What exactly is AI? A common definition is it's the simulation of human intelligence in machines that are programmed to think as humans do and impersonate their actions. AI can refer to any system that exhibits traits associated with the human mind, such as the ability to learn and solve problems.
Technology products and services bolstered with AI capabilities can lead to significant benefits for organizations: increased efficiency, automation of manual processes, enhanced decision making, improved customer service and experience, and the ability to solve complex problems—to name a few.
However, IT and business leaders need to be adept at separating the hype from the reality. As research firm Gartner notes, many organizations struggle to put a realistic value on AI. Business leaders tend to overestimate the impact of the technology and underestimate its complexity, it says.
The firm in early 2020 published a report on AI in organizations, based on a survey of more than 600 senior-level executives from organizations in the U.S., Germany, and the U.K. It found that more than one third of the organizations have deployed AI and are using it today, more than half planned to deploy AI in less than 12 months, and about one third planned to deploy AI within two years.
On average, organizations take nine months to develop AI initiatives from prototype to production, according to the report. Despite the widespread shortage of skilled workers in AI, a lack of talent is not the main barrier for the successful deployed of AI. The limiting factor of AI applications success is the lack of clear outcomes, it said.
Engineering a strategy
Enterprises also face challenges involving integration, security, and privacy issues that prevent them from efficiently moving their AI practices from prototypes to production, Gartner says.
That's where AI engineering can play a key role. AI engineering “is the processes, tools, and best practices of how to design, build, test, deploy, operate, and evolve reliable AI systems,” says Matt Gaston, director of the Emerging Technology Center at Carnegie Mellon University's Software Engineering Institute's (SEI) and leader of the National AI Engineering Initiative.
“AI itself is about specific AI techniques to create intelligent models or capabilities; for example machine learning,” Gaston says. “AI engineering turns those intelligent models or capabilities into fully functioning systems for providing business, operational, or mission solutions.”
AI engineering brings together the various disciplines from across an organization to provide a clearer path to attaining value when operationalizing the combination of multiple AI techniques, says Erick Brethenoux, vice president and analyst at Gartner, which named AI engineering one of the top strategic technology trends for 2021.
A robust AI engineering strategy will facilitate the performance, scalability, interpretability, and reliability of AI models, while delivering the full value of AI investments, says Brethenoux, who works with clients to adopt AI engineering strategies. “AI engineering includes the methods, the best practices, the capabilities, and the techniques to put AI in production; to govern, manage, and monitor its use” across the enterprise, he says.
Typically AI can be segmented into two main cycles, Brethenoux says. One is the development and training of machine learning models, for example rule-based models, optimization models, or natural language processing models—depending on the type of application. The second cycle is the implementation of the models within the organization.
“Companies have been notoriously bad at operationalizing AI models,” Brethenoux says. “They are great at developing models; plenty have developed lots of models. But going into production [with the models] is hard. We need a way to systematize the production of those models, and that's where the engineering comes in.”
AI engineering should not be an informal or ad hoc effort at operationalizing AI for a specific project, Brethenoux says. Rather, it should be formalized across the organization so that it becomes a best practice for all use of AI and machine learning.
Gaston notes that AI engineering “is a nascent discipline that will grow and mature over time and keep pace with the rapid innovations we are seeing in the field of AI. That said, organizations [are using] AI engineering to create and maintain business solutions that incorporate AI techniques.”
This includes ensuring that the system requirements are driven by business needs, that appropriate AI techniques are selected for those requirements, that testing and monitoring tools are in place to ensure that the system continues to function as desired, and that processes and frameworks are in place to update and evolve the AI system as requirements or the operating environment change.
Main components
One of the keys to successful use of AI is recognizing that moving to the operational phase is not just a technical issue, but a business one as well.
“When you put AI into production, you have to talk to subject matter experts about the applications they are running, the use cases for AI, and how these applications serve the business,” Brethenoux says. “Then you have to know whether the key performance indicators in place are actually working,”
Organizations need to “put the business users at the heart of their AI engineering practice” in order to ensure that AI delivers on what the organization is trying to achieve, Brethenoux says.
That's not to say technology isn't important. AI teams need to be discerning when it comes to selecting tools and techniques to ensure they deliver on what the company is looking for with AI, and that they can be easily integrated with existing products as well as other new AI tools.
Another important consideration is the quality of the data that companies will feed into models. “Where is the data coming from and how reliable is it?” Brethenoux says.
A relatively new discipline that can help organizations improve data quality is DataOps, an automated, process-oriented methodology used by data analytics teams to improve the quality of data and reduce the cycle time for analytics via the Agile methodology.
DataOps applies to the complete data lifecycle, from data preparation to reporting, and it acknowledges that the data analytics team and IT operations are interconnected.
Organizations looking to leverage AI also need to have the right skills in place. These include experience with programming languages such as Python, R, Java, C++, and others; machine learning including algorithms and libraries; data science; neural network architectures; applied mathematics including probability and statistics; robotics; computer vision. Professionals working on AI projects also need soft skills such as good communications and collaboration, critical thinking, and creativity.
One of the newer related skills emerging is machine learning validation, Brethenoux says. Model validation is the process in which a trained model is evaluated with a testing data set. The main reason for using the testing data set is to test the general ability of a trained model.
“Those people are going to be involved throughout the lifecycle of AI models, validating from a data perspective, and then from a development perspective, and then from an operational perspective,” Brethenoux says. “They will evaluate whether everything being done with the models is valid, legal, and technically feasible.
AI skills are not as difficult to acquire as many organizations think, Brethenoux says. Having the right talent discovery programs internally, the proper upskilling and education practices, and a select external talent mix is all they need to be successful, he says.
A way forward with AI
Many companies are realizing that AI can provide real business value in many ways, and are starting to build up their data science, machine learning, and AI teams, Gaston says.
“While these are important first steps, the function of most data science or machine learning teams is to produce models using data in very controlled environments,” Gaston says. “It is critically important that companies look more holistically at how to realize, operate, and maintain AI for business solutions and [they] will need to bring a diverse set of talent, thought, and experience to bear in doing so.”
Organizations can move ahead with AI projects without creating an AI engineering strategy. But if they do so the likelihood of long-term and comprehensive success with the technology might be limited. It's too strategic and powerful a tool to leave to haphazard approaches.
“I think this is the only way to get to get AI issues resolved,” Brethenoux says. “If you're not doing AI engineering then I'd say only about 35 percent to 45 percent of your models will find their way to production. So you're going to waste the rest the rest of them. Maybe 65 percent of your models are going to be wasted and then you're going to have to start from scratch every single time.”
One of the main issues with AI will continue to be moving from the development and training phase to the operational/product phase. And that will continue to be a problem for companies if they don't take a structured, engineering approach to AI.
“The organizations that have done AI well are doing it more systematically, and they are reaping really nice benefits out of it,” Brethenoux says. “I'm almost tempted to ask why do AI at all if you're not planning to reach the operational end? Before that it's really just an interesting intellectual exercise. When you deploy AI then it becomes an economic advantage.”