So your company has finally decided to adopt AI. It works with a promising AI technology vendor. It also hires a few data scientists to help build the AI models and manages to obtain all the resources needed. After a couple of exciting months of putting the AI project together, results from the proof of concept are not only astonishingly good but also show the potential of what AI technologies can do to vastly improve your company’s performance. All the hard work seems to have paid off. Your company is now standing at the dawn of a new AI era, and may even become an AI leader in your sector.
And then it all stops. There is no more progress. The proof-of-concept results are mothballed. The AI project is declared to be a failure.
What’s being described here is not an imagined scenario. It is real and it’s actually happening to quite a lot of businesses. So, what has caused the uptake of AI to slow to a screeching halt? The answer: it is difficult—in fact, very, very difficult—to integrate AI elements into the business systems that can churn out the benefits of using these AI elements. When used in a vacuum, AI technologies confer few advantages to businesses; therefore, they must be used in conjunction with well-designed IT systems and infrastructure, a setup known as the “production environment.” As an analogy, imagine that your AI model is like a powerful car engine. What your business is truly seeking is how to get from point A to point B, but you need the rest of the vehicle to help you make the journey and maximize the benefit from the engine. No matter how powerful your AI model is, it is relatively useless to your business if it doesn’t get the surrounding environment right to ensure that the model’s output can be converted into successful business propositions. This is why AI operations—or “AIOps” for short—is becoming essential for businesses.
The importance of AIOps
The term “AIOps” has evolved from “DevOps”—a software engineering culture and practice that aims to integrate software development and software operations. The idea is to advocate automation and monitoring at all stages, from software construction to infrastructure management. Deploying AI is not very different. Harnessing the power of this technology requires companies to have strong AIOps capabilities, which involves systems developers and engineers building a reliable and rigorous production environment.
It is necessary to note that some people use the term “MLOps”—or machine learning operations—in the same vein as AIOps. However, even though MLOps as a term is gaining popularity, it is not necessarily a precise description. This is because an AI system often uses a lot of other technologies, such as combining machine learning technologies with rules-based algorithms. MLOps is therefore too narrowly focused, as it implies that the production environment is only important when machine learning is involved.
Why is AIOps a must?
There are three reasons AIOps is vital for companies wanting to be AI-powered. First, a significant business challenge when unifying AI technologies and company operations is that the data scientists who build AI are often far more interested in the models themselves than the overall production environment. In other words, they are mostly concerned with model prediction accuracy. However, the true value of AI comes not from the models but rather from having access to a capable team of systems developers and engineers who can build and maintain a production environment that draws on the power of AI models. These people are the ones who can convert model outputs into genuine business propositions: for example, turning the results of predictive analytics into customer experience enhancements, or automating the reading of documents of different formats and quality levels and entering the extracted information into a database.
It is therefore critical that any business seeking to successfully deploy AI has access to a competent AIOps team. Among other things, this team would consider issues such as how to make the production environment more robust, how to stop the production environment from crashing, and how to prepare for contingencies. Experienced systems developers will also be able to anticipate the problems that might surface within the production environment, and hence avoid potential downtime or crisis.
The true value of AI comes not from AI models but rather from having access to a capable team of systems developers and engineers who can build and maintain a production environment that draws on the power of the models.
The second reason AIOps is so vital to companies seeking to be AI-powered is that AIOps is not merely integral to building robust systems: just as important is their ability to enable the production environment to adapt to new business needs and technical requirements. This results in another practice borrowed from the software realm, “CI/CD”—the combined practices of continuous integration and continuous delivery/deployment. A good AIOps team will think through the planning, coding and testing of AI models (“CI”), as well as the releases, deployments, operations and monitoring of the IT infrastructure (“CD”). In short, AIOps allows the system infrastructure to be refreshed and updated, while enabling health checks to be conducted on an ongoing basis.
The third and final reason that AIOps is so important is that embedding AI models into IT infrastructure is often both a complex business affair and a huge technical feat. Fortunately, however, knowledgeable AIOps teams will know best how to divide and conquer to meet the complexities of the challenge. They can build the right architecture by dividing it into manageable chunks (for example, a part that supports the rules-based models and another part that supports the machine learning algorithms). They can also ensure that each of these separate pieces works seamlessly together and functions like a well-rehearsed orchestra.
The value of a vendor: Production environment as a service
So how should you as a company get around what seems to be the largest roadblock on the path towards successful AI implementation? What we have observed is that any company with the intention to deploy AI successfully must first take AIOps seriously. In fact, given the mission-critical nature of AIOps, we would argue that you should consider diverting resources away from data scientists and model development to invest in systems developers and engineers, which will enable proper operationalization and execution. However, this means you will have to build a team as well as acquire expertise in AIOps. In this scenario, the upfront economic and organizational costs are very unlikely to be small.
Alternatively, you can look at setting up a partnership with an AIOps vendor. A vendor should be able to provide the required expertise to construct and run a production environment that sits well within your IT infrastructure and can support your own AI models. This is called “production environment as a service.” With such a service, companies can own a robust production environment without the enormous resources necessary to run their own AIOps—much like the arrival of cloud technologies, which means businesses no longer need to buy their own servers. Production environment as a service can turn initial cash outlay into operating expenses, making it easier to cover the associated costs.
While media and academia often concentrate on how AI is an indispensable engine of business growth, they have missed the bigger picture. What truly matters is not the AI itself; rather, it is the well-oiled machine, powered by AI, that takes the company from where it is today to where it wants to be in the future. Companies that are serious about using AI to gain a competitive edge must take AIOps into consideration, otherwise any AI projects will start—and end—at proof of concept.
This article was written by Terence Tse, Takaaki Mizuno, Danny Goh, and Mark Esposito.