Artificial Intelligence and middle managers

Last time I wrote about AI and how I see it as a natural extension of data-driven decisions. I alluded to big upheaval elsewhere, but generally left a “Business As Usual” feeling, which I think is true to until we reach a critical mass. At that point we start designing not just decisions, but whole organizations around AI. Once this change takes hold, I think the management profession changes in a profound way.

Most people seem to think managers will do pretty well in the coming robot uprising. In their “The Future of Employment” paper, Carl Frey and Michael Osborne identified “Management, Business, and Financial” professions as on the low end of risk for automation*. They call out the drivers of this: “Our predictions are thus intuitive in that most management, business, and finance occupations, which are intensive in generalist tasks requiring social intelligence, are largely confined to the low risk category.” Given their methodology and exactly what they try to predict (likelihood of an occupation being completely automated), this is internally consistent, but it raises the question: “What are all these managers are going to do when the people they manage are replaced by computers?” 

This is an important question. This 2016 HBR article uses BLS data to estimate 23.8M managers in the US or 17.6% of the workforce. Compare this to one of the most cited cases of automation - self-driving cars replacing drivers - which makes up only 4.4M or ~3% of the workforce**. Managers also over-index on compensation, representing ~30% of total comp. So, we’ve got almost a third of salaries (and, with progressive taxation… maybe half of the income tax base?) on the chopping block, and no one is talking about it. Weird. 

The HBR article mentioned above estimates that with human workers, we should be able to get to an average of ~10 direct reports (a span of control of 10) for managers in the US. I think it’s intuitively obvious that we won’t have the same middle managers supervising the AI equivalent of 10 workers (let’s define the unit “AI full-time equivalent” or AIFTE). Picture Office Space's Lumbergh from the meme up top taking “TPS reports” from 10 AITFEs and passing them along to upper management. Hopefully not the future.

However, TPS reports are an interesting jumping off point for our thought experiment. The meaningless, mindless exercise of compiling TPS reports will be trivial to an AI. You can get a TPS report every second, all with the right cover sheet, without slowing down your AIs appreciably. But what then? Lumbergh can’t process that much information. 

So, I think the first role of managers of the future will be creating the interfaces with their virtual workforce that allow them to gather insight about what’s going on. The first important point here is that the “dashboard”, once static, is immensely scalable. If we completely define the process of summarizing 10 AIFTEs TPS reports and that is set, there is no reason a single manager couldn’t have a dashboard to cover 10,000 AIFTEs.

Secondly, this is probably a technical role. There is some social intelligence required in terms of reporting upwards, but once the workers are automated, the management task is all about talking to computers. These managers are probably not hand-coding neural networks, but they are at the first human-machine interface. Once the interface is simple enough that a non-technical user can work with it, we've eliminated that layer of the organization. In the meantime, the boundary between what’s automated and not will be exceedingly complex. Even I’m not crazy enough to think we’ll have a generic manager-bot that can take on any task. Automating each layer of each function in an organization will likely be bespoke, error-prone, and need to be refreshed on a regular basis (shameless plug: this feels a lot like a BizOps skillset).

Additionally, we won’t want the AIs to actually handle every transaction. There will be anomalous transactions that fall outside the parameters we’re comfortable a computer making a decision on. This could just be something really weird. It could be a high-value customer who has cutting edge voice synthesis detection and insists on talking to a human. The anomalies will vary by function and industry, but dealing with them will be the job of the manager of the future (and maybe their team constituting the last remaining front-line employees). This primarily will determine the span of managers of the future: the rate of anomalies and the capacity to manage them. Of course, the process to manage each anomaly will be documented and fed back into the AI so this scope of responsibility will shrink over time as well. 

In summary, I foresee a group of technical people interfacing with AIs, handling anomalies (perhaps with a small team of people supporting them), and working to roll up these anomalies (and indeed their whole layer of the organization) so that it can be automated as well. This is a complete shift in focus for the management profession and a reversal from the focus on social intelligence that is supposed to protect managers from automation. If I'm right (admittedly a big if), the primary management skill of the future is managing AIs, something that, as far as I can tell, no one currently does or teaches. Even if management in general is protected, “managing AI's” seems poised to become at least an emerging skillset that will be sorely needed. How do we get started building this capacity into our workforce and who is best positioned to take over this profession?

*Figure III on page 37 summarizes the data by occupation, the light blue being the whole “Management, Business, and Financial” occupation.

**I realize that part of the reason replacing drivers is talked about so much (aside from autonomous vehicles being cool) is that it’s one of the last non-offshorable blue collar jobs and provides a toehold for first generation immigrants and many other marginalized populations. Losing that is a big deal.