In November 2022, OpenAI’s ChatGPT introduced the world to the power of generative AI (gen AI). Since then, companies have been scrambling to respond and capture their share of the estimated $2.6 trillion to $4.4 trillion in new value potential offered by this revolutionary technology.
That hasn't been easy, especially in operations. But research by MIT’s Machine Intelligence for Manufacturing and Operations (MIMO) and McKinsey has found emerging evidence that leading companies are now starting to generate value by applying a range of AI technologies to manufacturing, back-office processes, and other operations functions.
For example, a top 10 retailer by global revenue developed an in-store chatbot for associates at its nearly 2,000 retail locations. The chatbot makes thousands of pages of best-practice manuals easily accessible to store associates, reducing the time they spend on the phone with internal service centers to get questions answered. Ultimately, this can reduce training time for new employees and may lessen the impact of employee turnover.
Elsewhere, a global pharmaceutical company is using gen AI to verify that supplier invoices comply with contractual terms. The company’s R&D division spends more than $4 billion a year on external products and services, and its contractual relationships are complex, involving variable discounts for different scopes of work, multiple currencies, and different arrangements for inflation. The prototype tool, which replaces labor-intensive manual invoice analysis, can extract invoice line items from PDF documents with 95 percent accuracy. In just four weeks, the new system identified more than $10 million in value leakage, an average of 4 percent of the spend analyzed. Even better, the new tool also highlights recurring spend items not covered by contracts, giving procurement teams the opportunity to negotiate better deals.
Our research also suggests that AI isn’t easy to get right. Plenty of AI initiatives fail to live up to management expectations, and many companies are struggling to integrate these technologies into their processes, or scale them up across their operations. Those challenges contribute to the widening gap in AI adoption and impact between leading companies and the rest. In our previous study of AI in operations, we found that the performance improvement enjoyed by leaders in AI adoption was 2.7x higher than that of companies in the bottom half of the table. Our latest study finds a 3.8x performance gap (Exhibit 1).
The data for the new study comes from a detailed survey of how more than 100 companies have implemented AI in their operations over the last two years, coupled with in-depth interviews with 15 of them, in industries ranging from insurance to steel manufacturing. While some of the larger performance gap can be explained by higher investments in digital technologies, the latest research also identifies four critical factors that separate today’s AI leaders from the rest: executive sponsorship, a maturing ecosystem of partners, cross-departmental collaboration, and data investments.
Executive sponsorship keeps projects on track
The leading companies in our latest study have made progress in addressing some of the key challenges that were holding back their AI and machine learning (ML) ambitions. Lack of expertise and cultural barriers to AI adoption, the top issues in our previous survey, were mentioned by less than half the leading organizations in our latest research.
Other problems have arisen to take their place. Difficulty quantifying the financial return on investment was seen as the biggest barrier to successful AI/ML investments, followed by limitations on the time and resources available for projects (Exhibit 2). Uncertain ROI is a perennial challenge for digital projects: New technologies may take longer to implement than expected; benefits such as time savings for salaried employees may not immediately translate into revenue numbers; and “shiny” new technologies may turn out to be a poor fit for real-world business problems.
Despite these obstacles, some of the companies in our study estimated that their returns were five times the cost of their digital projects, less than five years after implementation. What are they doing differently?
A large multinational industrial manufacturing company that supplies systems and components for transportation and industrial markets wanted to automate its advanced process control (APC) manufacturing operation with AI-driven decisions. The first attempt failed. The external partner chosen to help with the project didn’t understand the nuances of the company’s continuous flow manufacturing process, and commercially available AI and ML tools couldn’t provide the speed of response the manufacturer needed.
Fortunately, the company’s chief technical officer (CTO), convinced of the significant potential of AI in manufacturing, realized that integrating new tools into the existing manufacturing process would require bringing development in-house. The company eventually built its own AI-driven APC model, which ran ten times faster—and was ten times cheaper to operate—than the preexisting system. The approach was so successful that the company has since spun it off into a subsidiary start-up that provides a subscription-based process control service to other manufacturers.
The road to success, however, was anything but straightforward. Development teams spent a year and a half proving the technical feasibility of the solution, then another year and a half proving they could build the necessary data infrastructure and scale the solution into an enterprise-wide application. It took significant R&D investment before the project began to pay off. Only the unwavering support of the CTO kept the project alive during that time.
This story is one we hear repeatedly from leaders in the field of AI/ML in operations. Companies that have achieved great returns often do so with only a high-level business plan to guide them. Rather than devoting time and resources to the same detailed ROI calculations they would for traditional projects, leaders use industry knowledge and domain expertise to identify areas of high potential, but often move forward without explicit expectations about when they will see a return on their investment.
Persuading the board of a large multinational to invest in a large project with uncertain timelines and returns is difficult without the support of an executive champion—a leader with the technical skills and internal political power to see the project through from concept to delivery. In our new study, 77 percent of ML implementation leaders had C-level leadership driving their projects. For 44 percent of leaders, digital projects were sponsored by the CEO or board of directors—more than double the rate of companies in the bottom 50 percent of our sample and a 17-percentage-point increase from our previous study (Exhibit 3).
Partnership ecosystems are growing up
Leaders are becoming more confident in their own abilities. In our study, 89 percent of leaders say they are now using internal development capabilities to build AI and ML solutions, compared to just 50 percent of companies in the bottom half of our sample. Similarly, the share of executives citing lack of expertise as a barrier to implementation dropped to less than 50 percent in our latest study, compared to nearly 70 percent in our previous study.
Yet even with strong internal software and data science capabilities, more than two-thirds of leaders in our latest study still use external partnerships to develop ML solutions, while only 50 percent of companies in the bottom half of the table do the same. It appears that leaders are better able to recognize their internal limitations and see the value in bringing in external expertise and resources. In addition, the nature of partnerships is changing. Collaboration with academia and start-ups has decreased—falling from 83 to 50 percent of leaders—while leaders maintain their reliance on consulting firms, vendors, and industry partners (Exhibit 4). This indicates that the market for AI partners may be maturing, away from “experimental” approaches and toward the mainstream.
Leaders approach these partnerships in a variety of ways. Some work with external partners to help define the strategic direction for priority ML projects and build proofs of concept for identified use cases. Others prefer to outsource the development and maintenance of their AI solutions, either to free up internal resources or to gain access to cutting-edge technologies.
An example of the latter approach is a major metals manufacturer that has partnered with an AI company for an alarm system that monitors its manufacturing operations. The metals manufacturer uploads real-time data to the AI company’s proprietary anomaly detection model. The model then sends alerts back when it predicts potential problems with the process or equipment.
Another trend in partnerships that emerged from conversations with leaders is cross-industry collaboration. Leaders are developing their AI expertise by collaborating with companies in other industries through conferences, journals, and even one-on-one meetings. These relationships drive the spread of knowledge across industry boundaries. In many cases, leaders have successfully applied analogous ML use cases from other industries directly to their own operations.
One example comes from a large mining company. Before investing big in AI, for example, the mining company visited pharmaceutical companies to learn about their experiences and apply learnings on using AI to map out molecules, to its own work mapping chemical compounds.
While it is important that companies use the right partnerships to drive results from AI deployments, the leaders we spoke with excelled at thoughtful collaboration—for instance, starting projects with clear goals for partners, and ending them with a thorough knowledge transfer.
In the case of our large multinational industrial manufacturing company’s initial AI efforts, the vendor it contracted was too focused on the data science side of the problem, as opposed to the operational requirements of continuous flow manufacturing. To prevent such gaps in understanding, it is essential to discuss solution objectives and requirements early and in detail.
There are different mechanisms for doing this. For example, one tech company develops documentation with its partners at the beginning of each AI solution. In this exercise, project managers sit down with customer leadership and co-author a document that frames what the solution would look like if it was complete today. Through exercises like this, leading solution providers work to understand the outcomes and metrics that are most valuable to their customer’s business. The two can then develop a common set of project goals.
Another critical factor in successful partnerships that leaders focus on is thorough knowledge transfer at the end of the engagement, especially when that engagement involves the development of an AI/ML prototype or proof of concept. In such cases, the vendor understands the model and knows how to update it, but the customer may not. This reality of external partnerships requires that a well-thought-out model transfer process be built into the project workplan. This often takes the shape of an internal resource or team that knows how to access, maintain, and train on partner-developed models, so that the full value of the solution will not be lost when the engagement ends. In contrast, less successful firms often bring in partners because of a lack of internal resources and AI leadership. After the initial excitement of new pilots and use cases subsides, there is unfortunately no one to pick up the reins when the partner exits.
Cross-functional collaboration is key
Leaders have demonstrated an increased ability to coordinate effective collaboration among the many internal stakeholders impacted by AI/ML solutions. This challenge cannot be overstated. A senior executive at a leading technology provider, whose organization has helped a significant number of companies through their digital transformations, says that the biggest barrier to overcome is team collaboration, particularly between operations technology (OT) and information technology (IT). Leading companies have shown time and time again that they can effectively manage these and other disparate teams and reap the outsized benefits they produce. This collaboration is also enabled by top-down sponsorship.
Leaders adapt their organizations to ensure that projects and initiatives are implemented consistently across them. One common approach is to create a center of excellence (COE), a cross-silo internal organization staffed with a critical mass of employees with data science skills. The COE ensures that AI projects get implemented efficiently and deliver value, while also addressing issues such as cybersecurity, data error, and compliance. COEs typically set common practices and standards, and keep the talent pipeline filled, through both hiring and capability building. They can leverage the greater collaboration, talent, and diversity of experience that come with a larger team.
An alternative approach is to create dedicated cross-functional teams of data scientists, operational gurus, and engineering experts within a business unit. Sometimes co-located at a production site, this group is given a mandate to address high-priority topics for the business unit, but with lower standardization with other parts of the company.
In both these models, companies are constantly trying to bridge data science skills with operational expertise. Some leaders include experts from the company’s technical operations functions within their central AI teams. They find that this helps foster mutual understanding across traditionally siloed business units.
Data matters: Quantity, quality, and accessibility
Finally, a company’s approach to data management has emerged as a critical determinant of success. The difference between industry leaders and the rest is not simply the accumulation of data, but its careful management and availability for strategic use. While many companies collect data, leaders differentiate themselves by maintaining data that is accurate, well-formatted, at the appropriate granularity, and available across the organization to make informed decisions that drive efficiency and innovation.
In manufacturing, data is often either incomplete, inaccurate, poorly formatted, or not granular enough, resulting in missed opportunities or poor decisions. The foundation of effective decision-making lies in data integrity and accuracy, and leaders appear to invest in systems and processes to ensure data is well maintained and reflects their operations. At the core is data collection—58 percent of leading companies in our latest study collect data from more than half their equipment, a practice significantly less common among lower-performing peers (Exhibit 5).
For example, a global cement manufacturer with operations in more than 25 countries exemplifies the power of well-maintained and strategically applied data. The company embarked on a digital transformation with the goal of not only collecting more data about its operations, but also ensuring that all data is accurate, organized, and actionable.
To build its data management capabilities, it integrated thousands of sensors into plant equipment, invested in cloud-based data storage, connected all its plants, and implemented cybersecurity measures to ensure those connections were protected. Responsibility for data management was given to a centralized team that has grown from three people to more than 30 data scientists and engineers today.
By making accurate operational data available and using new analytical methods, the company has been able to better predict potential equipment failures before they occur, significantly reducing downtime and extending equipment life. In addition, by carefully tracking and analyzing energy usage and waste generation, it was able to significantly improve both. Together these tools have returned five times the company’s digital investment.
It’s not too late to catch up
As the performance gap widens, some firms will be left behind in the race to apply AI, ML, and data to operations. There is still an opportunity, however, for companies to rethink their operating models and become leaders in this space. While today’s leaders are outperforming the rest in terms of ROI, payback periods are shortening. In our previous study, payback periods ranged from 12 to 18 months for leaders and 18 to 24 months for the bottom 50 percent. In our latest data set, the payback period had shrunk to six to 12 months for both groups (Exhibit 6). As companies adapt their operating models to support ML initiatives from the C-suite, leverage strategic partners, build connections between data science and operations groups, and establish reliable and accessible data management processes, we expect to see further significant performance improvements.
Moreover, the widespread availability of off-the-shelf solutions, analytics tools, and vendors with micro-vertical solutions is lowering the barriers to entry. Easy-to-use systems now allow data science novices to implement ML tools. As one executive notes, “We don’t have data scientists building models [anymore]—everyone is working on MLOps [machine learning operations], automating data processes, and building cloud apps.” If companies can train business leaders to use these tools, it may soon be unnecessary to hire large numbers of data scientists, which has been a daunting challenge.
At the large global retailer, the rollout of gen AI was made possible by significant prior investments in digital capabilities combined with a strong digital operating model. The initiative is led by the retailer’s SVP of data science, who reports directly to the CIO and has significant executive sponsorship. The company has established partnerships to leverage the tools and expertise available in the market. In addition, by bringing together key teams within the company and letting business units take the lead in generating ideas and identifying value for AI use cases, it has been able to move quickly on promising projects, such as the AI chatbot. Now, it remains to be seen whether the rest of the pack will use the lessons learned here to catch up.
