Several years ago, I lived along the coast and had this fantastic deck. It had this crazy breeze that always seemed to be blowing across it. I loved that deck, except for one thing. It could get very cold.
So, I decided that I needed a space heater to complete my vision of deck-bliss. But I just couldn’t bring myself to buy a space heater like those restaurants used on their patios. It just seemed so, well, unoriginal.
Which is why I was so excited when I found it. I had walked into a hardware store looking for something innocuous and there it was: this interesting, modern and completely unique outdoor space heater. I had never seen anything like it. I bought it on the spot and happily took it home with a Cheshire Cat grin on my face.
Alas, my bliss was short-lived.
It seemed that I wasn’t the only one taken by this new space heater design. Within weeks, they were popping up everywhere — on my neighbor’s decks, in public spaces and, of course, on restaurant patios. To this day, whenever I see one I sigh inwardly and pout a little.
In many ways, this is the same story of technology fads in general — and the rise of machine learning specifically. Like my momentarily unique space heater, machine learning was briefly this unique thing of wonder. But it almost instantly went from unique to commonplace.
Machine learning is now on the tip of every IT executive tongue and spewing forth from every vendor trying to sell those executives, well, anything. But like the plethora of technologies before it, machine learning is no panacea. And like most other over-hyped fads, there is both truth and myth, promise and peril lying just beneath the surface. It’s critical that you can tell one from the other.
The promise of machine learning
A recent survey of CIOs conducted by ServiceNow, in conjunction with Oxford Economics, found that 89% of organizations have deployed machine learning or are planning to deploy it. On the surface, this is a surprising number. In my conversations with IT executives, I would describe their forays into the cognitive sciences as nascent, at best.
But I believe it points to a broader underlying trend. It’s now getting impossible to not use machine learning as vendors increasingly embed it into virtually every category of technology.
Both its purported use and the fact that vendors are embedding it with abandon make a lot of sense. Machine learning, which is just one category of artificial intelligence, is well suited to solve some of the most persistent problems that have plagued IT for decades and which the rapid evolution of the enterprise technology stack are only exasperating.
“A lot of what got automated in the ERP-era was the routine cognitive work,” explained Chris Bedi, ServiceNow’s CIO. “What was left was the non-routine, cognitive work. Machine learning is addressing this, and this will be the next big step up.”
Most IT teams spend the vast amount of their time, energy, and resources handling these kinds of non-routine, cognitive activities. But the reality — and reason that machine learning can help — is that these non-routine activities do, in fact, follow definable patterns. Finding those patterns is where machine learning excels and why it will, eventually, change the way organizations operate.
The promise of machine learning, in this context, is that by identifying these patterns and the outcomes they produce, systems can take over much of the drudgery that IT organizations now spend most of their resources handling and enable IT professionals to make better decisions more quickly.
And, the thinking goes, with the drudgery removed and with insights at-hand, gleaned from patterns that would have otherwise gone unnoticed, the IT organization will deliver exponential value to the enterprise.
The perils of machine learning
While there is ample justification for the rosy optimism surrounding machine learning, things are never quite that simple.
I recently gave a speech at the AppDynamics Summit in New York entitled, Machine Learning: The Enterprise Pandora’s Box. In it, I had two primary objectives: explain what the heck machine learning meant from an IT Operations perspective and to squelch some of the irrational fear and exuberance surrounding it.
The reality is that we are still in the very early stages of the AI-powered transformation of IT (and the enterprise). Yes, it’s coming. But it’s neither here yet nor will it overtake organizations like a magical wave of utopian pixie dust.
Transforming the organization into a cognitive enterprise will be an arduous task and an evolutionary process. Jobs will not disappear overnight, and many organizations will outright fail to leverage the power of this technology — and will suffer the business consequences as a result.
This lack of inevitability is because there are two significant problems when it comes to leveraging machine learning in the enterprise: data and bias.
Machine learning only works with data. Lots and lots of data.
It’s called machine learning because the machine must be ‘taught’ by giving it data from which it can distill patterns, and, in most cases, the teaching data must be ‘clean’ — meaning that it must be accurate and represent the desired outcomes (this is called supervised learning).
This reality means that for machine learning to work, an organization must begin with lots and lots of good, clean data. For most enterprise organizations this is going to be a problem.
But even if you have the data, you then bump into bias, the second major machine learning challenge.
While there is a lot of talk about machine learning bias in the press right now, most commentators are focused on social issues like embedding gender or racial stereotypes into machine learning algorithms. That’s a genuine concern, but it’s not the only one.
There is also something called operational bias. A simple example:
Every time a particular type of incident occurs, you send it to one specific team. The reason is simple: that’s how you’ve always done it. That team, while perhaps not the best team to handle this incident, has accepted its fate and gets the job done. Fed data that shows that incidents of this type are successfully routed to and resolved by that team, a machine learning system will correctly learn that this is the current course of action and will begin to automatically take that action. But it may not be the right course of action.
Therein lies the risk of operational bias when it comes to machine learning. The technology can discern patterns and understand what has happened in the past, but it has no way of knowing if that is necessarily the way things should be done in the future. Bad data will lead to bad outcomes — just faster and automatic!
Finding your machine learning balance
Machine learning is a part of your future. You must begin the process of adapting your culture and get comfortable relying on machine intelligence as a significant source of your organization’s decision-making capability.
The technology holds tremendous promise for the enterprise. But it won’t just happen. You must begin critically evaluating the health of your data now. That’s also where embedding machine learning into purpose-built platforms can help. Having a captive set of data enables these platforms to build reliable and clean data sets more easily and more quickly and will help organizations reduce operational bias.
Nevertheless, enterprise organizations must seek to find their machine learning balance. There is no place for either blind fear nor optimism. Machine learning is, after all, just another technology — there’s no magic here. As such, enterprise leaders need to apply it like they should all other technologies: deliberately and only when and where it delivers specific and measurable business value to the organization.
[Disclosure: As of the time of writing, ServiceNow and AppDynamics are Intellyx customers.]