We’re still new to the AI era of software, and there is a lot of confusion about what “counts” as an AI company. AI companies are exciting because their products become more powerful with more data, making them more defensible over time-- at least, that’s how these businesses should work. In reality, the first wave of AI startups have been all over the map. Why is it that “fake AI” companies are taking off in terms of customer traction, but “true AI” businesses built around cutting-edge models stall out? Why have other “fake AI” businesses crashed and burned after stalling out at scale, and why does Google’s defensive moat around Search deepen further every minute?
We confuse ourselves because we talk about AI and Automation interchangeably. Customers are looking to invest in value creation, even though they may think they’re looking to buy AI or Automation. An AI model on its own does not create value. Automation on its own creates some value, but not in the runaway sense where the outcomes get better over time. We can tease apart the distinction by first mapping out all the steps in the process of creating value.
Completing this process is incredibly expensive to do entirely manually. When that was the case, only a limited amount of data was captured and used. Sensors, databases, and software reduced the cost of the earlier stages of the value creation process, saving manpower to identify optimization opportunities, apply judgement on whether to apply the optimization, and implement the relevant action.
These last three steps to value creation is where data science and AI/ML models, rules engines, and automation come into play. Data science and AI/ML models replace analysts and run optimization models to the best outcomes for each set of inputs-- at a dimensionality and speed far higher than what humans can do analyzing the data on their own. Rules engines replace the decisionmaker where the same decisions are made over and over -- “if A, then go down path 1, if B, then go down path 2.” Automated controls take the direction from the decisionmaker or rules engine and complete the work, replacing the human who would otherwise do that work. An automation system can be a physical machine or a computer script that completes the task designed to drive toward the desired outcome. These three elements are entirely separate things but have been commercialized both as standalone and packaged solutions and indiscriminately called “AI.” We’ll explore how these pieces impact the value creation process on their own in order to find out how they could best fit together.
The data science or AI/ML model is the most potentially valuable component of the value creation process, but it is also the most dependent on the rest of the steps. The model takes in data and returns a prediction or optimization, which provides insight to the decisionmaker. Insights, however, are not inherently valuable-- they must lead to some kind of action in order to create value. If the insight is discarded or disregarded, no value is created for the customer. This is why AI companies that sell AI models or “intelligence” to their customers may struggle with ROI -- there’s no guarantee that your customer will actually do anything with the insight you surface. If the insight from the model is used, however, it can lead to an outcome that gets measured by the sensors, which feed that data back into the model to make it more accurate over time. Each turn of that feedback loop makes that model more accurate. An AI model can replace analyst labor. It could also increase topline revenue opportunity by powering a better product or enabling greater scale due to the optimizations it drives.
Rules engines act like guardrails -- they encode process and frequently repeated decisions. They take inputs and route actions based on those inputs. Rules engines on their own do not drive to more optimal outcomes with more data, but they can be used to improve quality by preventing avoidable mistakes. Rules engines may also save a small amount of labor.
Automation occurs when a machine replaces a human in carrying out a task. When the automated controls systems receive a signal, the robotic arm or software script will complete the specific action associated with that signal. Automated systems on their own do not automatically improve performance or quality over time. The business value of automation is typically seen in lowering direct costs by removing labor or lowering fixed costs by accelerating throughput.
The most powerful AI startups accumulate data very quickly and improve the power of the model with every turn of the feedback loop. The goal for these AI businesses is to get as many turns of that feedback loop as possible, in order to train the AI model at the center and creating increasing value over time. Combining the model with a rules engine and automation is the fastest way to cycle through that loop the fastest of all. Just operating the AI model in a vacuum, however, risks creating no value at all. Likewise, just applying a rules engine and automation loses out on the opportunity of the ever-optimizing feedback loop at the center.