A few months ago I read Kavita Ganesan’s excellent book, The Business Case for AI: A Leader’s Guide to AI Strategies, Best Practices & Real-World Applications. I highly recommend it to anyone thinking critically about implementing an AI solution in IT operations because the concepts and best practices she outlines helped me to think more practically about implementing AI solutions in IT operations.
Recently I’ve been having conversations with IT (mostly networking) teams about incorporating AI solutions, and the first questions I have are:
1) What problem are you actually trying to solve with AI (that AI would solve better than a simpler/cheaper solution)?
2) What’s the nature of the available data? This includes follow-up questions like: is network and network-related data stored in disparate locations? Has it been cleaned and pre-processed for analysis? How and how often is the data updated? Do you have different formats of data? How is data ingested and stored?
3) What’s the makeup of your technical staff? Does your network team contain or have access to developers, data engineers, perhaps even data scientists?
The answer to the first question is critical. For some, AI is simply an initiative handed down to operations teams by leadership with no clear direction or no clear problem to solve. Also, sometimes the direction and problem are clear, but it can be solved more easily and for far less cost than a sophisticated AI solution.
The answers to the second group of questions are just as important. The data we use in network operations is both structured (metrics, flows, etc), and unstructured (tickets, email, TAC call transcripts, certain types of logs, etc). The workflow to deal with this in a unified manner is more complex than many realize and requires serious planning, software development, data engineering, and perhaps data scientists.
Much of the data we deal with in NetOps is real-time telemetry requiring significant effort to clean, process, store, and analyze quickly enough to be useful for a production environment. This is no small feat, and addressing these data issues can end up being 80% or more of an AI implementation project.
The answers to third question will shape how we approach the build vs. buy discussion and how feasible an AI initiative is in the first place. There’s real ROI to consider, whether that’s ROI from decrease in current network operations cost, or a perhaps a direct increase to revenue from some new product or feature. Getting there will dedicated require staff – expensive staff – and a significant amount of time and effort. PoCs will require careful planning, timelines, and measureable results.
Kavita explains that successful AI solutions are by nature iterative in that they require constant tweaking. I would argue that any good software solution is iterative to one extent or another, but certainly I agree that an AI solution is much more so. Her perspective is practical and based in real-world scenarios, something I appreciate very much.
I’m excited about the near and long term potential of incorporating AI solutions into IT operations to solve real problems and to provide real value, and I’m spending a significant time and effort researching and experimenting with applying AI to what we do in NetOps. However, I’ve found it’s essential to temper curiosity and experimentation with an understanding of the use-case, cost, and tradeoffs of implementing AI in our beloved niche of tech.
You can check out and purchase Kavita’s book here.
Thanks,
Phil
{networkphil} 
Leave a comment