Notes from the AI frontier: Applications and value of deep learning

Artificial intelligence (AI) stands out as a transformational technology of our digital age—and its practical application throughout the economy is growing apace. For this briefing, Notes from the AI frontier: Insights from hundreds of use cases (PDF–446KB), we mapped both traditional analytics and newer “deep learning” techniques and the problems they can solve to more than 400 specific use cases in companies and organizations. Drawing on McKinsey Global Institute research and the applied experience with AI of McKinsey Analytics, we assess both the practical applications and the economic potential of advanced AI techniques across industries and business functions. Our findings highlight the substantial potential of applying deep learning techniques to use cases across the economy, but we also see some continuing limitations and obstacles—along with future opportunities as the technologies continue their advance. Ultimately, the value of AI is not to be found in the models themselves, but in companies’ abilities to harness them.

It is important to highlight that, even as we see economic potential in the use of AI techniques, the use of data must always take into account concerns including data security, privacy, and potential issues of bias.

1. Mapping AI techniques to problem types
2. Insights from use cases
3. Sizing the potential value of AI
4. The road to impact and value

Mapping AI techniques to problem types

As artificial intelligence technologies advance, so does the definition of which techniques constitute AI. For the purposes of this briefing, we use AI as shorthand for deep learning techniques that use artificial neural networks. We also examined other machine learning techniques and traditional analytics techniques (Exhibit 1).

1

Neural networks are a subset of machine learning techniques. Essentially, they are AI systems based on simulating connected “neural units,” loosely modeling the way that neurons interact in the brain. Computational models inspired by neural connections have been studied since the 1940s and have returned to prominence as computer processing power has increased and large training data sets have been used to successfully analyze input data such as images, video, and speech. AI practitioners refer to these techniques as “deep learning,” since neural networks have many (“deep”) layers of simulated interconnected neurons.

We analyzed the applications and value of three neural network techniques:

• Feed forward neural networks: the simplest type of artificial neural network. In this architecture, information moves in only one direction, forward, from the input layer, through the “hidden” layers, to the output layer. There are no loops in the network. The first single-neuron network was proposed already in 1958 by AI pioneer Frank Rosenblatt. While the idea is not new, advances in computing power, training algorithms, and available data led to higher levels of performance than previously possible.
• Recurrent neural networks (RNNs): Artificial neural networks whose connections between neurons include loops, well-suited for processing sequences of inputs. In November 2016, Oxford University researchers reported that a system based on recurrent neural networks (and convolutional neural networks) had achieved 95 percent accuracy in reading lips, outperforming experienced human lip readers, who tested at 52 percent accuracy.
• Convolutional neural networks (CNNs): Artificial neural networks in which the connections between neural layers are inspired by the organization of the animal visual cortex, the portion of the brain that processes images, well suited for perceptual tasks.

For our use cases, we also considered two other techniques—generative adversarial networks (GANs) and reinforcement learning—but did not include them in our potential value assessment of AI, since they remain nascent techniques that are not yet widely applied.

Generative adversarial networks (GANs) use two neural networks contesting one other in a zero-sum game framework (thus “adversarial”). GANs can learn to mimic various distributions of data (for example text, speech, and images) and are therefore valuable in generating test datasets when these are not readily available.

Reinforcement learning is a subfield of machine learning in which systems are trained by receiving virtual “rewards” or “punishments”, essentially learning by trial and error. Google DeepMind has used reinforcement learning to develop systems that can play games, including video games and board games such as Go, better than human champions.

In a business setting, these analytic techniques can be applied to solve real-life problems. The most prevalent problem types are classification, continuous estimation and clustering. A list of problem types and their definitions is available in the sidebar.

Section 2

Insights from use cases

We collated and analyzed more than 400 use cases across 19 industries and nine business functions. They provided insight into the areas within specific sectors where deep neural networks can potentially create the most value, the incremental lift that these neural networks can generate compared with traditional analytics (Exhibit 2), and the voracious data requirements—in terms of volume, variety, and velocity—that must be met for this potential to be realized. Our library of use cases, while extensive, is not exhaustive, and may overstate or understate the potential for certain sectors. We will continue refining and adding to it.

2

Examples of where AI can be used to improve the performance of existing use cases include:

• Predictive maintenance: the power of machine learning to detect anomalies. Deep learning’s capacity to analyze very large amounts of high dimensional data can take existing preventive maintenance systems to a new level. Layering in additional data, such as audio and image data, from other sensors—including relatively cheap ones such as microphones and cameras—neural networks can enhance and possibly replace more traditional methods. AI’s ability to predict failures and allow planned interventions can be used to reduce downtime and operating costs while improving production yield. For example, AI can extend the life of a cargo plane beyond what is possible using traditional analytic techniques by combining plane model data, maintenance history, IoT sensor data such as anomaly detection on engine vibration data, and images and video of engine condition.
• AI-driven logistics optimization can reduce costs through real-time forecasts and behavioral coaching. Application of AI techniques such as continuous estimation to logistics can add substantial value across sectors. AI can optimize routing of delivery traffic, thereby improving fuel efficiency and reducing delivery times. One European trucking company has reduced fuel costs by 15 percent, for example, by using sensors that monitor both vehicle performance and driver behavior; drivers receive real-time coaching, including when to speed up or slow down, optimizing fuel consumption and reducing maintenance costs.
• AI can be a valuable tool for customer service management and personalization challenges. Improved speech recognition in call center management and call routing as a result of the application of AI techniques allow a more seamless experience for customers—and more efficient processing. The capabilities go beyond words alone. For example, deep learning analysis of audio allows systems to assess a customers’ emotional tone; in the event a customer is responding badly to the system, the call can be rerouted automatically to human operators and managers. In other areas of marketing and sales, AI techniques can also have a significant impact. Combining customer demographic and past transaction data with social media monitoring can help generate individualized product recommendations. “Next product to buy” recommendations that target individual customers—as companies such as Amazon and Netflix have successfully been doing--can lead to a twofold increase in the rate of sales conversions.

Two-thirds of the opportunities to use AI are in improving the performance of existing analytics use cases

In 69 percent of the use cases we studied, deep neural networks can be used to improve performance beyond that provided by other analytic techniques. Cases in which only neural networks can be used, which we refer to here as “greenfield” cases, constituted just 16 percent of the total. For the remaining 15 percent, artificial neural networks provided limited additional performance over other analytics techniques, among other reasons because of data limitations that made these cases unsuitable for deep learning (Exhibit 3).

3

Greenfield AI solutions are prevalent in business areas such as customer service management, as well as among some industries where the data are rich and voluminous and at times integrate human reactions. Among industries, we found many greenfield use cases in healthcare, in particular. Some of these cases involve disease diagnosis and improved care, and rely on rich data sets incorporating image and video inputs, including from MRIs.

On average, our use cases suggest that modern deep learning AI techniques have the potential to provide a boost in additional value above and beyond traditional analytics techniques ranging from 30 percent to 128 percent, depending on industry.

Visualizing the potential impact of AI and advanced analytics

Our interactive data visualization shows the potential value created by artificial intelligence and advanced analytics techniques for 19 industries and nine business functions.

View the interactive

In many of our use cases, however, traditional analytics and machine learning techniques continue to underpin a large percentage of the value creation potential in industries including insurance, pharmaceuticals and medical products, and telecommunications, with the potential of AI limited in certain contexts. In part this is due to the way data are used by these industries and to regulatory issues.

Data requirements for deep learning are substantially greater than for other analytics

Making effective use of neural networks in most applications requires large labeled training data sets alongside access to sufficient computing infrastructure. Furthermore, these deep learning techniques are particularly powerful in extracting patterns from complex, multidimensional data types such as images, video, and audio or speech.

Deep-learning methods require thousands of data records for models to become relatively good at classification tasks and, in some cases, millions for them to perform at the level of humans. By one estimate, a supervised deep-learning algorithm will generally achieve acceptable performance with around 5,000 labeled examples per category and will match or exceed human level performance when trained with a data set containing at least 10 million labeled examples. In some cases where advanced analytics is currently used, so much data are available—million or even billions of rows per data set—that AI usage is the most appropriate technique. However, if a threshold of data volume is not reached, AI may not add value to traditional analytics techniques.

These massive data sets can be difficult to obtain or create for many business use cases, and labeling remains a challenge. Most current AI models are trained through “supervised learning”, which requires humans to label and categorize the underlying data. However promising new techniques are emerging to overcome these data bottlenecks, such as reinforcement learning, generative adversarial networks, transfer learning, and “one-shot learning,” which allows a trained AI model to learn about a subject based on a small number of real-world demonstrations or examples—and sometimes just one.

Organizations will have to adopt and implement strategies that enable them to collect and integrate data at scale. Even with large datasets, they will have to guard against “overfitting,” where a model too tightly matches the “noisy” or random features of the training set, resulting in a corresponding lack of accuracy in future performance, and against “underfitting,” where the model fails to capture all of the relevant features. Linking data across customer segments and channels, rather than allowing the data to languish in silos, is especially important to create value.

Realizing AI’s full potential requires a diverse range of data types including images, video, and audio

Neural AI techniques excel at analyzing image, video, and audio data types because of their complex, multidimensional nature, known by practitioners as “high dimensionality.” Neural networks are good at dealing with high dimensionality, as multiple layers in a network can learn to represent the many different features present in the data. Thus, for facial recognition, the first layer in the network could focus on raw pixels, the next on edges and lines, another on generic facial features, and the final layer might identify the face. Unlike previous generations of AI, which often required human expertise to do “feature engineering,” these neural network techniques are often able to learn to represent these features in their simulated neural networks as part of the training process.

Along with issues around the volume and variety of data, velocity is also a requirement: AI techniques require models to be retrained to match potential changing conditions, so the training data must be refreshed frequently. In one-third of the cases, the model needs to be refreshed at least monthly, and almost one in four cases requires a daily refresh; this is especially the case in marketing and sales and in supply chain management and manufacturing.

Section 3

Sizing the potential value of AI

We estimate that the AI techniques we cite in this briefing together have the potential to create between $3.5 trillion and $5.8 trillion in value annually across nine business functions in 19 industries. This constitutes about 40 percent of the overall $9.5 trillion to $15.4 trillion annual impact that could potentially be enabled by all analytical techniques (Exhibit 4).

4

Per industry, we estimate that AI’s potential value amounts to between one and nine percent of 2016 revenue. The value as measured by percentage of industry revenue varies significantly among industries, depending on the specific applicable use cases, the availability of abundant and complex data, as well as on regulatory and other constraints.

These figures are not forecasts for a particular period, but they are indicative of the considerable potential for the global economy that advanced analytics represents.

From the use cases we have examined, we find that the greatest potential value impact from using AI are both in top-line-oriented functions, such as in marketing and sales, and bottom-line-oriented operational functions, including supply chain management and manufacturing.

Consumer industries such as retail and high tech will tend to see more potential from marketing and sales AI applications because frequent and digital interactions between business and customers generate larger data sets for AI techniques to tap into. E-commerce platforms, in particular, stand to benefit. This is because of the ease with which these platforms collect customer information such as click data or time spent on a web page and can then customize promotions, prices, and products for each customer dynamically and in real time.

5

Here is a snapshot of three sectors where we have seen AI’s impact: (Exhibit 5)

• In retail, marketing and sales is the area with the most significant potential value from AI, and within that function, pricing and promotion and customer service management are the main value areas. Our use cases show that using customer data to personalize promotions, for example, including tailoring individual offers every day, can lead to a one to two percent increase in incremental sales for brick-and-mortar retailers alone.
• In consumer goods, supply-chain management is the key function that could benefit from AI deployment. Among the examples in our use cases, we see how forecasting based on underlying causal drivers of demand rather than prior outcomes can improve forecasting accuracy by 10 to 20 percent, which translates into a potential five percent reduction in inventory costs and revenue increases of two to three percent.
• In banking, particularly retail banking, AI has significant value potential in marketing and sales, much as it does in retail. However, because of the importance of assessing and managing risk in banking, for example for loan underwriting and fraud detection, AI has much higher value potential to improve performance in risk in the banking sector than in many other industries.

Section 4

The road to impact and value

Artificial intelligence is attracting growing amounts of corporate investment, and as the technologies develop, the potential value that can be unlocked is likely to grow. So far, however, only about 20 percent of AI-aware companies are currently using one or more of its technologies in a core business process or at scale.

For all their promise, AI technologies have plenty of limitations that will need to be overcome. They include the onerous data requirements listed above, but also five other limitations:

• First is the challenge of labeling training data, which often must be done manually and is necessary for supervised learning. Promising new techniques are emerging to address this challenge, such as reinforcement learning and in-stream supervision, in which data can be labeled in the course of natural usage.
• Second is the difficulty of obtaining data sets that are sufficiently large and comprehensive to be used for training; for many business use cases, creating or obtaining such massive data sets can be difficult—for example, limited clinical-trial data to predict healthcare treatment outcomes more accurately.
• Third is the difficulty of explaining in human terms results from large and complex models: why was a certain decision reached? Product certifications in healthcare and in the automotive and aerospace industries, for example, can be an obstacle; among other constraints, regulators often want rules and choice criteria to be clearly explainable.
• Fourth is the generalizability of learning: AI models continue to have difficulties in carrying their experiences from one set of circumstances to another. That means companies must commit resources to train new models even for use cases that are similar to previous ones. Transfer learning—in which an AI model is trained to accomplish a certain task and then quickly applies that learning to a similar but distinct activity—is one promising response to this challenge.
• The fifth limitation concerns the risk of bias in data and algorithms. This issue touches on concerns that are more social in nature and which could require broader steps to resolve, such as understanding how the processes used to collect training data can influence the behavior of models they are used to train. For example, unintended biases can be introduced when training data is not representative of the larger population to which an AI model is applied. Thus, facial recognition models trained on a population of faces corresponding to the demographics of AI developers could struggle when applied to populations with more diverse characteristics. A recent report on the malicious use of AI highlights a range of security threats, from sophisticated automation of hacking to hyper-personalized political disinformation campaigns.

Organizational challenges around technology, processes, and people can slow or impede AI adoption

Organizations planning to adopt significant deep learning efforts will need to consider a spectrum of options about how to do so. The range of options includes building a complete in-house AI capability, outsourcing these capabilities, or leveraging AI-as-a-service offerings.

Based on the use cases they plan to build, companies will need to create a data plan that produces results and predictions, which can be fed either into designed interfaces for humans to act on or into transaction systems. Key data engineering challenges include data creation or acquisition, defining data ontology, and building appropriate data “pipes.” Given the significant computational requirements of deep learning, some organizations will maintain their own data centers, because of regulations or security concerns, but the capital expenditures could be considerable, particularly when using specialized hardware. Cloud vendors offer another option.

Process can also become an impediment to successful adoption unless organizations are digitally mature. On the technical side, organizations will have to develop robust data maintenance and governance processes, and implement modern software disciplines such as Agile and DevOps. Even more challenging, in terms of scale, is overcoming the “last mile” problem of making sure the superior insights provided by AI are instantiated in the behavior of the people and processes of an enterprise.

On the people front, much of the construction and optimization of deep neural networks remains something of an art requiring real experts to deliver step-change performance increases. Demand for these skills far outstrips supply at present; according to some estimates, fewer than 10,000 people have the skills necessary to tackle serious AI problems. and competition for them is fierce among the tech giants.

AI can seem an elusive business case

Where AI techniques and data are available and the value is clearly proven, organizations can already pursue the opportunity. In some areas, the techniques today may be mature and the data available, but the cost and complexity of deploying AI may simply not be worthwhile, given the value that could be generated. For example, an airline could use facial recognition and other biometric scanning technology to streamline aircraft boarding, but the value of doing so may not justify the cost and issues around privacy and personal identification.

Similarly, we can see potential cases where the data and the techniques are maturing, but the value is not yet clear. The most unpredictable scenario is where either the data (both the types and volume) or the techniques are simply too new and untested to know how much value they could unlock. For example, in healthcare, if AI were able to build on the superhuman precision we are already starting to see with X-ray analysis and broaden that to more accurate diagnoses and even automated medical procedures, the economic value could be very significant. At the same time, the complexities and costs of arriving at this frontier are also daunting. Among other issues, it would require flawless technical execution and resolving issues of malpractice insurance and other legal concerns.

Societal concerns and regulations can also constrain AI use. Regulatory constraints are especially prevalent in use cases related to personally identifiable information. This is particularly relevant at a time of growing public debate about the use and commercialization of individual data on some online platforms. Use and storage of personal information is especially sensitive in sectors such as banking, health care, and pharmaceutical and medical products, as well as in the public and social sector. In addition to addressing these issues, businesses and other users of data for AI will need to continue to evolve business models related to data use in order to address societies’ concerns.. Furthermore, regulatory requirements and restrictions can differ from country to country, as well from sector to sector.

Implications for stakeholders

As we have seen, it is a company’s ability to execute against AI models that creates value, rather than the models themselves. In this final section, we sketch out some of the high-level implications of our study of AI use cases for providers of AI technology, appliers of AI technology, and policy makers, who set the context for both.

• For AI technology provider companies: Many companies that develop or provide AI to others have considerable strength in the technology itself and the data scientists needed to make it work, but they can lack a deep understanding of end markets. Understanding the value potential of AI across sectors and functions can help shape the portfolios of these AI technology companies. That said, they shouldn’t necessarily only prioritize the areas of highest potential value. Instead, they can combine that data with complementary analyses of the competitor landscape, of their own existing strengths, sector or function knowledge, and customer relationships, to shape their investment portfolios. On the technical side, the mapping of problem types and techniques to sectors and functions of potential value can guide a company with specific areas of expertise on where to focus.
• Many companies seeking to adopt AI in their operations have started machine learning and AI experiments across their business. Before launching more pilots or testing solutions, it is useful to step back and take a holistic approach to the issue, moving to create a prioritized portfolio of initiatives across the enterprise, including AI and the wider analytic and digital techniques available. For a business leader to create an appropriate portfolio, it is important to develop an understanding about which use cases and domains have the potential to drive the most value for a company, as well as which AI and other analytical techniques will need to be deployed to capture that value. This portfolio ought to be informed not only by where the theoretical value can be captured, but by the question of how the techniques can be deployed at scale across the enterprise. The question of how analytical techniques are scaling is driven less by the techniques themselves and more by a company’s skills, capabilities, and data. Companies will need to consider efforts on the “first mile,” that is, how to acquire and organize data and efforts, as well as on the “last mile,” or how to integrate the output of AI models into work flows ranging from clinical trial managers and sales force managers to procurement officers. Previous MGI research suggests that AI leaders invest heavily in these first- and last-mile efforts.
• Policy makers will need to strike a balance between supporting the development of AI technologies and managing any risks from bad actors. They have an interest in supporting broad adoption, since AI can lead to higher labor productivity, economic growth, and societal prosperity. Their tools include public investments in research and development as well as support for a variety of training programs, which can help nurture AI talent. On the issue of data, governments can spur the development of training data directly through open data initiatives. Opening up public-sector data can spur private-sector innovation. Setting common data standards can also help. AI is also raising new questions for policy makers to grapple with for which historical tools and frameworks may not be adequate. Therefore, some policy innovations will likely be needed to cope with these rapidly evolving technologies. But given the scale of the beneficial impact on business the economy and society, the goal should not be to constrain the adoption and application of AI, but rather to encourage its beneficial and safe use.