AI: Augmentation, not Automation

By Elena Salova, Digital Insights Data Scientist, Hitachi Vantara

Elena Salova, Digital Insights Data Scientist, Hitachi Vantara

The buzz around machine learning is hardly new, we see it all around us in recommendation engines—like when Netflix recommends a movie, or Amazon suggests an item based on your recent purchase. But, these are fairly low-risk examples of automation, from a consumption of technology standpoint.

But, what if a situation arises that involves an ethical or emotional dimension—such as law enforcement, where judgment really matters? The cost of failure can be very high.

That’s where augmentation of human capability comes into play. Automation and augmentation can use the same machine learning techniques to solve completely different tasks. Augmentation is improving the efficacy and effectiveness of labor, while pure-play automation seeks to eliminate the presence of human capital altogether. It leaves us with the important question of whether we should look to automate or augment.

An easy starting point common to both options is to look at what repetitive processes and data you have to start with.

At that point, the key juncture on how to move forward is the tolerance to, significance of and frequency of errors. An airplane whose autopilot has decided it’s about to stall, despite what the pilot can rationally observe, can have a disastrous outcome. But, a product recommendation for another expensive guitar, rather than just guitar strings, would be a mere annoyance—a genuine first world problem.

A combination of human skills and AI effectiveness can be the most successful solution for areas where the cost of error is very high. Augmentation means mitigating and managing the risk of errors, low productivity, and low accountability. Keeping human interaction as part of the decision chain would allow for unexpected events and minimize mundanity. Human operators will also be able to make decisions even with incomplete information – while machine-learning models often struggle with such a task.

Self-driving cars have automation levels ranging from zero to five— where five is full automation—and a similar metaphor can be applied to augmentation in general. A base level would be a simple reduction in repetitive tasks—think auto-completion in coding or emails, navigation systems. Even this would provide a significant productivity hike and cost benefits. Then further up the scale would be the ability to make optimal decisions in complex environments. Although that’s far from easy given the range of attributes that often need to be accounted for. For example, a sea captain’s considerations around fuel consumption, speed, weather, wind, currents, and so on—humans can process information only in a few dimensions, which can lead to sub-optimal decisions. Conversely, an AI augmentation can take all these factors into account easily, when given enough resources to output the best combination of route, speed, reliability, and punctuality—not only helping the captain with routine tasks, but also leading to optimized navigation.

Another consideration for augmentation is that the barriers, costs and time to implementation, are lower than a fully automated system. Using the human element to examine exceptions would hugely boost productivity. And, of course, additional positive benefits would be the opportunity the collect more data, especially on human decisions for AI override. This data can power the next generation of algorithms and progress up the scale towards semi or fully autonomous systems.