There is a legitimate reason for the hype around machine learning, and it goes beyond winning an ancient board game (see AlphaGo) or recognizing faces in photos. Machine learning is delivering unprecedented business impact across a wide range of industries.
In general, there are three ways we think about making business impact with machine learning:
1) Reduce Costs
Spend less money to perform existing tasks. Since Profits = Revenue — Cost, reducing costs increases net profits. A widget-maker can sell the same number of widgets and keep more money. The reduction in costs creates a profitable situation even as revenue stays flat.
2) Increase Efficiency
Enable your current processes to happen faster. Think of this as giving employees superpowers to do their jobs, enabling them to produce more units of work in a given timeframe. Increasing efficiency could impact both revenue and costs: a business unit can produce (and thus sell) more widgets (increasing revenue) and could also produce and sell the same amount of widgets with fewer inputs (reducing costs). This shows both cost and revenue lines heading the right direction:
3. Achieve Breakthroughs
There’s a lot of opportunity to use machine learning to accomplish new types of tasks that weren’t possible — or even imagined — before, thus delivering new types of value that weren’t previously available. Breakthroughs are a potential revenue boost. A business unit can add new kinds of widgets or features to its product line, boosting sales. If costs stay flat as revenues increase, profits go up.
In practice, of course, a given ML application may fit more than one of the above impact categories. An ML algorithm that can analyze documents, for instance, might be reasonably said to increase employee efficiency by helping them process documents faster. But it might also be viewed as a breakthrough since it allows the company to offer new document analysis products it never could before.
Or consider a machine learning algorithm that detects security threats (like guns or knives) in an image generated by a baggage scanner. It may reduce costs by reducing the need for a person to watch a screen constantly. It may increase efficiency by finding more threats. And it may represent a breakthrough by finding threats that humans may have never been able to detect.
Furthermore, “breakthroughs” need not necessarily be associated with new products. What if the breakthrough enables a new kind of insight for the C-suite? Such reporting may lead to strategic choices that ultimately lead to increased revenue or lower costs, even if the immediate impact is not as clear.
Business impact thinking is not meant to rigidly constrain the way you think about machine learning projects. Think of it instead as a useful guide. You should certainly be aiming in the direction of at least one business impact — but you should be flexible enough to know when you need to change the way you communicate your goals and frame your success.
It’s not unheard of for a machine learning project to begin with a goal of reducing costs by automating away entire jobs. But this isn’t always easy. Even if 80% of an employee’s work can be eliminated, the remaining 20% may be stubbornly un-automatable. This doesn’t mean the project was a failure. If the humans can focus on their 20% of the task while being free of the other 80%, they can take on new business. This could represent a massive business efficiency gain — a huge win when the business impact is viewed properly.
Robbie Allen is a Senior Advisor to Infinia ML, a team of data scientists, engineers, and business experts putting machine learning to work.