NEW YORK — While making sense of a crowded field of vendors, a skilled chief information officer can tune out the hype and focus on where the value lies. In few fields is this task more difficult than in artificial intelligence.
IT leaders, be advised. If there's one thing AI is not, is magic, said Jeff McMillan, chief analytics and data officer at Morgan Stanley, speaking at The AI Summit in New York in December.
"Too much of the conversation is about artificial intelligence and not the problem you're trying to solve," said McMillan. "If you think you'll solve this problem by plugging in somewhere you're wrong."
To determine where an AI undertaking could lead an organization, it's helpful to break solutions into pieces.
Here are 5 AI traps:
1. AI that can add value but lacks required data
AI is only as good as the data that powers it and, for many companies, mountains of data sit unused.
"If data [isn't] accurate, it's not predictive of the future, you got a problem," said McMillan. "It's like having a Ferrari and not having any gas. For most of you, if you're failing, this is where you're failing."
The importance of the correct data sets led companies to flip the process on its head. Experian focused on identifying first which what data sets they'd need for specific outcomes — improving financial inclusion or building new customer-oriented services, for example.
2. AI that won't do what it says it does
Vendors will smooth over difficulties of AI deployment and promise their solution is a quick fix for specific problems, even without knowing what data is available.
"This is AI that is lying to you," McMillan said. Currently, 25% of companies say half their AI projects fail to deliver on their promise, according to an IDC report.
To identify whether a provider has value to deliver, leaders should ask for tangible results.
3. AI that works but adds no value, solving no underlying problem
Ultimately, the goal isn't to innovate or to deploy AI but to deliver business value.
"That can come from very sophisticated tools and sometimes very simple ones," said McMillan.
It's risky to constrain innovation and experimentation, but especially when it comes to AI the experiments should be shaped around definable and measurable outcomes. Otherwise, innovation for innovation's sake will lead to sunken costs.
4. AI that can add value, but is too complicated to use
Platforms are only as good as the utility they provide stakeholders. If a deployment proves too complex for the desired users, industry risks losing employer buy-in.
One way to avoid taking on these types of projects is by simplifying the initial goal and scope of the project and going back to basics, said McMillan.
The barriers to accessing internal talent with knowledge and expertise in AI makes it easy to enter this kind of project accidentally. Without internal knowledge, it's hard to supervise the outcomes if the system is too complex.
5. AI that works and adds value, but costs too much
AI may be complex, but figuring out a worthy investment is simple arithmetic.
If the cost of deploying a solution exceeds the opportunity it represents, it's just not worth the effort, McMillan said.
Part of the strategy to sidestep this pitfall of AI deployment involves having courage to stop a project in its tracks in favor of a less complex, less costly approach if the option exists.
Here are 2 AI tips to avoid traps:
1. Old stuff rebranded as AI
The difficulty of sorting through vendors in a crowded market has the adverse effect of making technology less accessible to the average company.
But in certain cases, there are other ways to obtain key data insights that don't require the same resources as building internal platforms or turning to a vendor.
"AI is not magic," said McMillan. "It is just math. It's complicated math.
"Our focus needs to be figuring out what works, what doesn't, and then how to move our business forward," he said.
Using means, medians and standard deviations can often glean actionable insights and turn them into competitive advantages.
"If you are not training and educating your employees to understand what a correlation coefficient is, understand how to interrogate the data, and understand how to look at outliers, you're doing an enormous disservice to your business," McMillan said.
Don't dismiss this kind of approach to AI, as it has the ability to deliver results to a business.
2. Real, usable, achievable and affordable AI
Identifying an AI project that checks all the boxes must start with attainable goals that deliver business value and engage with its stakeholders right away.
For Morgan Stanley, one example of AI that fit this elusive category was its Next Best Action system, which uses machine learning to offer financial advisors insight into which actions are likely to generate success for specific clients.
One differentiating factor? The project started with an accessible data set which held historically accurate information on the outcomes each decision has yielded in the past.
"I could give you my algorithms," said McMillan. "I could hand them to you on the way out. And you know what? You couldn't do anything with them. Because you don't have my data set."
Clarification: This article has been updated to reflect AI traps to avoid.