Data science and machine learning have tremendous potential business impact. They’re also rapidly becoming commodified table stakes.
So how do you outperform competitors who are embracing the same principles of machine learning and algorithm-driven decision making as you?
The answer isn’t just more or better data science. To get the most value from algorithms and data, you have to situate great data science in the tightest, most nimble outcome-driven OODA loops you can build.
OODA You Think You Are?
The OODA model is certainly applicable to artificial intelligence (AI)-enabled business. In this case, “observe” can be understood as the intake of data. Our algorithms then “orient” by making sense of our broad, chaotic data observations. This algorithmic product then allows our systems to automatically “decide” and “act” (although, for a variety of reasons, we often retain human engagement in these two phases).
Of course, unlike the military originals, our AI OODA loops are driven by business outcomes: customer conversion and retention, sales margins, supply-chain efficiency, return on capital, etc. We continuously evaluate our AI implementations to see if they’re delivering on their value promises. We keep recalibrating our algorithms and data inputs to optimize our business KPIs. And we try our best to respond to the ever-changing demands of the market.
The Rise Of The Non-Deterministic Application
While similar in some ways to the agile, DevOps and continuous delivery disciplines our organizations have recently come to embrace, the AI OODA loop is also substantively different. In both cases, we’re attempting to improve the speed, accuracy and efficiency with which we get feedback from the real world and use that feedback to improve our organizations’ digital behaviors.
But with conventional applications, we know what the code does. So when we have a specific new functional requirement to fulfill, we simplify modify that code as appropriate. There are certainly challenges associated with writing that code properly and making sure we don’t accidentally break anything else, but the behavior of that code is ultimately deterministic.
AI doesn’t work like that. Its inner workings are non-deterministic — constantly and autonomously reconfiguring themselves in response to new inputs. So instead of changing and testing procedural code, we have to keep monitoring outcomes and then forensically relate those outcomes back to algorithms, data inputs and other application parameters.
AI Ops: New Rules For New Loops
We’re all discovering how to keep getting the most out of our non-deterministic AI applications over time. Emerging best practices include:
• Define and measure your loop processes. One great lesson from DevOps is that a well-defined process with well-defined metrics is much easier to manage, troubleshoot and improve than unstructured ad hoc team behaviors. So start mapping out your AI processes and selecting some initial metrics for what will eventually become your “AI Ops.”
• Close working relationships between data science staff and business domain knowledge leaders. AI is a team sport. Advancements in algorithmic technology are coming so quickly and are of such great potential value that we can’t expect our data science experts to also be domain masters. They therefore need to be paired with people who really understand our business, our markets and our customers.
• Build outcome guardrails and bias tests. Good AI governance isn’t just about incrementally improving business key performance indicators over time. It’s also about avoiding potential autonomic disasters and the unintentional systemization of bad ethics. Make sure your process continuously addresses these types of issues as well.
The continuous optimization of non-deterministic AI applications is new to all of us. But it’s something we all will need to do extremely well — because there’s a very high-stakes battle going on across virtually all of our markets. And when you’re in a tough fight, your OODA loop can make all the difference.