I’ve used the term “unit value of labor” a lot in my blogs, because I’ve found it central to understanding the goals of automation in general and AI in particular. I find that most people who see the term have a pretty good idea of what it means, which is (roughly speaking) the cost of the worker when all factors impacted by their hiring or elimination are considered. In most cases, you can say that it’s the worker salary and benefits package cost. But how does it relate to automation and AI? That’s a much harder question, and while I can’t be absolutely sure I have the best answer to it, I have developed an answer from enterprise comments.
Most tech projects these days talk about “productivity”, which is a measure of useful output per unit time. The value of productivity, of course, depends on the unit value of labor of the empowered group. Where unit value of labor is high, it’s reasonable to spend a lot to improve productivity. Where it’s low, you’re very limited in how much you can spend to empower a group.
Productivity measurement has pitfalls that have to be considered. It’s easy to say that you’ve saved a given group of workers an hour per day, but does that mean that you can actually realize the gain? Only if it’s possible to do the same work with fewer workers. That means that the specific nature of the work done by a target group has to allow for reduction in worker numbers if work is facilitated. For example, if you have two workers in a target group, in jobs that require three-shift operation, you could only realize a savings from a productivity gain if it allowed the number to be reduced to two. If there is only one worker per shift, and if some staffing is needed 24×7, you can’t realize a savings unless you can do the whole job without any human actions, or give the worker other things to do with the “saved” time that can deliver actual workforce reduction elsewhere.
This is often the problem with making a business case for AI, say the enterprises I chat with. It’s not that it can’t save time, but that it can’t save time in a way that easily converts to an improvement in company profits in some way. Suppose, they say, you help workers cut a half-hour per day in activity by enhancing their ability to produce emails or other documents. That’s one sixteenth of their day, or about six percent. If you needed 16 workers in the group you empowered this way, and can now get by with 15, you actually realize a savings. If not, you’re probably just giving workers free time.
The unit value of labor thing plays here too. In most cases, companies find that this sort of email-and-document empowerment is valuable for workers with a fairly low unit value of labor, which means that even if you saved that one worker in sixteen, you didn’t save a lot of money, and to get it you had to pay for AI for 15 workers. That makes it hard to hit the ROI target.
Another (as if we needed one) complicating factor is the regulatory/political/contractual framework the target worker groups are working in. Suppose you can’t reduce the workforce, or can do so only with difficulty, given those factors? How do you generate benefits? Perhaps you can say that the current workforce, empowered with AI, doesn’t need to be augmented if more work comes along, but how long would that take and is there really a credible source of that new work?
Finally, you have to consider the question of how difficult it will be for the target group to learn the new processes and be proficient. AI is a long way from an easy technology to use, and use correctly. Most of us, even those with considerable technical skills, have experienced major problems in using AI. I know I have. It’s easy to find examples of this on YouTube, or on websites that use AI to digest information without human quality control. For example, one enterprise told me that in their experience, AI coding was like the work of a very junior programmer. If you tried to build a project using AI programmers, you could not get someone who was comparable to a mid-level type, so you needed more senior people to check the work. In the end, you often ended up with higher costs and no improvement in code quality. And one major mistake can have major consequences.
It would be useful to offer guidance on what unit value of labor would be needed to make automation or AI effective in making a business case, but this is difficult because of the differences in labor costs and policies, regulations, and so forth, that we find globally. Generally, though, enterprises offer information that points to something like a five-tier grouping of workers. Note that this grouping is within the economy and not per-company or per-vertical.
The top 20% of workers in unit value of labor are generally a group whose productivity can easily justify AI. My own modeling and experience in the US market sets this group as the top 15%.
The next 20% of workers can justify empowerment under many conditions, particularly if the worker groups targeted are scarce in the labor market. Even reasonable care here would likely justify AI in as-a-service form, and a bit over half of AI infrastructure projects enterprises have reviewed for this target group can make a business case.
The middle 20% group of workers are unlikely to justify capital projects, and special attention will be needed to ensure that any time savings projected can actually deliver benefits, or that improved work here can save effort for workers in higher tiers. It may be easier to empower this group by redesigning the overall business process than by giving workers AI tools to facilitate the current process.
The fourth quintile of workers will be difficult to empower with suitable ROI in a direct sense, meaning giving the worker an AI tool. Empowerment would have to be done by using AI to remake the work itself, to shift some burden to automation/AI and to make the overall work process less labor-intensive.
The bottom 20% of workers in unit value of labor are difficult to empower with AI, and significant improvements in AI cost will be required to address them. Today, in almost all cases, enterprises will be significantly challenged to fund a capital infrastructure project targeting this group, and even a hosted service will be useful only if the cost is low and special care is taken to address the complications I’ve noted above.
Note that, if we consider these five economic tiers as outlined as “layers” or “rows”, there’s another “column” dimension, which is whether the worker operates at a fixed position (like a desk), or moves around. The former group, which makes up 60% of the total workforce, is easier to empower since they have easy access to computerized devices. The latter group takes more effort, and would be more costly.
The point here is that we can already see, and should continue to expect to see, successful AI projects focusing on the top two economic “rows” within the “desk” column. This is what the most successful enterprises are doing, and why they tend to see AI as a component in a workflow that already touches these workers in some way. In any event, productivity value depends on the value of the workers whose time is being saved, and the extent to which that savings can actually be captured. Ignore that truth and you’ll never be able to assess the potential of AI or any other automation technology.
