Are You Using ML/AI for Automation? or for Innovation?

Determine whether an ML/AI project is for automation or innovation so you can prioritize it accordingly.

There are many ways to categorize your ML/AI projects: “easy-medium-hard” for expected effort, “red-yellow-green” for chances of success, or even “base scikit-learn classifier versus neural networks versus custom adaptation of an algorithm from a research paper” for the technology used.

One split I never hear about, but I find very important, is “automation versus innovation” in terms of the project’s purpose. Both are meaningful, but they serve different needs and therefore have different expectations in their outcomes.

Using ML/AI for automation

With automation, you build a tool to peform a task that you currently do by hand. Compared work performed by a person, machines will work 24x7 in a consistent, precise manner. (Granted, it’s possible for a machine to perform in a consistently bad fashion, but that’s still consistent.)

Agriculture automates through machinery. Manufacturing uses robots. And as of the last three decades, companies have increasingly turned to custom software to automate business processes. Ordering goods? Buying plane tickets? Sending form letters? Let the code do the heavy lifting.

Software is, deep down, a series of actions tied together through a series of conditionals: if X happens, then do Y. Otherwise, do Z.” Software-based automation requires that you have a clear, unambiguous, rule-driven business process that will reliably translate into if/then statements in code.

That, in turn, requires that the underlying business process is all of dull, repetitive, and predictable:

  • Dull, because it’s drudgework that humans won’t want to do or will get distracted as they do it.
  • Repetitive, because software scales best when it has to do the same task ad infinitum.
  • Predictable, because the software can only do what it’s programmed to do. It can’t adapt to a new situation. Developers need to bake in enough of those if/then staments to handle every possible scenario. (When software encounters a situation beyond its programming, all sorts of weirdness can happen.)

All of this means that software is great for automation when the conditions are simple or straightforward: “If the price falls below $100, buy” or “If the person booking has gold-tier status, and they’re booking at Hotel XYZ, and they’ve stayed at this hotel more than six times, then apply the special loyalty discount.” Code can become hard to read, and even harder to debug, when it’s a mass of nested if/then statements from more complicated business rules. It also stumbles when the the hard business “rules” are more like “guidelines.”

That’s where automation through ML/AI steps in. Unlike software’s if/then statements, ML/AI models are built by an algorithm that generalizes across what it sees in training data: “under this variety of circumstances, usually, X is the answer.”

That allows the models to act under conditions that aren’t exactly like what they saw during training. You don’t have to tell models about every possible situation that they may encounter, which makes them slightly more adaptable than their software-for-business-rules cousins. And since the model’s decisions are based on patterns it has found in the training data, it can handle tens or even hundreds of conditionals (expressed as features in the training data) that would be much more difficult to implement in code.

Using ML/AI for innovation

While automation is about doing the same work with less human involvement, innovation means using ML/AI to find new business opportunities and new ways approach how you operate.

Consider the world of trading. Until the 1990s, many practitioners focused on qualitative or visual tools to devise their strategies. That changed when physicists arrived on Wall Street and brought new levels of statistical and analytical rigor to the field. They built massive mathematical models to surface unexpected correlations and mispriced assets in a very large sea of data.

All of that analysis drove new trading strategies, especially in arbitrage opportunities. This quantitative analysis reinvented trading and it is still the norm today.

The 2002 Oakland A’s applied that same idea to baseball. Their statistical analyses uncovered over- or under-valued player attributes – ergo, under- or over-valued players – and they developed their recruiting strategy accordingly. Acquiring top talent on the cheap allowed them to compete with teams that had much larger budgets.

This is probably why, when Michaal Lewis wrote up this story in Moneyball, he subtitled the book The Art of Winning an Unfair Game. By applying statistical analysis in new ways, the Oakland A’s completely upended traditional baseball recruiting techniques.

Why it matters

Automation and innovation serve different purposes:

Automation ultimately serves to shift the underlying economics of a business process, by moving work from people to machines. This permits you to do more work without growing your team, or do the same amount of work with fewer people.

While automation is hardly guaranteed to work, your first few experiments in building the model will give you an idea of how things might turn out. You can then decide how much more time and effort to invest in improving performance, and how much money to invest in staffing up additional human supervision for the model.

By comparison, the goal of innovation is to find a new way of doing business. You have a much wider solution space to explore here, because you’re creating new pathways and procedures from whole cloth. Innovation is a riskier prospect than automation, but it can have a much greater payoff if it works out.

Above and beyond acquiring and analyzing data, innovation requires that someone imagine what else could be possible. Before physicists came to Wall Street, someone first had to raise the questions: “What if we were to use statistics to look for weird price relationships? And who would have that depth of mathematical knowledge?”

Why not both?

Automation through ML/AI is important and is useful. That’s probably why it dominates the landscape of ML/AI efforts: the lion’s share of today’s ML/AI work involves document and image classification, content moderation, and recommendation systems.

At least, that’s most of what we see in the public press. Perhaps there’s more innovation happening behind the scenes? The best way to use ML/AI for innovation in your company is to leave room in your schedule and your budget to explore. You never know what you’ll find until you try.