Readers: this blog is set in the future (sometime after the year 2020). Each entry assumes there has been a 5th revolution in the US — the Revenge Revolution. More about the Revenge Revolution and author, How the 5th US Revolution Begins and About the Author.  Many entries are formatted as conversations. Characters appear in a number of entries, with many entries building on previous conversations.  

Occasionally I do a “sense check” about the likelihood of a Revenge Revolution.  Entry #318 is the most recent “sense check.”  One more note — sometimes I write about another topic that does not quite fit the theme of the blog.  Those comments are available on the page titled “JRD Thoughts and Comments” as well as “Tech Tsunami”, which has more articles about how technology might affect US…and add a dimension to the Revenge Revolution.

Background to Technology Tsunami series focuses on how implementation of technology may change the family earnings structure in the US.

In “Technology Tsunami” (Entry #319) I noted that with the increased use of artificial intelligence, many current workers will need to increase skills in order to remain employed. But just what is artificial intelligence? And how can it be used? To make AI more concrete and less abstract, thought it might be interesting to allocate the next couple of entries to describing some personal experience developing AI and what happened as a result.today

First, let’s go back to define just what constitutes artificial intelligence, or AI? (Readers, please keep in mind this is not an article for an academic journal. The article is aimed at trying to help the general populace understand more about what AI is and how it might affect the workforce.)

The term “artificial intelligence,” which was first used in the 1950’s, seems to be applied to an ever-increasing range of computer-based applications.  Much of AI we hear about today has been developed by applying to very large data bases sophisticated multiple regressions (regressions look for an association between one action/word and another). The algorithms that result become the foundation for software to support an AI application. What has expanded the use of AI is the availability of very large databases and much more computing power.  However, as demonstrated by this example, a useful and effective AI program can be developed without an overly large database and/or staff.

A question associated with AI, “When AI is implemented, will people be replaced?” Yes, but people have always been replaced with the introduction of new technology. Farm hands were replaced by tractors and mechanical harvesting equipment. The printing press replaced scribes. The telegraph replaced the Pony Express. Trains replaced stage coaches. Cars replaced buggies…and endless other examples.

In the current wave of AI, the jobs that seem most vulnerable in the near-term are ones that involve repetition. Jobs where running a software program or using robot could perform most or all of the task. Such jobs might be assembling parts, loading/unloading shelves, providing certain types of information (clerks, including law clerks could be replaced by a more sophisticated Siri, for example), completing forms or completing some basic analysis (proof reading, financial analysis, etc.), steering vehicles and similar jobs.

The list of vulnerable jobs is quite lengthy and includes a considerable number of white-collar positions. For example, when General Motors announced in fall 2018 the intent to close five plants in the US/Canada, more white-collar workers were affected than assembly workers.

OK, how about a real-world example. In 1980…yes, that was many moons ago…I transferred to headquarters of Buick Division of General Motors. One of the staffs I managed was responsible for forecasting sales – short and long-term. The short-term forecast – 180 days – was used to set production schedules at assembly plants and suppliers.

When I arrived, the accuracy of the forecast was abysmal. Even though Buick had been in business about 75 years, it was not uncommon for forecast sales for the current month to miss actual sales by 30-40%…sometimes 50%. Such a variance made it extremely difficult to manage inventory. The forecast/actual discrepancies also caused frustrations with Buick dealers because arrival dates for cars ordered varied widely from the original schedule, which in turn frustrated customers.

To increase the accuracy of the forecast, we developed an application of AI. The AI-based forecast consisted of three key estimates: (i) industry sales; (ii) mix of sales by category – % small cars, % mid-size cars, % full-size, % SUV’s, etc. – within the industry; (iii) Buick % share within the general categories.

Unlike today, at the time most assembly plants were limited to a few models with little variation in size. Further, changing the production mix at an assembly plant could be time consuming and costly.

Buick’s solution to this dilemma (and common in the industry) was to “force” the dealers to take the mix of cars produced. Further, there was little recognition of differences in consumer preference by region of the country. Dealers in New England, where smaller cars were preferred, would end up with mix of small/large cars very similar to the dealers in say Texas, where larger cars were preferred. “Encouraging” dealers to take the production mix required the field staff to spend considerable time with the dealer and often involved some type of costly incentive – free financing, extra cash per car, etc. Dealers would then have to try to steer customers to these “unwanted” cars.

The solution to fixing the problem was conceptually simple: (i) a more accurate forecast; (ii) allowing dealers to order what cars they wanted. Improving the accuracy of the forecast was the critical first step. Doing so required building a math model that would predict more accurately upcoming changes in demand.

Previous sales forecasts had been based on changes in the rate of actual sales. Basing the forecast on “lagging indicators” – sales the past few months – is akin to trying to drive a car by looking only in the rearview mirror. Doing so reduces one’s speed and increases the chance of making a serious error. The previous method of forecasting was always “catching up” to changes in demand rather than being ahead of the curve.

Developing the AI model was remarkably easy – or so it seems now. We ran regressions of historical sales data for the industry as well as Buick. Fortunately, the auto companies had been reporting monthly sales for many years, so the data base was credible. The results of the regressions yielded useful, seasonal patterns. We also analyzed the shift in mix of sales over time. This helped determine if sales of smaller cars were increasing faster or slower than say mid-size or luxury cars. Another task was estimating how many people were switching from cars to what were then early-version SUV’s.

Finally, we had to determine Buick’s likely share of each category. At the time the overall car market was shifting to smaller cars. While Buick had competitive smaller car entries, it was more successful in larger cars. The effect of the shift in consumer preference was profound. Even though in a given month Buick could gain in market share in every major industry category compared to the previous year, that same month could show Buick’s overall share had declined compared to a year ago.   That phenomenon was always fun to try to explain. “Yes, we gained market share in every category…but, no we lost market share overall.”

Within about one year of starting the AI model, the US industry experienced a major economic downturn and vehicle sales took a nosedive. The AI model helped Buick management begin to make more informed decisions about setting production schedules and marketing plans. With the implementation of the AI-model, the accuracy of the forecast improved markedly. Rather than a variance of 30-40% between actual and forecast for a given month, the variance fell to less than 5%. The improvement helped smooth production schedules, reduce short-term layoffs and/or overtime at Buick and suppliers and made lead-times for deliveries to dealers much more accurate.

The increased forecast accuracy allowed Buick to migrate to what is called a “free-expression” forecast and production schedule. Dealers were allowed much more freedom to order the number and model of cars they wanted.

The decision to migrate to “free-expression” forecast/production caused great angst among staff members tied to the old “dealers-will-order-what-we-tell-them” system. In the end, however, most everyone became a convert because the overall production volume and mix were about what the dealers wanted.

Other benefits of the AI forecast model? The field staff was able to spend more time helping dealers with marketing programs, working on customer satisfaction and finding ways to improve profitability. The dealers then started to order more cars from Buick because the turnover rate improved. In the three-year period following implementation of the AI model, Buick increased market share more than any other manufacturer, domestic or foreign. While not all the gain in market share can be attributed to the AI model, the number of new products Buick introduced during the same period was limited, so most of the gain in market share came from “non-product” activities.

What happened to employment? Buick reduced the number of field offices from 26 to 20. Buick also started a call center to increase contact with dealers located outside urban areas. The non-urban dealers still received some personal visits, but less frequently.

Use of AI also changed the skills required of the office staff. To be effective in the new environment, staff members needed more skills in math, statistics, economics and marketing. If today’s computing power were available then, we could have cut the staff in half, possibly more. Even skills of and the number of senior managers would have been affected. At the retirement party of a key sales executive, who’d grown up in the days of gut-feel and seat-of-the-pants forecasts, the retiring executive told me – after several drinks – “I never understood what you were talking about, but I trusted you.” I appreciated the compliment but was a bit taken aback by the admission.

Does this example help us look ahead for what might happen when more AI is implemented? I think so. What did this rather simple application of artificial intelligence help Buick accomplish?

  • + Increased sales
  • + Increased market share
  • + Increased profits
  • + Increased customer satisfaction (dealer and buyer)
  • – Reduced employment
  • – Higher skills required of employees

If you’re a shareholder and/or your compensation is tied to profits, you will view the results of implementing the AI program as positive. If you’re an employee whose job was eliminated and/or you were unable to learn the additional skills required, you will view the AI program as negative.  The inherent conflict between perspectives, unless we quickly start to manage more effectively, will likely be another contributing factor to the Revenge Revolution.

(In the next entry, another real-world example of using AI – an early application of consumer profiling. While the profiling was not as sophisticated as done today by Google, Facebook, Amazon and many others, the effort allowed Buick to spend marketing dollars more effectively.  We’ll also address why it is important that the output of AI programs is understood and trusted. )