Adopt Decision Modeling for DecisionsFirst Analytic Success

Bill Schmarzo wrote a great post on 6 steps to AI Success recently. He identified 6 steps in what he called the AI Solution Decomposition process that work for analytic and AI projects. This is almost exactly the same process we use on projects and you should read the whole post as he makes some great points. Here’s the list with my additions:

  1. Identify and Understand Your Targeted Business Initiative
    Nothing to add, Bill nails it.
  2. Identify Your Stakeholders and Constituents
    Likewise – design thinking and stakeholders are key.
  3. Identify Key Decisions
    We agree with Bill that this is a critical step. However, we would add a couple of things:
    • We find it is not enough to simply list and prioritize decisions – you have to understand them. First and foremost this means capturing the question you have to answer to make the decision and the possible or allowed answers for this question.
    • Then you have to model these decisions to make sure you really understand them. We use the Decision Model and Notation (DMN) standard to model decisions so we can decompose the decisions, see how they are linked (making a decision often involves knowing the result of a another precursor decision) and understand them enough to both find analytic opportunities AND implement them. Framing the requirements for analytics with decision modeling is really effective – check out this white paper for more on that.

      In the example you can see a decision to be improved, the various decisions into which it has been decomposed, the data and regulations/policies involved in making it and where, exactly, an analytic might be used. Decision models provide critical context for your analytics.
    • Finally, it’s important that your decisions really are decisions. Knowing something (lifetime value score, retention risk) is not the same as making a decision as to what to do about it (next best offer, channel to approach the customer in). Make sure your decisions involve some kind of business action not just knowing something.
  4. Identify Predictive Analytics
    We identify three distinct ways to identify predictive analytics – Data First or DecisionsFirst with DecisionsFirst being broken down into Incremental and Disruptive:
    • Data First Analytics are the ones the analytic team finds by looking at the data without constraints or pre-conceptions. These will be the predictive, most radical insights you are going to find. Adopting them will be hard – often impossible for regulatory or organizational reasons – but they may offer dramatic improvements in results. While these are driven by the data, building a decision model of the decision making approach that would have to be adopted to use the analytic can be revealing and is often a critical building block for a successful adoption.
    • DecisionsFirst Incremental Analytics are predictive analytics that fit right into your current decision model – they just make it better. For instance, using analytics to predict that a customer has an undisclosed medical condition (something claims adjusters try to do) or analyzing data to determine which medical conditions really are straightforward (which may not be the same as the “official” list). These are easy to adopt – the decision model makes it clear where they go – and build trust with your business partners because they show that data and analysis can be used to improve the business. We often find these by looking at a decision model and asking “if only…” – if only we knew which customers had preexisting conditions, if only we knew which forms had been altered, if only….
    • DecisionsFirst Disruptive Analytics are analytics that fit in the decision model but require significant change to that model to be adopted. The analytic team often finds these by combining several distinct kinds of data across the breadth of the decision and finding interesting patterns between data elements normally not considered together. The analytics team can show the analytic, describe the potential benefit to the business and show the change required in terms of the decision model. These are harder to adopt – they require change to the decision-making approach after all – but are generally much less disruptive than Data First Analytics because the business has been involved in driving the decision modeling and so the change is described in their terms.
  5. Brainstorm Data That Might Be Better Predictors of Performance
    Often combined in our approach with the identification of predictive analytics. I’d only point out that everyone SAYS they want to predict customer attrition for next month but what they REALLY want to do is reduce it….
  6. Implement Technology
    Last, where it belongs. I’d only add that we find implementing a Business Rules Management System can really help get analytics and AI into production by automating more of the surrounding decision so it can be made as analytically as possible. A BRMS also allows decisions to be simulated with different analytics, something that can be particularly useful if business teams are cynical or worried about analytics or when adopting more disruptive analytics.

In summary, a business-led, DecisionsFirst approach to analytics and AI, especially one that uses decision modeling to clarify what decisions are being improved and how, is a powerful approach for success in AI and Analytics.