At its most basic, artificial intelligence (AI) refers to a computer making decisions that typically would have been made by a human. Or put another way, AI is very good at making predictions in which there are a large number of variables and complex interactions at play, Michael Schmidt, chief scientist at DataRobot told CMSWire.

Little wonder, then, that companies are eager to have AI applied to their particular business problems. And why not? It’s new(ish), shiny technology and — as the hype sometimes goes — absolutely necessary to remain competitive. But the truth is, it is the rare problem that can only be solved using AI.

AI Is Overhyped in Certain Situations

“Companies like Google and Facebook, they are doing a great job around neural network architectures,” Padraig Stapleton, VP of Engineering at Argyle Data told CMSWire. Unfortunately, he added, there is too much hype around the technology, especially around the need for AI. “I think about 90 percent of the problems out there that people are looking at AI to address can be solved by more traditional approaches like logical regression and anomaly detection,” he said.

It’s an admittedly counterintuitive concept to swallow, so Stapleton repeated it: “You don’t need the fanciest or the latest technology that’s being talked about or published in papers to solve most of your problems. Most of your problems are solvable by historical approaches.”

This isn't to say that AI is a waste of time for companies. AI is necessary for use cases that call for speech or image recognition, for instance. Anything, Stapleton said, that mimics human capabilities.

It is also necessary for companies looking to create new data, Allen Nance, CMO of Emarsys told CMSWire. “If a company is looking to use the data that they already have, they’re probably better suited for segmentation and analytics applications. If a company is saying ‘I want to use data that I don’t have,’ then that is a sign that an AI or machine learning solution will be necessary.”

When AI Might Not Be the Answer

But just as clearly, there are good reasons for a company to skip the AI and settle for a less complicated, perhaps older technology.

AI Is Complex

A shift to AI can create more problems, or at least a lot more work. “In some ways you can store up problems because if you jump to a neural network architecture, unless you actually understand what’s happening within the actual architecture itself, or you have the people that know how to architect, then more questions arise,” Stapleton said. “You need to figure out the answers to such questions as ‘What data gets fed in?’, ‘How do you encode it before you feed it in?’, ‘How do you train it?’” Unless you have people who understand these concepts you run the risk of actually applying a technology that may not be maintainable in production, he said. 

Learning Opportunities

Which brings us to ...

Data Scientists Are Scarce

Not every enterprise has access to data scientists, who are seen as the gatekeepers of this technology, Schmidt said. And indeed, data science talent is notoriously scarce and expensive. Element AI states fewer than 10,000 people in the world are able to tackle serious artificial intelligence research, as reported by The New York Times.

It May Not Be True AI

Some vendors layer in a superficial aspect of AI — such as an Alexa skill — into their products “solely for the purpose of justifying that their product does indeed provide AI capabilities,” Colin Priest, director of Product Marketing for DataRobot, told CMSWire. And sometimes vendors purport to have an AI-based solution when in fact they do not. “The vast majority of what’s being sold in the market today is neither machine learning nor AI,” Emarsys's Nance said. “These products are some combination of dressed-up segmentation or dressed up analytics.”

It May Not Solve Your Problem Anyway

Consider the process of a credit evaluation, Stapleton said. “In some parts of the world, you need to be able to provide insight as to why you denied someone credit. Some of the newer neural technologies still don’t have that transparency around the underlying features that result in a ‘yes’ or ‘no’ from a credit perspective.”

Older technologies work very well on even complex business problems, he continued, such as mean failure times for industrial automation or calculating fleet efficiencies. “You don’t have to go to AI to solve these problems.”