Why humans still hold the advantage in decision automation
From retail to healthcare to consumer goods; from global businesses to regional operators to Etsy makers and food trucks, businesses of all sizes make a head-spinning number of decisions – using up stores of valuable and finite energy in the process. It’s not surprising then that we try to preserve this precious commodity by simplifying or automating as much as we can.
Refining a process, establishing preferred sources for raw materials, and working with employees, partners and suppliers we know are all ways we “automate” decision making. In our modern era of decision automation, we increasingly use computers to process volumes of data not feasible for humans and to support decision making that we could not otherwise take on. Artificial intelligence holds out the promise of computers making human-like decisions for us.
In the data and analytics business, we exist to help our clients make superior decisions. Our approach embraces the reality that decision automation requires a partnership between technologists, data scientists, and business leaders (or those we like to call Geeks, Nerds, and Suits). But which decisions can we automate, and what can artificial intelligence really deliver?
The Decision Automation Continuum
Typically, we don’t entrust decisions to computers right away. We develop a hypothesis about something data and algorithms can help us with—say, understanding which clearance price points will maximize profit when products reach end of life. We may initially perform some studies using previous sales data to find patterns that interest us, including which groups of customers will buy at a given discount, or which store locations appear to sell through faster than others.
These insights may lead to some simple rules that guide inventory management as a product’s end of life approaches and may cause us to test variable clearance prices at different stores and points in time once replenishment is shut off. We witness how these tests perform and make adjustments.
After more refining, we may begin to trust our systems to make these decisions automatically. Over time, we may allow them to adjust inventory levels and pricing in highly dynamic ways—beyond the capabilities of human decision makers.
This might look like artificial intelligence if we only saw the final stage. But to the experts involved in its development, it doesn’t look that way at all. In fact, all of the intelligence is human. Through this learning process, we refine, consolidate, and concentrate human intelligence through computing machines.
Let’s take a closer look using a kind of maturity model I call the Decision Automation Continuum. We start with data and ad hoc analysis, proceeding to increasing levels of automation as we work out the kinks and build trust in our algorithms and systems. Eventually, the machines are doing the work well enough that we trust them to do it without supervision—decision automation nirvana!
This is the typical path decision automation follows:
First there is unassisted decision making, in which humans do all the thinking. The vast majority of decisions a business makes fall in this category. Data science is still involved at this stage, but the analysis conducted to support the decision is typically based on some hypothesis generated by a business leader (the Suit). This analysis may help us think about how we might automate a decision, such as setting simple decision rules about when to offer product discounts in the pricing example above.
Next, based on what we’ve learned, we can build new models, design metrics, and provide reports or dashboards that inform and assist human decision makers. Automated reporting or interactive dashboards can provide business leaders with deeper insights about the business problem they’re trying to address, and should enable more rapid responses to changes in data trends or customer behaviors.
After enough learning has taken place, we can then build models that will enable the machines to recommend specific decisions or actions. The models and algorithms developed in the first stage may still be used, or they may have been updated by learnings from previous stages.
In this stage, a human still takes the final action, often after adjusting the machine recommendation. Using our pricing example again, this would be where a buyer or pricing analyst reviews the clearance pricing plan for a product before it is put into effect.
In the final stage, we have refined the machine’s decision making to the point where we entrust it to act without intervention—the pricing system automatically deploys the clearance pricing plan without human review or approval.
How We Decide Who Decides
What we call data science actually plays its biggest role in the early stages, investigating hypotheses and developing “sandbox” solutions to complex decisions. This is labor-intensive work. These solutions may proceed slowly or rapidly to full automation, depending on the nature of the problem, and the organization’s ability to effectively integrate and deploy the technology (Geek), data (Nerd), and business processes (Suit) required to create a new decisioning capability. But during later stages, the analyst’s role is primarily tuning, refining, and monitoring the correctness of decision processes that were designed in the first stage.
IT teams—the Geeks—play a bigger role in the last two stages. As we begin to harden and scale up automated decision-making across the business, we need the skills and knowledge of IT professionals to ensure systems run reliably, data are clean and accurate, security is maintained, and the various systems and processes of the enterprise are integrated appropriately.
Finally, the strategists (Suits) are required in every stage to ensure that we are continuing to consider business priorities and ask the right questions along the way. Further, they should be driving the prioritization of which decisions to automate and when, considering how much of an impact the decision would have on the business (especially the potential risk of getting it wrong) and the frequency at which decisions need to be made (it’s not scalable to have humans making hundreds of real-time pricing decisions).
The AI Struggle is Real
When we see organizations struggling with how to apply machine learning or artificial intelligence to a business problem, the issues usually lie in failing to understand this learning process. They want a magic algorithm to take us directly from the first stage to the last, including an automated prioritization of which decisions are most important to automate.
But there are no shortcuts to the fully automated decisions—at least, not in the foreseeable future. While you’ll get there eventually, it takes human planning to ensure that they’re designed correctly in the first place, human prioritization to make the transitions to AI and machine learning, and human oversight to ensure they continue to function correctly over time.
By proceeding along the decision automation continuum, you can set up an effective and sustainable solution that will help us solve problems and assist us in decision making at scales an unaided human cannot reach. While machines can and will magnify our thinking and apply it to data faster and more accurately than we can without assistance, they still can’t think for us. Advantage: human.