Start adding items to your reading lists:or
Save this item to:
This item has been saved to your reading list.
Putting AI to work means fusing it with analytics, IoT and other enterprise systems and having the roles and processes to keep it running. This is one of five PwC priorities for making the most of AI in the coming year.
AI doesn’t do its best work when it’s isolated from other technologies, or when it’s siloed in a lone function or business line. First of all, AI needs data, and as it gets more quality data from more sources, it gains power.
Secondly, some of AI’s most valuable uses come when it works 24/7 as part of broader operational systems, such as marketing or finance. AI leaders are therefore operationalizing AI, across multiple functions and business units, in full integration with broader automation initiatives, data analytics or both.
Given that approach, it’s no surprise that the top-three AI data-related challenges all have to do with different kinds of integration: integrating data from across the organization (45%), integrating AI and analytics systems (45%) and integrating AI with IoT and other tech systems (43%). To solve these and other challenges as you make AI operational, it’s critical to realize that AI development is very different from software development and requires a different mindset, approach and tools. Whereas software development is based on rules of coding, AI model development requires a “test and learn” approach, in which the algorithms are continually learning and the data is being refined.
Although data is the key to operationalizing AI, it’s unfortunate that—same as last year—labeling it is low on executives’ priority list: Only one-third of respondents said it’s a 2020 priority, and just 13% view it as a key challenge. Even if you’re currently focused on bringing AI to a single function or process, it’s essential to begin cultivating secure, quality data from throughout (and outside) the organization. Likewise, you need to build the skills and the enterprise-wide governance to use that data responsibly. Indeed, the ability to show how data may (or may not) be securely and ethically used can be critical to getting a use case approved.
One way to meet the data labeling challenge is active learning: Data scientists do their work, and—by labeling and revising algorithms’ decisions and recommendations—teach machines to start doing it for them. Another approach is to take advantage of cloud-based services that include data sets, which can enable companies to capitalize on analytics and AI quickly.
AI development is very different from software development and requires a different mindset, approach, and tools.
Embed AI into your overall IT stack. Incorporate AI models that are responsible for automation or key decisions, while interfacing trained AI models into production applications to scale up use. This embedding of AI into IT should also support a common AI services layer that allows any application to integrate with AI models.
Develop machine learning operations. The key to making AI part of daily operations is a new capability, MLOps, which combines expertise in data science with software engineering and IT operations. For an effective MLOps function, most companies will need to hire and upskill talent.
Make your data trustworthy. To make AI operational at scale, it needs data that is not just accurate, but standardized, labeled, complete, free of bias, compliant with regulations and secure. Only then can you trust your data—and the results of AI models based on it.