Dispelling the myths surrounding machine learning’s ability to automate and augment business processes reveals a wealth of potential on the not-too-distant horizon.
Artificial intelligence (AI) and machine learning are making rapid progress with consumers, gaining a nearly ubiquitous presence in our daily lives. In the past three years alone, Amazon has sold 10 million Echo smart speakers, and Apple’s new iPhone X has an AI chip dedicated to handling tasks such as facial recognition and understanding voice commands.
And yet, AI is not nearly as pervasive for automating and enhancing business processes. Although an Oxford study says AI could automate more than 50 percent of all jobs over the next two decades, many executives—perhaps with perspectives clouded by media hype and misinformation—doubt the potential impact of this burgeoning technology.
In this paper, we provide a practical fact-based view of the AI trends and implications for enterprise automation over the next five to seven years. Based on our work with a wide range of companies and discussions with academia, start-ups, and enterprise users, we highlight four important AI facts, discuss the implications, and offer strategies on how to get started.
Fact 1: The AI boom is sustainable and should not be ignored
For the first time, machine-learning algorithms are beating humans in tasks such as image recognition and voice-to-text translation, and complex games such as Go. This AI boom is fueled by a convergence of three factors: a breakthrough in deep-learning algorithms, the proliferation of big data (structured data) to train these algorithms, and an exponential speedup in processing power for machine-learning hardware, such as the graphics processing unit (GPU) chipsets that cut down a machine’s training time from months to days and hours (see figure 1).
All three factors are expected to keep accelerating. In fact, by 2020, almost 70 percent of all enterprise data will be stored and processed in cloud-based data centers—providing unprecedented big data infrastructure to train machine-learning algorithms. Similarly, hardware chips that can speed up algorithm training are rapidly advancing. Google, NVIDIA, and Intel have announced next-generation GPU chip hardware that will accelerate training by 10 to 100 times. And based on the growing number of patent applications, advances in the underlying machine- learning algorithms will continue to pick up pace.
These trends offer strong evidence that AI will continue to improve. For business leaders, this means that ignoring AI is no longer an option.
Fact 2: AI is being used across organizations but with a limited scope
So what will AI be able to do for enterprise automation in the next five to seven years? Most experts say companies will adopt narrow AI, or supervised machine learning that is focused on one task. AI algorithms will be able to use training data to learn how to automate a task, but once the task is mastered, the solution will be narrow, and in most cases, the machine will not be able to generalize that learning to perform other tasks (see figure 2). Widespread use of broad, human-like general intelligence—in other words, unsupervised and context aware—could be decades away.
For example, consider a company that wants an AI algorithm that can scan PDFs and hand- written invoices, validate field formatting, and then automatically trigger the accounts payable process. With more advanced algorithms, or unsupervised learning, the solution could also work with unlabeled data, such as collecting past invoices without having fields that are labeled as valid or invalid. However, the AI solution this company adopted is limited to automating the task of text-field recognition and formatting. To use AI to do more advanced invoicing tasks, such as detecting fraud, the company would need to design and train a completely new solution by focusing on other underlying features and patterns.
In the foreseeable future, AI applications will fit a narrow paradigm of using training data for supervised machine learning. This has two strategic implications: acquiring labeled data for training becomes a strategic capability and a source of differentiation, and AI solutions require deep functional and domain-specific human co-creation and process redesign.
Fact 3: AI is ready for deployment on select activities
The Japanese insurer Fukoko plans to use AI to replace more than two dozen human agents who process claims, and Goldman Sachs used machine learning to transform its 600-person trader unit into a much leaner 200-person team between 2000 and 2016. However, not all organizational activities are suitable for AI automation under today’s narrow paradigm.
One way to characterize machine-learnable tasks is what computer scientist Andrew Ng calls A→B activities: processes that take a set of unambiguous inputs and produce a response (see figure 3). For example, retail demand forecasting is an A→B activity. By taking an array of information such as seasonality, regional sales history, point of sale trends, social media signals, and data about pricing sensitivity, an algorithm can predict demand as well as a human can. Financial trading is another good example. An algorithm can use historic prices, macro trend drivers, and arbitrage rules compiled from past traders and determine whether the best option is to buy or sell. Although making the right decision based only on these inputs may be a challenge given the unpredictability of the market, an AI solution is still attractive if it can outperform humans over a large number of trades.
Mapping enterprise processes and activities into A→B versus non-A→B categories can help managers pinpoint opportunities to use AI automation and augmentation strategies.
Fact 4: Adopting AI is about more than technical feasibility
Some AI applications will be adopted faster than others, even though the technical requirements are comparable. Broader solutions can ensure that a company’s portfolio of AI initiatives can unlock value in the near term while also paving the way for long-term aspirations.
Consider the following when deciding where to use AI for enterprise automation:
- One-time costs. Assess the initial capital outlay for a new AI solution, such as developing an algorithm and acquiring training data. Open-source access to algorithms and pay-as-you-go “AI as a service” platforms can lower the fixed-cost hurdles, but access to training data can be either an expensive bottleneck or a powerful source of differentiation.
- Switching costs. Evaluate the costs associated with displacing the existing solution with an AI solution, including technical hurdles such as the ability to open the AI algorithm’s black box to trace and explain decisions and human obstacles such as political and cultural resistance to change.
- Ecosystem requirements. Determine if an integrated solution will require any complementary technologies. For example, an AI solution that must be integrated with innovative IoT sensors and emerging robotics technology will be more complex to adopt.
- System externality hurdles. Consider the extent to which the AI solution could negatively affect third parties that did not choose to use the new technology, bearing in mind that the value of the solution will increase as more users adopt it.
Companies face a variety of challenges when adopting AI solutions, along with different potential timelines for achieving a broad impact. For example, a call center we worked with was assessing one solution that would convert handwritten forms into structured database records (see figure 4). The software development and switching costs were modest because the solution could be launched by using off-the-shelf machine-learning modules. In addition, the company already had a large amount of labeled data to use for training, optical character recognition scanners to process the information, and extensive databases. The solution faced relatively low externality barriers, and proof-of-concept solutions in certain functions and geographic units could serve as training data and points of reference for other units. The company expects to have broad use of the solution by 2021.
Another solution would enhance the call center’s agent interactions by automating sentiment analysis from consumers’ voice calls or chats. This solution has much higher switching costs in terms of cultural and risk barriers. The company opted to start small to minimize any negative impact on its customers. If the solution works, the company will need to redesign its end-to-end training process to allow the AI engine to push suggested responses to agents. This solution has relatively high network externalities, where higher adoption will generate more training data to improve performance but collecting the initial critical mass of training data will take time and require company leaders to take a leap of faith. Given such complexities, an AI sentiment analysis bot will likely take seven to 10 years rather than the first solution’s two- to three-year adoption.
Charting the way with an AI road map
AI and machine learning will continue to enjoy a boom phase, powered by more sophisticated algorithms, greater availability of data, and more advanced hardware. However, forward- thinking companies take a disciplined approach in evaluating the opportunities. Based on the above trends and their implications, four steps can help organizations embark on a comprehensive AI automation strategy:
Map the A→B activities. Rather than using one-off AI solutions, systematically identify and map all A→B activities across the value chain for potential machine-learning automation. This ensures a broad view for downstream considerations such as value potential, time to impact, development costs, and alignment with the overall business strategy.
Prioritize where AI will be used. Because AI solutions are not generalizable and require specific data resources and training, prioritize the A→B activities to be automated based on the potential business impact along with the complexity of adoption to ensure smart resource allocation.
Adopt a portfolio approach. Manage the AI automation pipeline like a product innovation portfolio to optimize both near-term and long-term value creation while mitigating the risks. Segment the initiatives based on available data and adoption risks.
Consider the full spectrum of options to make, buy, or partner Using AI often requires multidisciplinary capabilities and insights. Capitalize on internal experts, consultants, and corporate partners as well as AI technology start-ups to adopt best-of-breed solutions and develop nimble proofs of concepts that can be launched in weeks instead of years. For sectors where AI has disrupting potential, have a clear overarching acquisition strategy to avoid expensive rushed bets or late entries in a landscape of record-high valuations.
AI automation is rapidly becoming a reality across organizations and value chains. Now is the time for forward-thinking business leaders to adopt a disciplined, portfolio-based approach to develop machine-learning capabilities, data, and partnerships to remain relevant.
The author would like to thank Eric Gervet, Florian Dickgreber, Denis Bassler, and Joseph Edwar for their valuable contributions to this paper.