In July 2026, a16z published a piece called "The Next AI Gold Rush: Tokens, Loops, and Management." The argument was simple but not what most people expect.
A lot of people have been rushing to learn how to build with AI. That's right and necessary. But the next trillion dollars of value may not come from building. It may come from managing.
This isn't a shift away from building. Building remains essential. It's the floor. But managing is the ceiling. And right now, almost nobody's talking about it.
For MBA students betting $200,000 on a degree, that gap matters.
The Railroad Analogy
The a16z piece draws a parallel to the 1830s.
The railroad was the biggest infrastructure buildout the world had ever seen. American track mileage multiplied 120x in a decade. Then the system broke.
In 1841, two trains fatally collided in Massachusetts. The cause was a coordination failure. Individual conductors couldn't keep the system safe at scale.
So the railroad companies invented modern management. They hired managers for each geography. They defined roles in writing. They established clear reporting lines.
The railroads became one of the world's first billion-dollar industries. At their peak, they represented about 60% of the stock market.
The pattern: build first, then manage. The building creates the capability. The management creates the value.
We're at the 1841 moment with AI.
What "Managing AI" Actually Means
Building with AI means knowing how to use tools. ChatGPT, Claude, GitHub Copilot, Cursor. You learn the interface. You write prompts. You generate output.
That matters. It's where everyone needs to start. And it's where most AI education stops.
Managing AI means something different. It means defining what good looks like before you delegate. It means building checks to see whether the output is correct. It means understanding a process deeply enough to encode it for a machine.
a16z puts it this way: maybe 1 in 100 employees knows how to give AI real context. The other 99 produce "loops" — agents calling themselves to fix themselves because the task was never specified cleanly.
The token spend isn't the problem. The problem is that most teams haven't built the management habit yet. Nobody taught them how to write a clear task spec for an AI. That's not a personal failing. It's a missing system.
A Simple Framework
If you want to start practicing AI management, here's a basic cycle:
DEFINE — Write down what a good result looks like before you ask AI to do anything. Be specific. "A one-page summary of this paper for someone who hasn't read it" is better than "summarize this."
DELEGATE — Give the AI the context it needs. Audience. Constraints. Examples of what good looks like. The more specific you are, the less the AI has to guess.
EVALUATE — Check the output against your definition. Did it hit the criteria? If not, was the problem your instructions or the tool's limits?
ITERATE — Update your instructions based on what you learned. Over time, you build a library of task definitions that work.
This cycle works whether you're using Claude, GPT, Gemini, or something that doesn't exist yet. That's the point. The framework outlasts the tools.
The Non-Obvious Skill
Here's why this matters for MBA students.
Building with AI is obvious. Everyone knows they should learn it. There are courses, bootcamps, and YouTube tutorials. The market is saturated.
Managing AI is non-obvious. It's not a coding skill. It's not a prompting skill. It's a management skill. And it compounds.
Ethan Mollick, a professor at Wharton, has been making this argument. He watched executive MBA students use AI tools to build working startup prototypes in four days. His conclusion: the scarce skill isn't doing the task anymore. It's defining work clearly enough that a fast, imperfect system can produce useful results without creating a mess.
That's management. Not prompt engineering. Not vibe coding. Management.
A KPMG report from June 2026 found that 92% of technology executives believe AI management will be a vital skill within five years. The executives who run companies are telling you what they'll be hiring for.
Why Business Schools Haven't Caught Up
Most MBA programs teach AI in one of two ways.
Some treat it as a technology topic. A module on machine learning fundamentals. A case study on digital transformation. Useful but distant from the daily reality of using AI tools.
Some teach hands-on building. Prompt engineering workshops. AI app-building electives. These are valuable. But they teach the floor, not the ceiling.
What's missing is the middle layer. The layer where you learn to define what good looks like. Where you build evaluation systems. Where you figure out which tasks to delegate and which to keep. Where you manage AI output the way you'd manage a junior analyst.
That layer doesn't have a course code yet. But it's the layer that determines whether AI creates value or noise inside a company.
The Jagged Frontier
Research from Harvard Business School illustrates why management matters more than building.
In a field experiment with 758 Boston Consulting Group consultants, AI helped on tasks inside its capability range. Consultants using AI completed 12.2% more tasks, finished 25.1% faster, and produced over 40% higher-quality work.
But on tasks outside that range, consultants using AI were 19 percentage points less likely to reach the correct answer.
The same tool helped on one class of work and hurt on another.
A manager who can't tell the difference will scale errors along with output. A manager who can tell the difference will capture the upside and avoid the downside.
That judgment isn't a building skill. It's a management skill.
What This Means for Your MBA
If you're in an MBA program or heading into one, the implication is practical.
Keep learning to build with AI. That's table stakes. But don't stop there.
Start treating AI like a workforce you manage, not just a tool you use. Practice the define-delegate-evaluate-iterate cycle on your own work. Pick a recurring task. Write down what success looks like before you ask AI to do it. Check the output against your definition.
When you do group projects, volunteer to be the person who structures the AI workflow. Not the person who runs the prompts. The person who decides which prompts need running, in what order, with what criteria for success.
You're not behind. This skill is new for everyone. The a16z piece came out yesterday. Mollick's framing is months old. Nobody has a 5-year head start on you yet. But in 5 years, the people who started practicing now will have one on everyone else.
The Bull Case
The evidence for "management is the next layer" is building.
GOLDMAN SACHS — In July 2026, Goldman economist Elsie Peng published research showing AI productivity follows the same J-curve as the PC revolution. Modest drag for the first four years. Flat for four more. Statistically significant gains only at year eight. Peak impact around year twelve. If ChatGPT's 2022 launch was year zero, the real payoff starts around 2030. The reason for the lag isn't the technology. It's the reorganization work — the management layer — that takes years to build. Goldman estimates each dollar of hardware investment required at least $1.70 of complementary investment in software, data systems, and organizational overhaul.
ENTERPRISE ADOPTION DATA — McKinsey's 2026 survey shows 88% of firms use AI in at least one function. 62% are experimenting with agents. But only 23% are actively scaling. That 88% to 23% gap is the management gap. Companies have the tools. They don't have the management layer to scale them.
KPMG — A June 2026 report found that 92% of technology executives believe AI management will be a vital skill within five years. The people running companies are telling you what they'll be hiring for.
HARVARD BUSINESS SCHOOL — The BCG field experiment with 758 consultants showed AI helps on tasks inside its capability range. 12.2% more tasks completed. 25.1% faster. Over 40% higher quality. But on tasks outside that range, consultants using AI were 19 percentage points less likely to reach the correct answer. The same tool helped on one class of work and hurt on another. Managing that frontier — knowing which is which — is the skill.
WHARTON — Professor Ethan Mollick watched executive MBA students use AI tools to build working startup prototypes in four days. His conclusion: the scarce skill isn't doing the task. It's defining work clearly enough that a fast, imperfect system can produce useful results without creating a mess.
The Bear Case
The thesis could be wrong. Here's what the skeptics have.
NBER — A February 2026 working paper based on a survey of nearly 6,000 executives found that roughly 90% of firms actively using AI reported the technology had no impact on productivity over the prior three years. Moody's chief economist Mark Zandi told Business Insider that "so far the productivity impacts from AI appear to be small and haven't really moved the dial on aggregate productivity growth." If the gains never show up broadly, there's nothing to manage.
WORKER RESISTANCE — An April 2026 survey found 29% of employees admit to actively sabotaging their company's AI strategy. Among Gen Z workers, that figure was 44%. A separate survey found 54% of workers had bypassed their company's AI tools in the past 30 days to do the work manually. Harvard Business School researchers documented "symbolic adoption" — surface compliance while quietly undermining the technology. If the workforce never adopts, management skills don't matter.
MODEL PLATEAU — Multiple sources suggest diminishing returns on compute scaling. The parameter-count arms race may be flattening. If models stop getting meaningfully better, the "compounding demand for transformation" argument weakens. The frontier stops moving. Companies catch up. The management skill becomes static rather than compounding.
BUBBLE RISK — Jeremy Grantham says the AI bubble is about to burst. A Wharton paper warns that if productivity doesn't materialize, "the current buildout will be the largest misallocation of capital in history." If capital pulls back hard, companies may stop investing in transformation before the J-curve pays off.
TIMING — The PC revolution took 15 years to show up in productivity data. We're 4 years into AI. Even if the thesis is right, the timeline may be slower than people expect. MBA students graduating in 2028 might be early to a market that doesn't mature until 2032 or later.
Our Assessment
We think the thesis is structurally sound. Maybe 65 to 70% confident it holds for 10 years or more.
The gap between building and managing is real, documented, and growing. Even the skeptics — Goldman, NBER, Moody's — are saying the gains haven't shown up yet, not that they'll never show up. The J-curve pattern suggests we're early, not wrong.
The main risk isn't that the thesis is wrong. It's that the timeline is slower than the hype implies. If gains don't materialize until 2030 or later, MBA students graduating in 2028 are slightly early. But slightly early is better than too late for a compounding skill.
The management layer is being built right now. The people who start practicing it — defining what good looks like, building evals, managing AI output the way they'd manage a junior analyst — will have a head start that compounds for the rest of their careers.
The Compounding Advantage
Building skills can depreciate. Tools change. The framework you learned this semester may be obsolete by graduation.
Management skills compound. The ability to define what good looks like transfers across every tool and every model. It works whether you're using Claude, GPT, Gemini, or something that doesn't exist yet.
a16z argues that "AI transformation companies" will be 10x larger than any neofirm. The reason: every use case an organization adopts surfaces ten more. The more AI-enabled a firm becomes, the more transformation it consumes.
That's a compounding demand for people who can manage the transformation. Not build it. Manage it.
FAQ
What is AI management?
AI management is the practice of defining what good output looks like, building evaluation systems to check AI work, and deciding which tasks to delegate to AI and which to keep with humans. It's a management skill, not a technical skill. Professor Ethan Mollick at Wharton has framed it as the key AI capability for business leaders.
Is AI management more important than learning to build with AI?
No. Building with AI is the foundation. You need to understand the tools before you can manage them. But building is the skill everyone's learning. Management is the skill almost nobody's learning. That's where the opportunity gap is.
Should MBA students learn to code?
Coding remains valuable. But the specific coding skills matter less than the systems thinking that coding teaches. AI management requires the same kind of structured thinking — defining inputs, outputs, success criteria, and failure modes. If you can code, you'll pick up AI management faster. If you can't, you can still develop the management skill through practice.
What companies are hiring for AI management skills?
According to a KPMG report from June 2026, 92% of technology executives see AI management as a vital skill within five years. The role doesn't have a standard title yet. It shows up as "AI product manager," "AI operations lead," "AI transformation manager," or simply as part of a general management role. The function matters more than the title.
How do you start practicing AI management?
Pick a recurring task in your work or studies. Before you ask AI to do it, write down what a good result looks like. Be specific. Then check the AI's output against your definition. If it falls short, figure out whether the problem was your instructions or the tool's limitations. Over time, you'll develop a library of task definitions and evaluation criteria. That library is your AI management skill.
Will AI management skills still matter in 10 years?
The underlying skill — defining what good looks like and checking work against that standard — is timeless. It's the same skill that makes someone a good manager of people. The specific tools will change, but the management discipline transfers. Goldman Sachs research suggests AI productivity gains may peak around 2034, which means the demand for people who can manage AI output could grow for at least the next decade. Even in the bear case — where AI productivity gains are slower than expected — the management skill still applies. You'd just be early rather than wrong.
What if AI productivity gains never materialize?
A February 2026 NBER working paper found that 90% of firms using AI reported no productivity impact over three years. That's a real risk. But the same pattern happened with the PC revolution — flat productivity for years before the payoff arrived. The Goldman J-curve research suggests the lag is about reorganization, not technology. If the gains do materialize, the people who built management skills early will be positioned first. If they don't, the skills still transfer to managing people and processes. The downside of practicing AI management is low. The upside is high.
Want to learn even more?
We've been covering the intersection of AI and MBA admissions on the blog. If you're thinking about how AI fits into your MBA journey, these posts may help:
- AI and MBA Admissions: What Applicants Need to Know — how business schools are updating their AI policies for applications
- Why AI Doom Fears Could Make This the Best MBA Cycle in Years — why anxiety about AI in admissions may actually help applicants
- What the WSJ Got Right (and Wrong) About MBA Demand — how to read MBA news critically, a skill that applies to AI hype too
And if you're preparing for the GMAT® and want to talk strategy, we're here to help.