There is a mountain of hype around big data, artificial intelligence (AI), and machine learning. It’s a bit like kissing in the schoolyard – everyone is talking about it, but few are really doing it, and nobody is doing it well (shoutout to my friend Steve Totman at Cloudera for that line). There is certainly broad consensus that organizations need to be monetizing their data. But with all the noise around these new technologies, I think many business leaders are left scratching their heads about what it all means.
Given the huge diversity of applications and opinions on this topic, it may be folly, but I’d like to attempt to provide a practical, useful definition of artificial intelligence. While my definition probably won’t win any accolades for theoretical accuracy, I believe that it will provide a useful framework for talking about the specific actions that an organization needs to take in order to make the most of their data.
The theoretical definition
If you asked a computer scientist (or Will Smith), AI is what you get when you create a computer that is capable of thinking for itself. It’s Hal from 2001: A Space Odyssey or Lt. Commander Data from Star Trek: The Next Generation (two of the greatest masterpieces of all time). These computers are self-aware: thinking, independent machines that are (unfortunately) very likely to take over the world.
While that definition may be strictly accurate from the ivory tower, it’s not particularly practical. No scientist created such a thing, and no business is really considering utilizing such an entity in their business model.
Laying aside that definition, then, let’s look to something much more practical that can actually move the conversation forward in business.
AI is not machine learning
There are two main concepts, according to my definitions, that are important. AI is one, and I shall define it shortly. Machine learning is the second. There’s just as much confusion about the definition of machine learning as there is about AI, and I think it’s important to point out that they’re not the same.
Machine learning is known by other names. Harvard Business Review called it data science, and dubbed it the sexiest job of the 21st century – which is a pretty bold claim, given that there are a lot of years left until the 22nd century. Years ago, it was called “statistics” or “predictive modeling.”
Whatever you call it, machine learning is method of using historical data to make predictions about the future. The machine learns from those historical examples to build a model that can then be used to make predictions about new data.
For example, credit card companies need to detect fraudulent transactions in real-time so that they can block them. Losing money to fraud is a big problem to card providers, and detecting fraud is an ideal machine learning problem. Credit card providers have a mountain of historical transactions, some of which were flagged as being fraudulent. Using machine learning, the historical transactions can be used to train a model. That model is basically a machine that looks at a transaction and judges how likely it is to be fraud.
Another common example in the healthcare space is predicting patient outcomes. Suppose a patient goes to the ER and ends up getting an infection while they’re in the hospital. That’s a bad outcome for the patient (obviously), but also for the hospital and the insurance companies and so on. It’s in everyone’s interest to try to prevent these kinds of incidents.
Healthcare providers frequently use past patient data (including information on patients that both did and did not have a bad outcome) in order to build models that can predict whether or not a particular patient is likely to have a bad outcome in the future.
Machine learning models are very narrowly defined. They predict an event or a number. Is the patient going to get sicker? How much pipeline will my sales team generate next quarter? Will this potential customer respond to my marketing message? The models are designed to answer a very specific question by making a very specific prediction, and in turn become important inputs into AI solutions.
Artificial intelligence combines data, business logic, and predictions
Having a machine learning model is like having a superpower or a crystal ball. I can feed it data and it will make predictions about the future. These models can identify potentially bad loans before they default. They can forecast revenue out into the future. They can highlight places where crimes are likely to occur. The AI system is how you put them to practical use.
Let’s go back to the credit card fraud example. Suppose I could tell you by means of a machine learning model whether or not a transaction was likely to be fraudulent. What would you do? Even thinking about it for a minute makes it obvious that there’s a lot more work to do before you can start getting value out of that model.
Here are some questions that you need to consider in this example:
- What data is available to me at the time of the transaction?
- How much time do I have in order to process the data and reject the transaction?
- What regulations restrict my ability to block potentially fraudulent transactions?>
- Nobody likes having legitimate transactions blocked. What customer experience concerns do I need to address?
- What false positive rates and false negative rates am I comfortable with?
- …and so on
There are many more questions that a credit card provider would need to consider before implementing a system to block potentially fraudulent transactions.
That system, though, is what I call AI. It’s the combination of all the business logic, all the data, and all the predictions that I need in order to automate a decision or a process.
- Business Logic: Business logic is probably the most important aspect of implementing an AI system. It encompasses the user experience, the legal compliance issues, the various thresholds and flags that I may need, and so on. It’s basically the glue that holds together the whole process
- Data: AI systems reach out for data. They might need to aggregate customer data, summarize transactions, collect a measurement from a sensor, and so on. Regardless of where it comes from, data drives the AI system; without it, the system comes screeching to a halt.
- Predictions: Not every AI system uses data, but all of the good ones do. Anyone that has ever called their cable provider has dealt with the endless automated phone system. They’re trying to automate a process, but they’re not being smart about it. It’s dumb AI. Smart AI might make predictions about why I was calling and attempt to route me to the right place, for instance. Predictions are the technology that makes AI truly smart.
Get started today
There are lots of resources for organizations that are trying to get started with AI. DataRobot offers education programs that can help anyone quickly get up to speed quickly. There are many organizations out there that can help accelerate the process.
Remember, though, that it’s not about finding the perfect use case with the perfect data. It’s about taking on as many small projects as you can handle in order to generate value quickly. Getting a few wins under the belt will build up substantial momentum and make it possible for you to iterate and expand your monetization of your data.