I have written a fair amount about how AI agents compress work; this one is about the other side.
I once described it this way: if I lost it one day, I would probably feel something like being crippled, and that is not a figure of speech but crippled in the literal sense.
1. Dependence on AI is irreversible#
As you gradually replace your own brain, hands and eyes with a machine, a new risk appears. When the agent is unavailable, because the network is down, the service is out, policy forbids it, the vendor raises prices, or a model update breaks the existing process, my output falls more sharply than other people’s.
For someone without the prosthesis, that day is like any other; someone with it cannot go back to how things were before, because the original muscles are no longer being exercised.
I have entrusted my floor to a machine I do not control.
There is no good solution to this, only a few ways to soften it: not treating a single vendor as the only path, keeping a human-executable version of critical processes, regularly asking myself whether I could still do the job if it went down today.
But these are only mitigations. The fact is that I have already made this trade, and it only goes one way.
2. My abilities are shrinking#
This is harder to admit than the first.
Just as a manager’s front-line skills and familiarity with the floor fade over time, once many things are delegated, my knowledge of how they actually stand rests on trust rather than on the confidence of having done them by hand.
The difference between those two kinds of “knowing” is invisible normally and only shows when something goes wrong. The person who has done it knows where it tends to break and which number looks off; the person going on trust only knows that it should be right.
I have not found a way to keep this without giving up the leverage. The only thing I do at present is to deliberately do a few things myself, not because it is faster but so as not to lose judgement, because once a reviewer loses judgement, all that is left is pressing the confirm button.
And that has a price: it really is slower, so I am trading efficiency for something whose value I cannot be sure of.
3. This is not only happening to me#
The first two are my personal problems and I can carry them; this one is not.
The price is that I, and the whole of society, really cannot go back to the old way of working.
Whether that is good or bad for any individual, we will all hit that wall eventually. When everyone’s output has been raised, what meaning is left in the eight-hour day as an institution?
The eight-hour day is not a law of nature; it is a number negotiated over a hundred years ago, based on how much one person could do in a day back then. When that base is raised several times over for everyone, the institution does not adjust by itself; it quietly raises the standard, and everyone keeps working eight hours.
On this one I have no answer.
The only thing I am sure of is that pretending it does not exist helps nobody, and that almost all current discussion of AI productivity stops at how much time you can save and never asks the next question: whose will the saved time be, in the end?
So why do it anyway#
Because choosing not to does not make these costs disappear; it only means you carry them without getting any of the benefits.
The first and second are prices I can see with my own eyes, and I have paid them. The third is not one I can afford, nor one I can decide alone.
I write it down because a record that only tells the upside is not worth trusting. If the earlier posts made it all sound too good, this is the other half.
If you are evaluating how deep to take this for a team, these three are worth putting on the table first, especially the second, which usually only shows itself a year or two after adoption.
If I can help with your situation, let’s talk: scope, pricing, how to start →