Using artificial intelligence (AI)—and especially one example of AI: machine learning (ML)—is all the rage these days with enterprise IT. But could it also turn the reactive "have you tried restarting?" corporate helpdesk into a mechanism that could anticipate and predict technology problems before they're readily apparent? Quite possibly.
What brings this to mind is a recent attempt by Facebook to use an AI scavenger hunt to train robots to better understand physical homes and where items are most likely to be situated. That approach is quintessential ML, where the software is given a vast amount of (seemingly) unrelated data and the software tries to make sense of it all. Could the same approach that trained robots to do their thing also help IT predict problems courtesy of helpdesk activity analysis?
At its most simple level, machine learning is throwing a mass of data at software and letting the software try and find patterns and deviations from patterns. Theoretically free of most human thinking biases, the software will often discover patterns that few humans would have detected.
What if IT applied that ML model to the petabytes of data that IT helpdesks receive every day? Not just help tickets, but every text question, every comment during a phone call. If grabbed and analyzed in real time, would certain kinds of questions flag the earliest stage of a D-DOS attack? Perhaps even at an early enough stage where the bulk of the attack could be thwarted? More innocuously, could the inquiries suggest restarting a server before serious problems happen?
Bryce Austin, a security consultant who wrote the book Secure Enough?: 20 Questions on Cybersecurity for Business Owners and Executives, said that he thinks the approach has potential.
"If applied to the unstructured data contained in customer service call notes, ML could bring about a new generation of UI tweaks and other end-user experience changes that will reduce calls to help desks and provide a better customer experience," Austin said. "The trick will be to give the ML engine access to end-user application usage data and customer service call data. If it can find interesting correlations, users and software developers will all benefit."
Another security consultant, Virginia Tech Professor Wade Baker, said the idea has potential depending on how far companies are willing to push the data. "The information related to someone calling into the helpdesk, those things make for excellent pre-warning signs. This is especially true if it's outside the flow of what a helpdesk person typically would be following," Baker said. "I would definitely expect ML to see patterns that are highly predictive and diagnose some pretty interesting things."
Columbia University Professor Sal Stolfo said the technology makes such predictions practical. "I don't know if anyone has achieved that level of utility using ML automation, but it seems well within the range of possible provided ticket system logs that are available today," Stolfo said.
Security consultant and podcaster Ashwin Krishnan, however, said the benefit would come not merely from the analysis and the recommendations, but from how willing enterprise IT executives are to accept those recommendations and to do so very quickly. That, Krishnan argued, could go against the psychological grain of many in IT.
"If the data model is sufficiently broad, this could have substantial value," Krishnan said. "Where the challenge is going to be is with the IT psyche. This is a psyche that is used to putting out fires, to acting very reactively."
Krishnan likened such a helpdesk predictive analytics model to weather systems that predict weather disasters. Most of the time and for most consumers, the weather system's tornado warning predicts a disaster that doesn't happen. He added that even though he lives in an area with multiple geological faults, "I've never done an earthquake drill in my life."
Helpdesk analytics could reveal much. For example, a large number of users complaining about slow web responsiveness might reveal a drain on resources. But is it local (the user's machine) or a LAN or WAN issue? The distribution and severity of such complaints—and how sharp a spike of the number complaints the system sees per minute—could provide helpful clues. Is it one bandwidth-intensive application running or is it something more ominous?
Even a rash of phishing efforts aimed at getting employees to click on email links could reveal much. Are the complaints coming from employees in similar roles—say perhaps finance—across a wide range of geographies? That might suggest not only that the company is under a coordinated attack, but what the specific target is—for example, an attempt to steal company financial details.
But there's no reason to stop at helpdesks. ML analysis could unlock predictive clues by examining almost every area of corporate life. Feed it a few years of data on conference room scheduling and let it recommend if some of those rooms could be reassigned without invoicing employees? Explore all accounts payable records to seek repeated activities that could be automated for more efficiency? What about maintenance? Would a cross-compare of all repair requests against hours worked point to other efficiencies?
ML can be used to find the hidden patterns in almost every corner of the enterprise. Perhaps the ultimate ML application is to let the software review all enterprise assets and recommend where ML would work best?
Evan Schuman writes a weekly column for Computerworld and security pieces for SCMagazine and PCMagazine. He also moderates podcasts, webcasts, and live events on B2B tech topics.