• It is easy to think of all the ways in which advice advocating for equal treatment of humans and agents is not right: Don’t expect the AI agent embedded in your marketing workflow to buy Girl Scout cookies. They aren’t going to care if you forget their birthday. They aren’t going to bet on the March Madness pool
• Even the analogy is weak, but let’s think if any aspects of it sound right.
• Where AI agents differ from a computer is that it is much simpler to get them to do the one thing you want done.
• The right way to think about AI agents in the workplace is that they make us—the employees who work with them—something like a computer programmer, with the big plus that we don’t need to know any programming skills. Here’s an example.

• The problem is that you might not get a good response at least on the first pass. That is true for virtually any tool. The difference is that with other tools, including software, if our tool wasn’t giving us what we want, we would raise our hand, call the IT department and have someone fix it. That is not what is happening with agents. For most agents, you fix it yourself.

   
• That is where employees have to deal with the agent to figure out what is wrong with the answer it has given you and how to describe it in a clear way
• We hope the employee can change the prompt to get a better answer. Over time, the quality of answers may change because the outside information changes (e.g., what counts as a “truck” now) or what we need changes. The employee has to keep monitoring the output and improving it.
• The agent has to be supervised. But the agent is like an extremely dumb employee that will not know if it is producing something good or bad, or how to make it better, unless the employee takes that task on
• So, in what sense is an AI agent like an employee? Only in the sense that its output needs to be monitored and corrected.
• We aren’t setting “goals” for the agent, but we might be for the employee using it. We aren’t doing performance appraisals on the agent, although we might be for the employee using them. We shouldn’t expect it to get better on its own. We can’t switch it to another task without a ton of work. We shouldn’t get angry at it.

• And we shouldn’t invite it to happy hour.

   

    

9 more annotations...

• He calls it "performative productivity" - that feeling at the end of a busy AI-assisted day where you've shipped a new landing page, rewrote the sales deck, drafted a content calendar, and fixed the email sequence, yet somehow can't name one thing that actually moved you forward.
• it's a problem we dealt with constantly in manufacturing, and we have a name for it.
• In lean thinking, there are seven deadly wastes - muda, in Toyota's terminology. Overproduction, waiting, unnecessary transport, over-processing, excess inventory, unnecessary motion, and defects
• AI is starting to feel like a very fast conveyor belt that doesn't ask where the belt is going.
• The classic lean metaphor is automating a broken process. If you have a defective operation and you automate it, you don't fix the process - you just produce defects and unnecessary work faster
• You've just accelerated the production of waste. The velocity of output goes up; the value doesn't.
• the Jevons Paradox. William Stanley Jevons observed in 1865 that more efficient steam engines didn't reduce coal consumption - they increased it, because efficiency made coal-powered processes cheaper, so people used more of them.
• he same dynamic is playing out with AI. Cheaper cognition means more projects, more experiments, more long-tail ideas that were previously too expensive to pursue.
• The lean antidote to overproduction isn't speed control - it's the discipline of asking "should we be making this at all?"
• downstream demand triggers upstream production
• Applied to AI-assisted knowledge work, pull might mean asking a few questions before firing up the prompt: What decision does this output support? Who actually needs it and by when? What happens if we don't do it?
• The resource dimension makes this more urgent than it might seem. AI inference is not free.
• What this really comes down to is that AI amplifies your prioritization ability - or your lack of it. If you're already good at distinguishing high-value work from low-value work, AI makes you significantly more effective.
• motion is not progress. Looking busy is not the same as adding value. AI doesn't change that distinction. It just makes the illusion easier to sustain.

12 more annotations...