Compared to e-commerce enterprises and companies that deal with clickstream data, IT management software companies operate in relatively low data environments. Generally speaking, IT management software vendors don’t have graphics processing units on their customers’ on-premise environments, which can make it difficult to roll out AI tools.
Building out effective AI tools for smaller companies with low data environments and minimal processing capabilities can be a challenging endeavor. However, by accessing data from external sources, you can build out effective AI tools for your software.
Bootstrap your model using globally available data
When starting out, one particularly effective technique is to use globally available data, such as material from publicly available restaurant and movie review websites. For example, when creating your AI model and beginning your sentiment analysis process, you can bring in up to 50 percent of your data from movie review websites.
During this process, remember to remove salutations, and any words, articles, or phrases that don’t contribute to sentiment.
Slowly deploy AI tools into your customers’ environments
While A/B testing your AI model, we recommend initially opening it up to roughly 20 percent of your users. Throughout the quality insurance process, be sure to adjust the algorithm parameters; after that, adjust the algorithm itself, and then balance out the dataset.
In regard to balancing out the dataset, let’s look at a quick example. During the algorithm training process, perhaps half of your dataset is comprised of spam emails and the other half is legitimate. In reality, however, perhaps only about 5 percent of emails in your customers’ environments are likely to be spam. In this case, you’d need to balance out the dataset to account for this discrepancy.
Provide confidence intervals for each and every recommendation from the AI system
It’s vital that any AI recommendation be explainable. That is, the AI system must explain exactly why it has recommended any given course of action, and these explanations should come with confidence measures. If the AI tool is 80 percent confident in a recommendation, it should certainly recommend this course of action. Alternatively, if the confidence level is in the 50-80 percent range, it should frame the recommendation as a suggestion. Lastly, if the confidence level is less than 50 percent, it’s inadvisable to share the recommendation at all.
AI has to adapt to changes
Generally speaking, any well-tuned AI system should be at least 80% accurate. Also, if an AI offers a recommendation with a 90 percent confidence level, and the user says it’s incorrect, this feedback should be taken very seriously.
Within the realm of IT management, there isn’t a one shoe fits all, silver bullet method, as every customer’s environment is different. The data within these environments constantly changes, especially if you are engaged in IT operation management or infrastructure monitoring.
By incorporating data from external sources, you can bolster your AI tools
Ultimately, 50 percent of your data may come from the globally available data (e.g., Amazon reviews); another 25 percent can originate from your historical service desk data, and the last 25 percent could come from your customer’s environment.
In addition to using globally available data from e-commerce or movie review sites for sentiment analysis purposes, you can also rely on other available data. As an example, let’s say you have a website monitoring solution running on a Linux server. If you’re trying to predict outages in this Linux server due to a stack overflow problem, there are only a few places that might be responsible.
Even if this customer’s Linux machine has never gone down due to a stack overflow issue before, you can still effectively predict this outage because you’ve trained your AI tools on an external dataset that has dealt with such issues.
As another example, perhaps you’re running log management software for the first time on a customer’s environment in the Middle East; even without a great deal of data, your AI tools can make decisions based on historically standard patterns of security breaches. In the Middle East, the workweek is generally Sunday to Thursday, so a log-in on a Saturday may trigger an alert.
By using publicly available, external sources, you can train and bolster your AI tools—even in low data environments with minimal processing capabilities. And remember, once your AI tools are up and running, identify where the AI recommendations are falling outside of your desired confidence intervals, and then adjust your processes accordingly. If your AI system is consistently delivering recommendations with at least 80 percent accuracy across the board, you can rest assured that your AI system is working well.
Ramprakash Ramamoorthy, product manager for AI and ML at Zoho Labs, is in charge of implementing strategic, powerful AI features at ManageEngine to help provide an array of IT management products well-suited for enterprises of any size. Ramprakash is a passionate leader with a levelheaded approach to emerging technologies, and a sought-after speaker at tech conferences and events.