AI at scale: Propelling your organization into the next normal

Before COVID-19, it took years, even decades, for game-changing technologies to become truly useful to organizations, but the pandemic has motivated companies to speed up that transition by several years. The focus on digital will continue into the recovery, and artificial intelligence (AI) will be a game changer during the journey, helping companies to make better decisions about what their customers want, what their employees need, and the risks their organizations face.

In fact, McKinsey research has shown that AI and analytics have the potential to unlock a global total of $10 trillion to $15 trillion in value—and most of that potential value lies not in chatbots or autonomous vehicles but in the core activities of the business, such as sales and marketing or supply-chain management.

However, a recent McKinsey global survey indicates that most companies have far to go: only a small subset of companies, across a range of industries, are capturing significant bottom-line impact from AI. During a McKinsey Live webinar, experts Michael Chui and Helen Mayhew explained what these companies tell organizations about getting the most from AI.

After studying AI transformations in companies around the globe, Chui and Mayhew have observed that many businesses are able to successfully capture value from AI at the individual-use-case level (optimizing their logistics network, for example), but scaling is proving to be a challenge. The companies that receive the greatest bottom-line impact from AI at an enterprise level have a set of best practices that include: aligning their AI strategy with broader corporate strategy; finding, developing, and empowering the right talent; boosting organizational effectiveness by changing ways of working; employing advanced technology, tools, and pipelines; having a clear data strategy and using data to enhance deep learning; and changing organizational behavior to enable AI adoption.

In their presentation, Chui and Mayhew focused on a few underappreciated but important actions within these best practices:

• Scaling from lab to factory. Making the difficult transition from successful AI pilots to effective enterprise-wide AI requires an organization to adopt a scaling mindset; take a modular, or component-based, approach to building AI assets; and shorten the analytics-development life cycle by using MLOps and increase a model’s reliability by automating repeatable steps in the development workflow.
• Integrating humans and machines. Leaders must frame AI as a way to augment, not compete with, human capabilities, and vice versa. Human–machine integration requires companies to create processes that engage the entire workforce, thus developing AI literacy at every level of the organization. Often, this integration is more of a cultural than a technical challenge.
• Reimagining risk. It’s important to embed risk at every phase of the model-development life cycle. High-performing companies are more likely than others to recognize, mitigate, and prioritize AI risks such as cybersecurity, privacy, and physical safety. But they also recognize less prominent risks such as a lack of explainability, or an inability to explain to employees how and why AI systems make certain decisions. Another risk to be tackled often is bias that could limit an organization’s equity and fairness.

Questions and answers from the webinar

1. What are your thoughts about the common fear that people will lose their jobs to AI and machines?
   It’s rare for an entire job to be automated away. Particular activities of a job may disappear, but then the job will evolve, perhaps to include more interesting tasks. Over time, technology will create entirely new jobs and job categories. Our colleagues have been helping CEOs look ahead at the new skills, talents, organizational structures their workforces will need in five to 15 years, and our next McKinsey Live webinar will focus on this topic and the McKinsey Global Institute’s research on the future of work after COVID-19.
2. Could you elaborate on the tactic of using a "digital twin"?
   Digital twins are virtual replicas of physical assets that can be used to run simulations before actual devices are built and deployed. They are also changing how technologies such as AI, analytics, and the Internet of Things (IoT) are optimized. In the case of the sailing example we shared during the webinar, prior to the development of the digital twin, testing new boat designs required the team’s sailors to perform test runs in a simulator, which was inefficient and costly. The answer was AI bots used in the digital twin that could test new designs by sailing them on the simulator. Using the bots with the digital twin we built sped up testing and the rate at which designs are optimized by a factor of nearly ten.
3. Although AI is capable of a speed and processing capacity far beyond those of humans, it cannot always be trusted to be neutral and fair. How can we eliminate AI bias?
   Minimizing unfair bias is critical if AI is to reach its potential and increase people’s trust in AI systems. CEOs should provide guidance for translating company values, policies, and applicable sector-specific regulatory frameworks into AI development. Processes and practices must be established to test for and mitigate bias in both AI systems and human decisions. Underlying data rather than the algorithm itself are most often the main source of bias, and this bias can be amplified when machine-learning models are deployed at scale. To mitigate bias, companies must ensure that data sets reflect real-world populations. Above all, each organization should have a tailored, thoughtful fairness framework against which all AI models must be tested before being deployed in a live production environment.

* * *

For more on this topic, watch the webinar recording and read the articles “Global survey: The state of AI in 2020” and “Executive’s guide to developing AI at scale.”