This is a wild statement that does not seem to be supported by any actual data.
What does it mean? Does clicking on a link counts as labor.
What they will have done is asked a human who should be knowledgeable approximately how much time they spend on the activity (e.g.: how long do you normally spend copy and pasting? How long do you spend looking for the files you need? etc.). When you ask someone whose job it is, they tend to overestimate, and on top of that, you break down the questions as much as possible, so these small overestimations compound without it being obvious to the person that they're making a mistake. It's easy enough to spot that you've said 'four hours' but you know a task doesn't to take you a full morning.
Once you've got all these answers, you ask how often the person has to do it.
Then you ask someone in HR for an average salary you can use. Now that AI is doing that work, you multiply the number of hours saved by the average salary, and report that as your savings.
Something, as usual, stinks about these numbers. $1.6M in saved labour costs, and 3.25 years of work in 4 weeks are basically two ways of saying the same thing (labour costs vs improved productivity).
So let's say 3.25 years of work is around 160 weeks, minus the four weeks they actually took, so 156 weeks of productivity savings. Assume 40 hour weeks, that's 6,240 hours 'saved'. Which works out at around $250/hour, which... well. You decide if that's plausible.
I think we might be seeing what happens when people are being paid too much to spend all day emailing each other and jockeying excel/gantt charts/org charts. Yeah for some definition of "work" I guarantee that a LLM could perform 3.25 years worth in four weeks.
> people are being paid too much to spend all day emailing each other
Hmm, this does not sound exactly right. Also, does anybody seriously think that communication is not work, or is not important? A number of really impactful things started from people emailing each other. (Hell, Linux kernel development is still much about people emailing patches each other.)
Coordination consumes a larger and larger amount of employee time to the point that, in the absolute largest organizations, the vast majority of employee time is internal coordination vs. actual improvement/selling of the customer offering.
So if you go from 100 employees to 1,000 employees, they can MAYBE do 4X the work. Not 10X like you'd think. And this effect gets even worse as you scale further.
So if an AI can do 10X more labor in a human day, and can coordinate instantaneously via a central context ledger (say a git repo), it doesn't just create 10X gains in productivity for large orgs. It creates a multiple of that 10X due to also removing the human coordination overhead.
This is why having less people and more agents actually makes sense but the coordination problem remains either way.
And you cannot escape it because it is simply mathematical.
Here's an easy non-AI example:
In the past, a 'computer' was literally a person [1]. If you needed to synthesize large amounts of data, you needed to split the task among a team of people writing things down and then a team of people to check their work after the fact and then a team of people to combine all the work and then a team to double-check the combined work.
Tasks that in the past would have taken a room full of people coordinating with pencils are absolutely done by 1 machine today (what we know as computers) that no longer needs to split that task and coordinate, which is exactly what will happen with 'agents' who can take on vastly more work per unit of time.
The math doesn't care whether the nodes are people, CPUs or language models. If agent A's next action depends on what agent B decided, you've introduced a sequential dependency.
The computer flattened the coordination dependencies of that room full of people by doing all the calculations by itself. As they get smarter, you can theoretically assume 1 agent could eventually run the entire US federal government.
In the historical [human] computer example; if 15,000 calculations needed to be done, a CPU doesn't need to wait on Bob to come back from lunch to do the next 20 calculations...and doesn't need to wait on Alice to combine his work with the 20 calculations done by Jane...and doesn't need Bill to wait for everybody to be done to double check Jane's work.
The CPU does all 15,000 calculations instantly, by itself. This will be similar with AI agents.
1) The purpose of algorithms is ultimately to create value, not compute some fixed value X. This is important as it gives flexibility to choose different value producing tasks where parallelism dominates over serial tasks, whenever the the latter becomes a bottleneck.
2) In terms of producing value, perfect accuracy or the best possible solutions are not always necessary. Many serial tasks can become very parallel tasks when accuracy or certainty do not have to be complete.
3) Solutions that are reusable changes the math further. No matter how serial a calculation is, if something is calculated that can be reused, that serial part becomes effectively order O(1), after calculation if reused exactly, but as neural network demonstrate, many serial tasks become very parallelized after training a model that can be reused for now a wide class of specific problems. Resulting in very amortized serial computing costs.
It doesn't matter how many steps something takes, if those steps are now in the past and the value is "forever" reusable.
4) The economics of serial and parallel computation are not static, but improve relative to economic value achieved. Meaning that demand for cheaper serial time and currency costs result in improved scaled up hardware that delivers cheaper serial costs. This may have less impact than the previous points, but over years makes a tremendous difference on top of all those points.
This can go on.
The point being Amdahl's law certainly applies to specific algorithms, but is not the dominant determinant of computing in general, and not useful application of computing to a significant degree, where problems can be strategically chosen, strategically weakened or altered, and can be strategically fashioned to create O(V) of value - to balance any O(S) cost of serial computing, via direct reuse and generalization.
Theoretically, each of those steps is parallelizable to some extent. Amdahl's law equivalent here would be that some delays are outside the reach of an organization to improve. For instance, a building permit will take the time it takes to be examined based on an external public administration.
I’ve been in that situation and died a little inside everyday. It’s not like being a rentier, because you still have to lose most of your day at the office and pretend to work, and be available in case some higher up needs something so you don’t get caught.
It might sound sweet but it’s hell.
1. https://davidgraeber.org/articles/on-the-phenomenon-of-bullshit-jobs-a-work-rant/
2. https://en.wikipedia.org/wiki/Bullshit_Jobs
3. https://theanarchistlibrary.org/library/david-graeber-bullshit-jobs(Hint: if you have to prompt it to write an email -- you could have saved everyone some time and emailed the prompt instead.)
The lingering question is if the intermediate LLM translation steps will actually make our communication more efficient - or just amplify the already inefficient parts.
I bet in 2000 years they will still be writing about it - yeah, technology changes our lives (for better or worse).
Take construction work. Incredible improvements through power tools, gasoline-powered mobile cranes, etc. The productivity per worker has exploded. A lot of this has been captured by induced demand: we build bigger, taller, grander. But the improvements aren't distributed equally. Which means that crafts that haven't seen much improvement are now more expensive in comparison to everything else. Which has contributed to our buildings having less elaborate facades and becoming more "bland"
The same in clothing. Clothing has become dirt cheap. Even the poorest people can afford new clothing multiple times a year. But in the same transition we have gone from everything being custom tailored to most things only kind of fitting, being made for variations of the most common body shapes. Not necessarily because tailored clothing has become much more expensive (though higher labor costs from higher average productivity haven't helped), but because every other step has become cheaper and tailoring hasn't.
I wonder what we will say about the trajectory of software in a couple decades
I'm sure whatever path this takes will seems obvious in hindsight
We should definitely seek ways to turn things over to AI. Then turn the AI off, and figure out what actually needed to be done.
For example, one task takes a document with data, charts, and metrics, and Perplexity Computer was tasked with creating a 10-page slide deck for a presentation. Prior to AI, that took human capital and labor costs.
I can't say whether the $1.6M in labor costs is legit or not, but these tools are not just clicking links in 2026.
I want to know pre-"personal computer by perplexity"
send me the data and ill ask my own AI to do it in my favorite silly voice.
I think their numbers of $1.6M and 3.25 years is still probably a massive overestimate, but the order of magnitude seems plausible.
The typical market research , Google analyze , put into spreadsheet is almost gone job. Imagine how many people were doing that as major part of their work