15

Today’s AI tools are far from the machines with human-like intelligence of science fiction (often referred to as “artificial general intelligence” or AGI, or “strong AI”). The current generation of AI provides tools for prediction and little else.

This view of AI does not diminish it. As Steve Jobs once remarked, “One of the things that really separates us from the high primates is that we’re tool builders.” He used the example of the bicycle as a tool that had given people superpowers in locomotion above every other animal. And he felt the same about computers: “What a computer is to me is it’s the most remarkable tool that we’ve ever come up with, and it’s the equivalent of a bicycle for our minds.”1

Today, AI tools predict the intention of speech (Amazon’s Echo), predict command context (Apple’s Siri), predict what you want to buy (Amazon’s recommendations), predict which links will connect you to the information you want to find (Google search), predict when to apply the brakes to avoid danger (Tesla’s Autopilot), and predict the news you will want to read (Facebook’s newsfeed). None of these AI tools are performing an entire workflow. Instead, each delivers a predictive component to make it easier for someone to make a decision. AI empowers.

But how should you decide whether you should use an AI tool for a particular task in your business? Every task has a group of decisions at its heart, and those decisions have some predictive element.

We provide a way of evaluating AI within the context of a task. Just as we suggested identifying tasks by breaking down a workflow to find out whether AI might have a role, we now suggest taking each of those tasks and decomposing them into their constituent elements.

The AI Canvas

The CDL exposed us to many startups taking advantage of recent machine-learning technologies to build new AI tools. Each company in the lab is predicated on building a specific tool, some for consumer experiences, but most for enterprise customers. The latter type focus on identifying task opportunities within enterprise workflows to focus and position their offering. They deconstruct workflows, identify a task with a prediction element, and build their business based on the provision of a tool for delivering that prediction.

In advising them, we found it useful to separate the parts of a decision into each of its elements (refer to figure 8-1): prediction, input, judgment, training, action, outcome, and feedback. In the process, we developed an “AI canvas” to help decompose tasks in order to understand the potential role of a prediction machine (see figure 15-1). The canvas is an aid for contemplating, building, and assessing AI tools. It provides discipline in identifying each component of a task’s decision. It forces clarity in describing each component.

To see how this works, let’s consider the biotech company Atomwise, which offers a prediction tool that aims to shorten the time involved in discovering promising pharmaceutical drug prospects. Millions of possible drug molecules might become drugs, but purchasing and testing each drug is time consuming and costly. How do drug companies determine which to test? They make educated guesses, or predictions, based on research that suggests which molecules are most likely to become effective drugs.

Atomwise CEO Abraham Heifets, giving us a quick explanation of the science, said, “For a drug to work, it has to bind the disease target, and it has to fail to bind proteins in your liver, your kidneys, your heart, your brain, and other things that are going to cause toxic side effects. It comes down to ‘stick to the things you want to stick to, fail to stick to the things you don’t.’”

So, if drug companies can predict binding affinity, then they can identify which molecules are most likely to work. Atomwise provides this prediction by offering an AI tool that makes the task of identifying potential drugs more efficient. The tool uses AI to predict the binding affinity of molecules, so Atomwise can recommend to drug companies, in a ranked list, which molecules have the best binding affinity for a disease protein. For example, Atomwise might provide the top twenty molecules that have the highest binding affinity for, say, the Ebola virus. Rather than just testing molecules one at a time, Atomwise’s prediction machine can handle millions of possibilities. While the drug company still needs to test and verify candidates through a combination of human and machine judgments and actions, the Atomwise AI tool dramatically lowers the cost and accelerates the speed of the first task of finding those candidates.

Where does judgment come in? In recognizing the aggregate value of a particular candidate molecule to the pharmaceutical industry. This value takes two forms: targeting the disease and understanding potential side effects. In selecting the molecules to test, the company needs to determine the payoffs of targeting the disease and costs of the side effects. As Heifets noted, “You are more tolerant of side effects for chemotherapy than for an acne cream.”

The Atomwise prediction machine learns from data on binding affinity. As of July 2017, it had 38 million public data points on binding affinity plus many more that it either purchased or learned itself. Each data point consists of molecule and protein characteristics as well as a measure of the binding between the molecules and the proteins. As Atomwise makes more recommendations, it may get further feedback from customers, so the prediction machine will continue to improve.

Using this machine, given data on protein characteristics, Atomwise can predict which molecules have the highest binding affinity. It can also take the data on protein characteristics and predict whether molecules that have never been produced are likely to have high binding affinity.

The way to decompose the Atomwise molecule selection task is to fill in the canvas (see figure 15-2). This means identifying the following:

• ACTION: What are you trying to do? For Atomwise, it is to test molecules to help cure or prevent disease.
• PREDICTION: What do you need to know to make the decision? Atomwise predicts binding affinities of potential molecules and proteins.
• JUDGMENT: How do you value different outcomes and errors? Atomwise and its customers set the criterion regarding the relative importance of targeting the disease and the relative costs of potential side effects.
• OUTCOME: What are your metrics for task success? For Atomwise, it’s the results of the test. Ultimately, did the test lead to a new drug?
• INPUT: What data do you need to run the predictive algorithm? Atomwise uses data on the characteristics of the disease proteins to predict.
• TRAINING: What data do you need to train the predictive algorithm? Atomwise employs data on the binding affinity of molecules and proteins, along with molecule and protein characteristics.
• FEEDBACK: How can you use the outcomes to improve the algorithm? Atomwise uses test outcomes, regardless of their success, to improve future predictions.

Atomwise’s value proposition lies in delivering an AI tool that supports a prediction task in its customers’ drug discovery workflow. It removes the prediction task from human hands. To provide that value, it amassed a unique data set to predict binding affinity. The prediction’s value is in reducing the cost and increasing the likelihood of success for drug development. Atomwise’s clients use the prediction in combination with their own expert judgment of the payoffs to molecules with different binding affinities to different kinds of proteins.

An AI Canvas for MBA Recruiting

The canvas is also useful in large organizations. To apply it, we break down the workflow into tasks. Here, we consider an AI canvas centered on the decision of which MBA applicants to accept into a program. Figure 15-3 provides a possible canvas.

Where did the canvas come from? First, recruiting requires a prediction: Who will be a best or high-value student? That seems straightforward. We simply need to define “best.” The school’s strategy can help identify this. However, many organizations have vague, multifaceted mission statements that lend themselves well to marketing brochures but not so well to identifying the prediction objective for an AI.

Business schools have many strategies that implicitly or explicitly define what they mean by “best.” They may be simple indicators such as maximizing standardized test scores like the GMAT or broader goals such as recruiting students who will boost the school’s rankings in the Financial Times or US News & World Report. They may also want students who have a mix of quantitative and qualitative skills. Or they may want international students. Or they may want diversity. No school can pursue all these goals simultaneously and must exercise some choice. Otherwise, it will compromise on all dimensions and excel at none.

In figure 15-3, we imagine that our school’s strategy is to have the greatest impact on business globally. This subjective notion is strategic in that it is global rather than local and is looking for impact rather than, say, maximizing student income or creating wealth.

For the AI to predict global business impact, we need to measure it. Here, we assume the role of the reward function engineer. What training data do we have that might be a proxy for global business impact? One option might be to identify the best alumni from each class—the fifty alumni from each year who have had the biggest impact. Choosing those alumni is, of course, subjective, but not impossible.

While we may set global business impact as the goal for a prediction machine, the value of accepting a particular student is a matter of judgment. How costly is it to accept a weak student who we wrongly predicted would be among the elite alumni? How costly is it to reject a strong student who we wrongly predicted would be weak? The assessment of that trade-off is “judgment,” an explicit element in the AI canvas.

Once we specify the objective of the prediction, identifying the input data needed is straightforward. We need application information for incoming students in order to predict how they will do. We might also use social media. Over time, we will observe more students’ career outcomes and can use that feedback to improve predictions. The predictions will tell us which applicants to accept, but only after determining our objective and judging the cost of making a mistake.

KEY POINTS

• Tasks need to be decomposed in order to see where prediction machines can be inserted. This allows you to estimate the benefit of the enhanced prediction and the cost of generating that prediction. Once you have generated reasonable estimates, rank-order the AIs from highest to lowest ROI by starting at the top and working your way down, implementing AI tools as long as the expected ROI makes sense.
• The AI canvas is an aid to help with the decomposition process. Fill out the AI canvas for every decision or task. This introduces discipline and structure into the process. It forces you to be clear about all three data types required: training, input, and feedback. It also forces you to articulate precisely what you need to predict, the judgment required to assess the relative value of different actions and outcomes, the action possibilities, and the outcome possibilities.
• At the center of the AI canvas is prediction. You need to identify the core prediction at the heart of the task, and this can require AI insight. The effort to answer this question often initiates an existential discussion among the leadership team: “What is our real objective, anyhow?” Prediction requires a specificity not often found in mission statements. For a business school, for example, it is easy to say that it is focused on recruiting the “best” students, but in order to specify the prediction, we need to specify what “best” means—highest salary offer upon graduation? Most likely to assume a CEO role within five years? Most diverse? Most likely to donate to the school after graduation? Even seemingly straightforward objectives, like profit maximization, are not as simple as they first appear. Should we predict the action to take that will maximize profit this week, this quarter, this year, or this decade? Companies often find themselves having to go back to basics to realign on their objectives and sharpen their mission statement as a first step in their work on their AI strategy.