Solutions

Decision decomposition

Break a decision into its component parts

Organizations need to establish a framework in order to make effective decisions. DI treats the establishment and refinement of that framework as an engineering problem, structuring elemental decision components to optimize decision making.

Drucker described decision making as the “specific executive task.” He also said, “Only executives make decisions, to make decisions that have significant impact on the entire organization, its performance, and results defines the executive. Effective executives therefore make effective decisions."^xv

Business leaders can model their decisions around Drucker’s five essentials for effective decision making:

Recognize generic problems. Use the decision to establish a rule/principle. (Good advice where applicable, but the context around decisions changes rapidly)
Define boundary conditions. (Still good advice)
Define success criteria beforehand. (Still needed)
Build the action into the decision. (Synthesis of decision elements benefit from further decomposition)
Use feedback to test validity of the decision against reality. (Still needed, but much more involved)

Decompose the decision-making process into the individual elements, then synthesize those elements into a coherent framework. ML and AI predictions alone don’t constitute a decision. Traditionally, executives have combined prediction and decision-making, a practice poorly suited to complexity. Separate the decisions and predictions and start modeling the DI framework.^xvi

Pratt explains:

“You can draw a bright line between predictions and decisions, but it's not widely understood that they're quite different. A decision is something that leads to an action, which leads to some outcome through a chain of events; whereas a prediction — like with weather or stocks — doesn't have to do with actions that lead to outcomes. The way to think about predictions is that they are one component of the action-to-outcome pathway. For instance, I might choose to invest in a stock because I predict that this stock is going to appreciate. The decision I make to trade that stock is informed by that prediction, but it's also informed by the price of the stock, how much money I have in my bank account, my investment goals, my personal tolerance for risk, my need for liquidity, and other considerations that go beyond just the prediction of the stock forecast. This ‘big picture’ is often overlooked by simple predictions.”

Relationships that are relevant to the situation train prediction models. A framework combines predictions from the models with judgement, which can then be used to act. This action leads to an outcome. This outcome produces feedback that can be used to improve subsequent predictions and judgement.

Figure 5 Decision-making process^xvii

Leaders should develop a decision model to make impactful decisions. The resulting framework should incorporate principles of DI and utilize data and AI in the broader context of the decision as a whole.^xviii

Pratt’s “Causal Decision Diagram” (CDD) provides a model for decision frameworks. The CDD breaks the decision down into its component parts:^xix

Figure 6 Decision Framework Elements

Source: Quantellia LLC^xx

Levers: The set of choices that affect an outcome or goal. Only the most simplistic decisions have a one-to-one impact from levers to outcomes. Rather, levers are more often choices that start a chain of events, leading to one of many potential outcomes.
Externals: Factors that affect the outcome but are not in the decision-maker's control. These may be simple facts (like the price of an umbrella) or may also be predictions. For example, one might choose to bring an umbrella based on the prediction of rain—but no decision affects the likelihood of rain.
Outcomes: The results of decisions. They are potential future states that result from the interaction of decisions and externals.
Intermediates: Additional event chain values that link decisions to outcomes. These intermediates can change the intended course of a decision, leading to feedback loops and unintended consequences.
Goals: Tests against outcomes that represent success. For instance "net revenues are 10% higher 12 months from now".

Pratt emphasizes the importance of decision intelligence in complex environments:

“The hard problems of our time are these multilink decisions whose effects invisible, exponential, and involve feedback effects – problems like Covid-19 and climate change. These problems also require multiple disciplines to be integrated. So for Covid-19 for instance, our systems integrate system dynamics with computational fluid dynamics, agent-based modeling, building designs, and physics. This is the role of DI: it’s like a ‘glue layer’ that integrates these pre-existing components.“

Urbint uses a three-step approach to build DI models with predictive input:

Build a model of the world that includes data about the infrastructure and relevant negative impact.
Train ML models to predict the probability of the negative outcome occurring.
Build DI frameworks that can adjust decisions based on various levels of risk.^xxi

It then defines three use case patterns to take the probabilistic output from the model and use it as an input to improve decision making:

Scoping: In this situation one considers which asset(s) in the network to treat. The model uses predictions to determine which adjustments could maximize the expected benefit. The choice may be constrained due to a fixed budget from the client or it may be flexible — in which case the selection will be the set of assets with highest marginal benefit. Benefits relevant to utility assets may be intangible: safety or financial, for example.
Sequencing: Within infrastructure environments that regulate mandatory treatment of all, the predictions help determine what order of asset maintenance maximizes benefit over the long-term.
Selection: When multiple maintenance options exist, the predictions assist by identifying the optimal mix of treatments over the entirety of the network for highest expected value. For example, utilities can maintain a gas line in multiple ways. The choice is complicated because different treatments have different costs and benefits, which vary between gas mains across large networks. The machine learning model predicts which selection and sequences of treatments maximizes benefits.^xxii