You have already been in this meeting.
A department costs $2 million a year. Someone says AI can do 40% of the work. The next slide cuts 40% of the salaries. It’s labeled as “strategic productivity enhancements.” It clears the board pack because it looks like arithmetic.
It’s actually firing the wrong person.
Let me be clear about what an AI agent actually is, because the word gets used like we all agreed on a definition. An agent is software that does a chunk of work on its own. It drafts the email. It pulls the report. It writes the first version of the contract summary. It works fast and it never gets tired.
Here is the part that matters. It does not always know if the work is actually any good.
Somebody has to. Somebody has to read the draft and decide it can go to the customer. Somebody has to look at the report and know that number is wrong, because that number is never that high in March.
That judgment is the scarce thing in the room. Not the headcount.
The three-intern problem
If you have ever managed people, you already understand this, and probably better than most of the people writing about AI.
Give a manager three eager interns. Do you now have four managers? No. You have one manager and a pile of new output that has to be checked before it goes anywhere. On simple, repeatable work, that manager gets a lot more done. On messy work that takes judgment, the manager gets less done than before, because now they are doing their own job and reviewing three attempts at it.
Agents are the interns who never get better, never learn from your customers, and never walk over to say “hey, I wasn’t sure about this part.”
That last one turns out to be the whole problem. I will come back to it.
My first take on this was wrong
A few weeks ago I wrote in my notes that one employee running three agents has the capacity of four employees. It felt like the right answer. Keep your people. Multiply them. Fight the layoff slide with better math.
Then I went looking for evidence, and my math had the same flaw as the layoff slide. It was just pointed the other way. Both treat an agent as a person you can put on a line next to a salary. Add three, subtract three, same mistake.
An agent is not a person. You cannot put it on the org chart and you cannot bill it out. What you can count is how much work one accountable human is willing to put their name on. That is the real number. Everything else is a guess with a decimal point.
The cuts are running ahead of the results
In McKinsey’s 2025 survey, 17% of companies said AI had already reduced headcount in their function, and 30% expected it in the year ahead. When the follow-up came, 14% said it had actually happened. Companies expected roughly twice the cuts they made.
And the companies making real money from AI were not the ones cutting. PwC surveyed 1,217 senior executives and found 74% of the economic value going to just 20% of organizations. Those leaders were two to three times more likely to point AI at growth and new revenue than at cost. Same tools. They used them to do more work, not to employ fewer people.
Klarna is the case everyone cites and most people stop citing halfway through. Their AI handled 2.3 million customer chats in the first month, work they equated to about 700 people. Hiring froze. Headcount fell 22%.
Then service quality dropped and they started hiring people back.
Two details matter. The 22% came through a hiring freeze and attrition, not layoffs, so even the poster child for AI replacing people never actually fired anyone. And the fix was not abandoning the AI. It was putting people back on the hard conversations and leaving the software on the simple ones. They had stretched their people across too much work. They pulled it back in.
What it costs when you cut the checker
This is the part that never shows up on the slide.
Researchers at Stanford and BetterUp surveyed 1,150 office workers about something they named workslop: AI output that looks finished and isn’t. 40% had received some in the past month. Each incident took an average of one hour and 56 minutes to sort out. Call it $186 per affected employee per month.
Now notice who pays. In a 2026 survey of more than 2,000 American workers, 57% of managers said they had to fix or redo work a colleague produced with AI, against 38% of individual contributors. At VP level and above it was 63%.
The work gets produced at the bottom. The cleanup lands at the top.
So when you cut the junior line because AI can do it, the checking does not disappear. It climbs. It lands on people whose hours cost three times as much, on top of everything they were already doing.
Speed without a quality check is not productivity. It is work you pay for twice.
It’s impossible to eyeball this
There is one more study worth your time and it is the uncomfortable one.
Researchers at METR ran a careful experiment with 16 experienced software developers, working on real tasks in code they knew well. With AI tools, they took 19% longer to finish. Afterward, those same developers said AI had made them about 20% faster.
They were off by roughly 40 points about their own work, while they were doing it.
That study was about developers, but this is not a developer problem. It is a people problem. We are bad at judging our own speed. Your team is bad at it. Your executives are bad at it. I am bad at it, which is exactly how I ended up writing “four employees” in my notes without testing it.
So when someone tells you a department got 40% more efficient, ask how they know. If the answer is that it feels faster, that is not a measurement. That is a mood.
Three agents do not become a team
Here is where my original thinking broke hardest.
I had also written that a marketing team of three, each running three agents, becomes a nine-person team, all of them coordinating with each other. It sounded right. It is wrong, and the reason takes no technical knowledge at all.
Think about how your team actually works. Somebody starts down the wrong path and says so out loud. Two people figure out in a hallway that they had assumed different things about the same project. A junior person asks the question everyone else was too proud to ask. Most of what makes a team function is people catching and repairing misunderstandings before they turn into deliverables.
Software does not do that. Two agents working the same project will quietly make opposite assumptions, and neither one will ever raise a hand. The work still gets produced. It just does not agree with itself, and nobody finds out until a customer does.
The research is blunt about this. A 2025 study catalogued fourteen distinct ways these systems fail, drawn from more than 1,600 real execution logs, and noted that stacking agents together often produces minimal gains. The engineering teams who build them for a living, at Cognition and at Anthropic, arrived at the same conclusion from opposite directions. One refuses to let agents work in parallel on anything that touches shared work. The other keeps its agents from knowing the other agents exist. Neither will let them negotiate with each other, because that is where it falls apart.
Nine agents can chase nine questions at once, and that is genuinely useful. Producing does not split. Nine agents writing one campaign is a committee that cannot hear itself.
My three marketers were never the surplus. They were the judgment.
Whose name is on the permit
Every org chart I’ve built or inherited did one job underneath the boxes. It said who is accountable for what.
That is why drawing agents onto the org chart does not work. A box on that chart means somebody is accountable, and an agent cannot be accountable. You can put it in a box, but what you have drawn is a diagram of who to blame with the blame taken out. And when you draw a line through the human who used to sit there, you have not automated the job. You have made it nobody’s.
Think about a construction site. Plumbers, electricians, framers, all doing the actual work. The general contractor’s name is on the permit. When the inspector shows up, the subs do not answer. The GC does. That is not a formality. It is the thing that makes the whole arrangement work, because somebody with something to lose has to walk the site before the walls get closed up.
That is the unit you should be planning around. One accountable person, the agents working under them, their name on what goes out. The capacity belongs to that unit. The signature belongs to the person.
Microsoft named part of this in 2025 with the “human-agent ratio.” Useful vocabulary, and I use it. But a ratio with nobody’s name attached is still just a slide.
Now look at the scale coming. Gartner projects the average Fortune 500 company will be running more than 150,000 AI agents by 2028. Today, 13% of organizations believe they have the governance to manage them.
Sit with that. Companies are being asked to plan headcount against a workforce they cannot count, cannot assign an owner to, and cannot switch off. Then they are being asked to fire the owners.
Why forecast capacity from a number nobody has measured? Why cut staff against savings nobody has verified? Why govern a workforce you cannot even list?
What to do about it?
Make every agent belong to somebody. Walk the list of AI tools running in your company. For each one, write down a person’s name. Not a team. Not a department. A person who would answer for it if the output landed in front of a customer or a regulator. If you cannot name one, that is not capacity you have. It is exposure. And if the name you write down is somebody on the layoff list, you just found your mistake.
Find out how much one person can actually carry, and do it per function. It will not be one number. Somebody processing invoices might comfortably oversee ten agents. Somebody handling escalated complaints might struggle with two.
Here is the trap: the number quietly shrinks as the work gets harder, and nothing on your chart changes. The person who could handle twelve when the work was routine is drowning at four when it turns messy, and the plan still says twelve. Nobody raises a hand, because nobody can see the line. Measure it before you staff to it. Measure it again before you cut against it.
Count what shipped, minus the time spent fixing it. Not tasks completed. Not seats purchased. Work that actually went out the door, with review and rework subtracted. If nobody is tracking the rework, you do not have a productivity number. You have a hopeful one.
My take
Recently, I wrote that if you are hiring a Head of AI in 2027, you are not ahead of the curve, you are building 2025’s org chart. I still think that is right. I just had the replacement wrong.
It is not a bigger chart with agents drawn onto it. It is not a smaller chart with the judgment cut out. It is a much simpler question, and most companies cannot answer it for a single team right now.
Not “how many people can this replace.” Not “how many agents can we deploy this quarter.”
Who owns this, and can they tell me what it did on Tuesday?
Go back to that $2 million department. The 40% was never sitting in the salary line. It was sitting in how much more your existing people could carry once you gave them the tools and left them the authority to say “no, that is not good enough to send.” Cut them and you did not find the 40%. You moved it onto somebody else’s calendar, at a higher hourly rate.
Capacity was never the constraint. The constraint is how many people are willing and able to put their name on what the machine produced. That number is smaller than your plan assumes. Cutting it is not an AI strategy. It is how the rest of the work gets more expensive.
Count that first. Everything else is arithmetic on a workforce nobody can find.