat claims get extra scrutiny. 

Operational AI is easy to get right if you go at it DecisionsFirst and combined with Interface AI it can really make your systems hum.