Is Your Team Working With AI, or Performing With It?
AI agent deployments grew 15x in a year. 21% of workers use AI. Among those who do, how much is real work, and how much is volume theater?
Enterprise AI agent deployments grew 15x in a year. Only 21% of workers actually use AI at work. And among those who do — how many are creating real value, and how many are just creating more stuff? I went through the data. Then I looked in the mirror.
My team has been writing code with AI for a year. My own PR count really did double. Sounds like a win — until I noticed I'd become the bottleneck.
Every day I open the code review page and a dozen AI-generated PRs are waiting. Clean structure, perfect commit messages, CI all green. But no human can read them all. So we made a decision that seemed reasonable: let AI review the code AI wrote.
And then? We lean on tests. We hope coverage is high enough and complete enough.
Looking back, the bottleneck moved from writing code to reviewing it — and then we automated the review too. The human in the loop turned into a rubber stamp. We sped up output. We didn't speed up understanding.
This isn't a thought experiment. I know because it's what my team is living through — and the data says it's becoming normal.
The illusion behind the numbers
If you've read my earlier posts, you know I'm wary of stories that sound good. So let's start with the data — because the data itself is messy.
Microsoft's 2026 Work Trend Index says enterprise AI agent deployments grew 15x in a year. Sounds like a revolution.
Pew Research's 2025 update says only 21% of US workers actually use AI in their jobs. Sounds like the revolution hasn't started.
Stack Overflow's 2025 developer survey says 84% of developers use AI tools, up from 76% the year before. Sounds like it depends who you ask.
I've written about this gap before — the distance between leadership ambition and frontline reality. But there was a layer I didn't dig into: among the people who "use AI," how many are creating real value, and how many are just creating more things?
That question led me to an uncomfortable answer.
Three forms of AI fake work
After months of running a team of nearly ten AI agents, and reading plenty of cases of companies pushing AI without managing it, I started seeing the same pattern everywhere. I call it AI Fake Work. It comes in three kinds.
1. Volume theater
The most common. AI makes "producing more" effortless — more docs, more analyses, more slides, more emails. A team goes from one report a week to five.
But nobody asked: was one enough?
Volume theater is output going up while impact stays flat. The AI did its job. The humans just had no system for deciding what was worth producing.
2. The automation illusion
This one hides better. A company automates a process — data handling, customer replies, report generation — and declares victory. Automated. Done.
The problem: they automated the wrong process. The bottleneck was never the task itself. It was the decision the task was supposed to drive. Automating a broken process just gives you a faster broken process.
I nearly fell into this myself. Early on, I built a data pipeline for an AI analysis system that ran every day — collect data, run analysis, write a summary. It ran perfectly. But nobody read the first versions of the report. The pipeline was running, but its output wasn't connected to any decision. It took several rounds to get right — filter out the noise, show only the changes worth acting on. The tool was never the problem. The problem was automating the output before thinking through what decision it should serve.
3. Delegation without oversight
The most dangerous. You hand a task to AI, it produces something that looks reasonable, you approve it without really checking. Not because you're lazy — because you're drowning.
That's what happened when I went from a few agents to nearly ten. By afternoon I was rubber-stamping things I should have read carefully. Most of the AI's work was good. Most. But "mostly good" plus no oversight means errors pile up quietly.
Why it keeps happening
Here's what I think the root cause is: AI deployed without a management framework creates an illusion of productivity.
When leaders roll out AI tools — and deployments grew 15x in a year — they think about capability. Can this tool do this job? Usually, yes. What they don't think about is the management overhead AI brings.
Remember Graicunas? I wrote about his 1933 formula: management relationships don't grow linearly with team size, they grow exponentially. A team of 5 means 100 relationships. 8 means 1,080. 10 means 5,210.
AI agents create the same dynamic. Every agent that produces on its own is an agent that needs reviewing, guiding, and quality control. The more AI you deploy, the more management bandwidth you need.
And here's the paradox: most companies deploy AI to reduce workload — but don't invest in systems to manage the new workload AI creates.
The result? People fall back on defaults — volume theater ("more can't hurt"), the automation illusion ("automated means under control"), and delegation without oversight ("looks fine, ship it"). Nobody is faking work on purpose. The management infrastructure just hasn't kept up with the tools.
Look at Amazon. This year it is putting more than $200 billion into capital spending — mostly AWS data centers and AI infrastructure — after cutting around 30,000 corporate roles since October 2025, with official statements citing AI adoption as one of the drivers, and hiring heavily for AI and ML roles. Three parallel lines: invest in future revenue, optimize current operations, build new capabilities. It's a remarkable transformation. But it's worth asking: of that enormous spend on what AI can do, how much goes into teaching the organization how to manage what AI produces? The tools got a huge budget. Management practice got a blank page. That gap — between the tools you deploy and the management system you build around them — is where AI fake work grows.
The moment I caught myself doing it
I want to be honest: I went through this stage too.
When I first grew my AI team, output went up, clearly. Articles came faster. Data got processed in hours. Infrastructure was maintained around the clock.
Quality, though… to put it kindly, "flat." To put it plainly, "drifting."
One early article draft came back with solid data and clean structure — and read like a product manual. No personal view, no emotional arc to keep a reader going. Before anyone else saw it, I already felt something was off: the kind of article you call "not bad" at first glance and forget the next day. Another time, a process that was supposed to run on its own kept producing results that nobody looked at — not because the output was wrong, but because there was so much of it that checking each one became impossible.
Admitting this took longer than I expected: the problem wasn't the AI. I was managing nearly ten agents the way I'd managed three.
I was delegating without oversight. I was celebrating volume instead of measuring impact. I was doing AI fake work, in my own way.
What real collaboration with AI looks like
So what's the alternative? I don't have a complete answer — but I've spent months building a system, and here's its outline.
Not every task should run on its own
The biggest lever I found: don't treat all AI output the same.
Some tasks AI can finish on its own — low risk, clear edges, easy to check afterward. Some need a human glance — a quick check that the output makes sense before it goes out. Some need a human decision — where judgment, context, and values matter more than speed.
Knowing which is which is the real skill. For example: a daily data summary can run on its own — if something's off, I'll see it in the numbers. An article draft needs a human look before anyone else sees it. And the decision of what to publish, and when? That has to be mine, every time.
Intuition can't carry this. I had to build a tiered system, because my brain couldn't hold all the distinctions.
Role design matters more than tool choice
Something nobody talks about: which model you use matters less than how you design roles.
My AI team's quality didn't improve because I switched to a stronger model. It improved because I separated the "creator" role from the "reviewer" role — the agent that writes the code isn't the one checking whether it runs in production. When I added a role dedicated to research, content quality jumped a level — not because the writing got better, but because the data underneath got more rigorous.
This is basic organizational design. Human teams solved it a hundred years ago. Yet somehow, the moment AI enters, we forget everything we know about how teams work.
Cognitive load is the real constraint
George Miller's finding — human working memory holds about 7±2 items — is the ceiling I hit as my AI team grew. The fix isn't fewer agents. It's fewer decisions that need you.
The goal of any AI management system should be to cut down what needs your direct attention — not to zero, because some decisions really do need human judgment — but to a number your brain can handle.
I've been designing around that principle: the right structure lets you add capacity without adding cognitive load at the same rate. It isn't perfect yet, but it works better than anything else I've tried.
A different angle
I'm not the only one thinking about this. Someone recently proposed a framework that defines AI trust with a math model — score each agent, give it more autonomy when it does well, less when it does badly. The core idea: trust should be earned, not switched on.
An interesting idea. Math gives you precision and consistency.
But my experience says something a bit different: trust isn't a score. It's a context. The same agent, I trust completely on routine data work and not at all on editing that needs fine judgment. That isn't about the agent's track record — it's about the nature of the task, the cost of failure, and whether you'd notice if something went wrong.
Scoring has its place. But managing AI agents — like managing people — ultimately takes judgment that can't be reduced to a formula.
The real productivity question
Here's what I keep coming back to.
The gap between 15x deployment growth and 21% actually using isn't just about adoption speed. It reflects a basic misunderstanding of what AI productivity means.
Real AI productivity isn't doing more. It's doing the right things — and having a management system that can tell which things are right.
Doubling PRs is "more." A data analysis that changes how you think about a big decision — that's "the right thing."
Most companies are optimizing the wrong metric. They measure AI usage, tasks automated, output volume. The metric that matters is: did AI help us make a better decision?
If nobody can answer that, AI isn't being used. It's performing.
So here's a question worth asking yourself: is your AI making you more productive, or just making you look more productive?
The answer may be uncomfortable. But it's the only question that matters.
Part of an "AI in practice" series. Earlier: How My AI Team Turned One Sentence Into a Full Data Analysis, Everyone Says They Use AI at Work, and I Manage a Team of AI Agents. It Almost Broke Me.