Every Technology Changes the Rules
There is a lot of nervousness around artificial intelligence, and it’s hard to blame people.
Every major technology seems to bring the same questions. Will it eliminate jobs or create new ones? Will it improve quality of life or fundamentally change society? People asked those questions about factories, electricity, automobiles, personal computers, the internet, and social media. Some predictions were right. Many were completely wrong. Almost nobody predicted exactly how those technologies would ultimately be used.
Part of the uncertainty around AI comes from trust. Most people have seen examples of someone using AI to write an article, only to discover later that some of the facts or references were completely made up. ‘Hallucinations’. A bigger concern is not knowing where the technology will lead or what it might mean for their job. Beneath all of that is a simpler question: What is this technology actually for?
To practitioners who build AI solutions, discussions too often become about the technology itself: algorithms, software, licenses, models, and clouds. That is understandable, but it sometimes feels like standing in a factory and talking endlessly about the machinery while paying little attention to what the factory is supposed to produce.
History suggests that the biggest gains from new technologies rarely come from doing the same things the essentially same way. Real gains come from developing new processes and systems that fundamentally change how things get done.
Take electricity as an example. Early factories often replaced steam engines with electric motors while leaving everything else unchanged. The real benefits came later, when people realized they no longer needed a massive engine in the center of the factory driving everything through belts and shafts. Machines could be placed where they made the most sense, workflows reorganized, and entire factory layouts redesigned. The biggest gains came not from the motors themselves, but from rethinking the system around them.
Measuring the Unmeasurable
While Generative AI may be relatively new to most people, decision support and optimization tools are not. In 1947, George Dantzig developed the Simplex algorithm, a method for finding optimal solutions to planning problems and many other applications. As the story is often told, his wife looked at this new mathematical technique and asked a perfectly reasonable, skeptical question: “What in the world could that possibly be used for?” In response, Dantzig formulated what became known as the Diet Problem — finding the least expensive combination of foods that would satisfy a person’s nutritional requirements.
The Diet Problem was a simple example, but the idea turned out to be much bigger than diets. The same mathematics could be used to allocate resources, schedule workers, route vehicles, plan production, and manage supply chains. Today, the Simplex algorithm and its descendants are utilized in making planning and operational decisions in nearly every industry.
Not everything can be measured quantitatively. Consider personnel scheduling. Labor costs, staffing shortages, and compliance with laws and labor contracts are relatively straightforward. Things become less clear when trying to account for quality of life, employee preferences, schedule fairness, work-life balance, or long-term retention.
Mathematically, the solution seems simple. Let’s put an equivalent dollar value to every decision and outcome. Scheduling a worker to back-to-back shifts may incur penalty equivalent to $100. Denying a time-off request incurs a penalty of $40. The scheduling software balances those indirect and artificial costs against labor and other real costs to produce the “best” schedule. Unfortunately, few organizations can agree on what those dollar values should be – not just the artificial costs but often the real costs.
A common workaround is to use penalty weights instead. Back-to-back shifts cost 5 points while denying a time-off request costs 2 points. Rather than claiming each point has a specific dollar value, several alternatives can be generated and compared side-by-side – costs vs. penalties. But that just substitutes one problem for another.
Who decides that one penalty should be 5 points and another should be 2? Why not 10 and 3? Or 100 and 20? Perhaps all we have done is take a planning problem people understand and could solve with a whiteboard and pad of post-its or a spreadsheet and turn it into a discussion about penalty weights and organizational values.
That is not necessarily a bad thing. Before algorithms can determine the best schedule, people must first decide what “best” means. In many cases, that question is the real problem. What does the organization value and how does it measure it? If an organization cannot define or measure what it cares about, how can it perform against those values?
Math is Easy. People are Hard.
Gene Woolsey, late Professor of Operations Research at the Colorado School of Mines, identified the real challenge when he said, “The mathematics is often the easy part; the people are the hard part.”
Many organizations would say they have efficient and equitable processes for generating personnel schedules. Without disagreeing with any organization making that claim, the truth is that many do not. Personnel scheduling can easily devolve into a political exercise or a source of frustration.
So how does a debate about penalty weights and implied costs make this less complex? It doesn’t—and that is where applied data scientists and decision scientists too often miss the mark.
Applying a model or algorithm can introduce some degree of impartiality, but it also creates a new way to game the system. People can lobby for the factors they care about or, more subtly, for the factors that produce the outcome they preferred all along, protected by the convenient excuse that “the model said so.”
The challenge is not simply producing a schedule. Most organizations can produce a schedule. The challenge is producing a process that consistently generates good scheduling decisions. That’s different. The schedule is simply the visible result. Behind it are policies, priorities, tradeoffs, information, and human judgment. Change those inputs and the schedule changes with them.
The trick is how to use Data Science and Artificial Intelligence (DS/AI) to improve the process without creating a “SkyNet” solution that impassionatetly rules over people’s work schedules, quality of life, and organizational performance.
There is a concept used far too infrequently: Human-in-the-Loop decision science. Instead of treating software as a black box within an otherwise manual workflow, an alternative is to use software to do the heavy lifting of managing and processing data while treating human decision-makers as an equally critical part of the process. People can account for factors that cannot be measured quantitatively or expressed directly in data.
The underlying power behind DS/AI is the ability to identify and learn from complex patterns in data. So why not design solutions that leverage people for the parts they do well and let software do the parts it does well? Rather than asking users to explicitly quantify every goal, preference, and tradeoff, we can create systems where they interact naturally with models and algorithms to shape the outcome. The software can then learn from those interactions and produce better recommendations over time with less human intervention.
Humans & Algorithms
This leads to a well-known challenge in DS/AI. If all we do is use algorithms to automate decision-making without improving those decisions, the return is limited to reducing the amount of work people have to do. Viewed negatively, that can look like taking jobs away from people. The skeptical practitioner would also ask whether the users we hope to learn from are making good decisions to begin with.
From experience, one of the most telling moments in gaining user acceptance is when people first look at the results from a model and see individual decisions they would have made differently. Even if the goal were complete automation and perfect replication of user decisions, that is usually impossible to achieve in complex, large-scale systems.
More importantly, it may not even be desirable. If the goal is to produce better decisions, then some of those decisions should be different from what a user would have made. That is not a failure of the model. In many cases, it is evidence that the process is working.
Because of this, many DS/AI practitioners remain reluctant to think beyond models and algorithms. Users are often viewed as obstacles to overcome rather than components that should be deliberately architected into the flow of information and decision-making.
The case for Human-in-the-Loop optimization can be summarized by another Gene Woolsey observation: “Any system requiring actions not consistent with human nature will never operate up to expectations (if at all).”
In the end, a solution that creates incremental value and that people will actually use is more important than optimizing an objective function with subjective penalty weights to one more digit of numerical precision.
c14blog 