How to learn from inevitable AI failures
Most artificial intelligence projects fail—that’s the bad news. The good news is learning from AI failure is exactly what your company should be doing right now.
We’ve seen this movie before. Artificial intelligence (AI), like big data and [insert the name of your favorite technology trend here] before them, is destined to change the world. NOW. Except, of course, that it’s not. Not now, and not anytime soon. Not at scale, anyway. You can see this in the conflicting data pulled from user surveys, which essentially scream: “Everyone thinks this is important but few have figured out how to flip the ‘on’ switch.”
Given the rampant confusion in AI, what should an enterprise do today to take advantage of AI tomorrow?
Making stuff up with AI
Everyone wants to be like Google these days, with CEOs touting their companies’ various AI/ML projects on analyst calls and press releases. Meanwhile, as Ben Lorica has highlighted, patent filings for AI-related innovations are off the charts (especially relative to publications on the topic). For those companies that have been doing AI for a while, 43% expect to spend more than 20% of their IT budget on AI projects.
This is big!
Or not. These kinds of figures sound great until you ask companies how they’re faring with those efforts. The tl;dr? Not so well.
Indeed, according to IDC survey data, upwards of 25% of companies report a 50% failure rate for their AI projects. This isn’t too surprising since just one-quarter of enterprises have implemented a broad AI strategy, according to the same data.
Even less surprising, much of the interest in AI isn’t being driven by folks on the ground within the enterprise but, as TechRepublic Premium survey data suggests, it’s being pushed by the C-suite 33% of the time. This is a recipe for failure, says analyst Lawrence Hecht: “These projects are destined to fail if there is no underlying technology need. Yes, I understand that c-levels are needed to lead everyone towards change, but sometimes it seems it’s just for change’s sake.” The other way to look at that same data is analyst Sam Charrington’s view: “[It] could also be ‘our lunch is gonna get eaten if this is real and we miss it, so here’s some $$ go figure it out.'”
Regardless of whether this glass is half-full or half-empty, the reality of AI within the enterprise is that it remains more aspiration than reality. Gartner, for example, has estimated that up to 85% of all AI projects will “not deliver,” a number confirmed by more recent research.
Making AI work
This isn’t to suggest enterprises should sit on the sidelines until AI/ML comes of age. The harsh reality is that it won’t without enterprises investing in it. Why? Because one of the biggest hurdles to AI success is people: There’s a shortage of skilled data science personnel.
Yes, and no.
Partly this is a problem of skills: To do well with AI or any area of big data, you need a mix of math, programming, and more. That kind of unicorn doesn’t readily gallop by. However, it’s also the case that finding someone who understands data science may be easier than finding someone who understands your business and the data that makes it hum. This calls to mind Gartner analyst Svetlana Sicular’s advice from years ago about big data: “Organizations already have people who know their own data better than mystical data scientists.” Therefore, look within your organization because “Learning Hadoop is easier than learning the company’s business.”
Many AI projects fail precisely because the technology is considered in a vacuum. As noted by Greg Satell in Harvard Business Review, any AI project should have a clear business outcome identified, with the right data culled to serve that end. This, in turn, requires (you guessed it!) involving smart folks within the enterprise who understand the business intimately and know where to find the best data.
AI, in other words, while ostensibly about replacing people, can’t succeed without involving your company’s best people. So get them involved sooner rather than later, with a high tolerance for failure as they (and the enterprise) learns from those failures how best to use AI within the context of a particular business need.