In recent months, Artificial Intelligence (AI) has leapt off the pages of Science Fiction and into the headlines. The Economist focused an entire Technology Quarterly on the dramatic advances made possible through Machine Learning. Recent articles by Tom Davenport in Harvard Business Review and Deloitte have pushed moving beyond the “artisanal,” human-driven analytics of the past toward a bountiful, automated future. With talented Data Scientists scarce but vital, the value proposition for AI seems to be clear. So, should companies hire Data Scientists (especially consultants) if the computer can do all the work?
A Moment for the Machines
The rationale for AI adoption given by Davenport is straightforward:
- Companies are collecting data and need to analyze it to create value.
- The talented Data Scientists who can do the modeling and analytics are scarce and expensive to engage.
- Increasingly voluminous and diverse datasets require sophisticated modeling techniques which create models too complicated for humans to construct or explain.
Businesses understand unlocking value in their data through analytics, but according to McKinsey “a great deal of that value is still on the table” (see Figure1).
Now (according to Tom Davenport and others) it is possible to perform analytics through automation. Computers can take in data, generate models, and produce results, all with minimal human intervention
In many respects, this is Christensen’s disruption theory at work in the Data Science field. The “good enough” quality of mass-produced, automated models is, in the market place, beating out the higher-quality but more costly “best” of the handcrafted, established experts. If it happened to the cobblers, the haberdashers, and the tailors, why not the Data Scientists?
A Case for Craftsmanship
To make the case for tailored analytics solutions we turn to an analogy from the clothing industry. For clothing manufacturers and consumers, off-the-rack production makes sense. It’s readily available, efficient, and tends to be less expensive than the tailor-made equivalent. The downside to this convenience is a loss of any personal relationship with the manufacturer, resulting in an acceptable fit that is rarely prefect.
The value of tailor-made craftsmanship in an off-the-rack marketplace is the flexibility of choice—typically of top-quality materials, personalized styling, and fit. Working with a tailor builds a relationship. Over time, they learn your likes and dislikes, and can recommend future garment options. The real benefit of tailor-made clothes is that they are made-to-measure; the fit and style is unique, individual, and unmatched, but comes at a higher price. All too often we encounter clients who have tried off-the-shelf analytic solutions or purchased analytic software only to discover that it didn’t meet their needs. Getting value from data requires that the tools and analytic solutions fit the unique needs of the business. At the limit of the hardest analysis cases, only a custom solution can provide a useful answer at all, or can provide one an order of magnitude more valuable than an automated answer.
A recent episode in healthcare involving IBM’s celebrated Watson system highlighted some of the difficulties with applying AI for analytics. The MD Anderson Cancer Center chose Watson after a carefully-considered search. They expected Watson to produce value by integrating their many disparate data sources and improving patient outcomes. What happened?
An audit revealed that Watson did not work with MD Anderson’s new Electronic Medical Record (EMR) system and that Watson was never used in a patient treatment clinical environment. In healthcare, there are infrastructure and data integration hurdles resulting from the need to protect patient data and satisfy government regulations. Data access must be closely maintained, and any analytic models must be interpretable to satisfy compliance. The implementation of Watson at MD Anderson did not account for these realities.
Watson has gained a lot of favorable press for some spectacular outcomes in healthcare, but these were extreme cases. For all its strengths, Watson is not a simple “turnkey” solution. Like our off-the-shelf clothing example, Watson’s fit is often not quite right. This high-quality AI tool still needs a human touch to generate real value. Whether a human or computer builds an analytical model, it will rely on the domain knowledge of experts to generate value. While tailors can make bespoke garments, they can also make alterations to off-the-shelf clothing.
Advising non-experts to trust models that they do not fully understand is flawed logic, and unlikely to generate buy-in. Trust is built when statistical models are adopted by decision makers and the action taken on the model results generates business value. If an off-the-shelf solution works, a tailored solution may not be necessary. Craftsman, like tailors and Data Scientists, build trust through quality workmanship, knowing their clientele, and building a relationship. Like a good jacket, if the model fits, the customer will use it.
Finding the Right Fit
Just as with clothes, “tailored” analytics and automated solutions each have their place. It is a false dichotomy to place human Data Scientists, or consultants, in opposition to AI. As Data Scientists, we want this shift to happen! AI makes repeatable aspects of our job easier, enabling us to focus on more interesting aspects, for which there is no good automated solution. As with other fields disrupted by automation, our value as Data Scientists will shift, and our roles will change.
Framed another way, automation is either a displacing or a disrupting technology. For example, factory robots displace workers from the floor, given their greater skill at performing repetitive manufacturing tasks. AI tools will displace Data Scientists from similarly repetitive and low-skill tasks that computers are better able to accomplish. Humans Data Scientists conceivably will be freed from responsibility in areas like Data Preparation and Modeling. AI solutions are also increasingly strong at predictive tasks, but they are still lacking when it comes to descriptive tasks.
Automation may also be a disruptive technology: it changes the way that skilled-workers perform their jobs. Sewing machines did not make tailors obsolete; they made them more efficient. By analogy, AI makes Data Scientists more efficient; freeing them from time-consuming, menial tasks and enabling them to spend more time on the important, difficult work of description, visualization, hypothesis generation, brainstorming, and communicating results.
The value of Data Scientists in an automated world is their workmanship. Data Scientists use automation in creative ways to extract value from data. AI for Data Science, like the sewing machine for tailoring, is just a tool. At Elder Research, we use automated feature selection and data profiling tools to accelerate our services and deliver value more efficiently to our clients.
Contrast the Watson example above with our recent engagement with the development arm of a major-medical center. Major-gift philanthropy is a deeply relational field. Because of this, automated model building through AI would have made our model no more effective or valuable to this client. If anything, it would have been counter-productive to adoption. The success of our model was driven largely by the relationships that we built with development professionals, through soliciting their feedback regularly throughout the model building process, and carefully describing the significance of the model results to our clients. This collaborative partnership gave them early ownership of the model, and invested them in its incorporation and implementation into their processes.
Current hype for AI is understandable, but its general applicability is questionable. There is scope (and rationale) for automating aspects of the predictive modeling process for which computers are superior to human analysts (e.g., data preparation). However, AI falls short when it comes to describing model results to users, especially in environments that are less analytically sophisticated or heavily-regulated.
Ultimately, the problem must fit the application of AI. While an off-the-shelf solution may provide a close fit, depending on the sensitivity of the application to your core business, a tailor-made approach may be needed to find the right fit. The requirements for successful application of AI are no different than those for analytics generally, and require experienced craftspeople to get the fit just right.