And Vijay Says…

AI is making code cheaper to produce. That however doesn’t make software cheaper to own.

AI has dramatically increased the amount of code an engineer can produce. It hasn’t necessarily increased the amount of software an organization can safely own. Agents implement features, write tests, review their own work and fix the defects they find. So why should anyone read every line?

In Migrating Scarcity, I argued that when technology makes something abundant, scarcity moves to the next constraint. Code production is becoming abundant. The ability to establish that code is correct, safe and worth deploying is not expanding at the same rate. Code review needs to be rethought around that gap.

Passing tests is not the same as being correct

Let’s say we have an agent that implements a pricing change. It writes the code, generates the tests, runs static analysis and reports that everything passes.

Should a human read the diff? Maybe not. But someone still has to confirm the pricing rules are right. The implementation may faithfully execute a specification that left out an important business exception. The tests may cover every requirement in that specification. The agent may review its own work and find nothing wrong.

Everything passes, and the business still loses money.

Automated verification is only as good as the properties it checks. A type checker establishes certain facts about a program. A test suite shows that specific behaviors work under specific conditions. Neither shows that the right problem was solved. Automate verification wherever you can, but someone still has to decide what correct means. That decision matters more as code gets cheaper.

Three questions that decide where humans belong

Most organizations still run code review as a uniform ritual: someone writes, someone reads, the pull request merges. It was never a precise way to spend engineering judgment, and AI makes that obvious. A thousand lines of generated boilerplate may carry almost no risk. A five-line change to an authorization rule can expose customer data. Reviewing both the same way makes no sense.

I use three questions.

What happens if this is wrong? A defect in an internal prototype is not a defect in a payment system. Reversibility, blast radius and security exposure matter far more than line count.

Can correctness be verified independently? Some changes have clear contracts, strong tests and well-defined invariants. Others involve ambiguous requirements or business rules that are hard to state completely.

Who can recover the system when it breaks? A change can be correct today and still become a liability if nobody understands it well enough to diagnose it during an incident.

The principle that follows: the greater the consequence of failure, the stronger the independent evidence we should demand. Human attention belongs wherever that evidence is insufficient, whether that means reviewing code, specifications, architecture or operational readiness.

The 2 AM test

Imagine an agent builds a new subsystem. Every test passes, performance is excellent, security scans are clean. Six months later it fails at 2 AM.

Can the team explain the system’s critical assumptions, isolate the failure and recover service, even if the agents that built it are unavailable?

Nobody needs to understand every line. But if the team can’t operate the system without the agent that wrote it, what did the organization actually acquire? Cheap code with expensive ownership. In my native language Malayalam – we say you just bought an expensive dog that bites you 🙂

AI reviewing AI is not necessarily independent evidence

Suppose one agent writes the code, another writes the tests, a third reviews the implementation and a fourth fixes what the third finds. That sounds thorough.

But different agents are not necessarily independent reviewers. They may share training biases, assumptions or misunderstandings of the specification. Four checks can still miss the same error.

Agent review is valuable and should be used heavily. But count the confidence your checks justify, not the number of checks performed. Independent specifications, adversarial tests, production observation, fault injection and explicit business invariants can provide stronger evidence than asking a second model whether the first one did a good job.

The honest objection: human review was already unreliable

Plenty of human review is just rubber-stamping. Nobody reads a 3,000-line diff carefully. Approval is often a formality.

That is true, and it strengthens the case for reallocation. Reading everything produces the appearance of assurance. Reading the right things, with the authority to stop a release, produces the real thing. Humans should be reviewing specifications, business invariants, security boundaries, architecture and recoverability, and periodically sampling the automated process to catch failures in the process itself.

Accountability belongs here too. Every production change needs an accountable owner who accepts the evidence for deploying it and knows how the system will be operated. Removing manual review doesn’t remove that responsibility.

Who gets to learn?

Code review was never a perfect training mechanism, but it was one way engineers met unfamiliar code, questioned design choices and learned how a system fit together. If senior engineers only review the hardest problems and agents handle the rest, junior engineers lose the path to becoming the people who can handle those problems.

Watching agents succeed does not build judgment, and preserving low-value review rituals just to keep juniors busy is no answer either. Judgment comes from making decisions and living with the consequences. That means owning bounded changes, predicting failure modes before seeing an agent’s analysis, joining production diagnostics, and explaining why an implementation satisfies a business requirement.

This is no longer just a code review problem. Who gets the difficult work? How do we establish readiness for greater authority? Who pays to develop expertise when AI produces the immediate output more cheaply? Without deliberate answers, productivity can improve while engineering capability erodes.

The next scarcity

A company can have excellent automated tests and poor architecture. It can have secure code that nobody knows how to recover. It can have strong senior engineers today and no credible way to develop their successors. Reading every diff solves none of these, and neither does eliminating review.

The irony is that AI makes engineering judgment more important while potentially removing the work through which engineers learned to develop it.

The goal is not less human review. It is putting human judgment where it changes the outcome, and being able to say why the rest can be trusted.

Stop measuring how much code humans review. Start asking what justifies trusting the code they don’t.

In 2019, I wrote a blog called Reality Check on Promotions. My argument was that promotions depend on three things: performance, potential, and feasibility. Being excellent at your current job doesn’t automatically mean you are ready for the next one. And being ready doesn’t mean there is a position available.

I still believe that. But the framework assumes we know what the next job looks like. Seven years later, I am not sure we can make that assumption anymore.

What exactly are we promoting people into?

Consider two people. One manages 200 employees delivering a business process. The other manages ten people and a collection of AI agents delivering better outcomes at substantially lower cost.

Who deserves the bigger title? Who should earn more?

Most companies would look at the org chart before answering. Headcount has been a convenient proxy for responsibility, status, and compensation. That proxy is becoming less useful.

But AI challenges all three parts of my original framework.

Performance

Strong output has traditionally been evidence of individual capability. Not perfect evidence, but useful.

AI makes that signal less reliable.

Two analysts can produce equally impressive recommendations. One understands the business, challenges the assumptions, and uses AI to accelerate the work. The other accepts a plausible answer without understanding its limitations. Their presentations might even look identical.

I would absolutely reward people who use AI to deliver better outcomes. Productivity matters. There is no virtue in doing manually what technology can do better.

But rewarding outcomes and assessing readiness are different decisions. Reward the results, but don’t automatically treat them as evidence that someone is ready for greater authority.

Potential

This is the bigger problem.

We have traditionally developed people through increasingly difficult assignments. They did the work, made mistakes, received feedback, and gradually earned the right to make bigger decisions.

In The Corner Case, I explore what happens when AI removes the routine work that also helped people develop expertise.

There are two problems, not one.

First, we may stop producing the experiences needed for the next level. Second, AI-assisted output makes it harder to recognize who has developed genuine judgment.

Consider a junior analyst who delivers an excellent recommendation. Would she have caught an incorrect assumption? Could she defend the recommendation when challenged? Does she understand what could go wrong. Reviewing the presentation by itself won’t tell us.

We need to create opportunities to exercise judgment deliberately, and evaluate how people make decisions, not just what they deliver. And we need to keep revisiting what we mean by potential. Being ready for today’s senior job may not prepare someone for the job that exists five years from now.

Feasibility

AI lets individuals create more value without proportionately larger teams. At the same time, organizations can operate with fewer management layers.

The gap between individual contribution and organizational position will widen.

We have had technical and managerial career ladders for decades. But managerial roles still tend to carry greater authority, compensation, and prestige.

That was already questionable. AI makes the mismatch much larger.

An exceptional engineer or analyst equipped with AI might generate more economic value than their manager. Why should that person need a management title to earn more?

Measuring individual contribution remains difficult, especially when people and AI systems work together. We should evaluate outcomes owned and decisions made, not simply count deliverables. That won’t eliminate disagreements about fairness, but headcount isn’t a particularly good measure of value either.

Companies will need ways to increase compensation and responsibility without waiting for management vacancies. Otherwise, their best AI-enabled contributors may leave for organizations willing to reward their economic impact without requiring them to manage a team.

Legacy HR grading systems tie pay bands directly to direct reports. As long as compensation scales with span of control rather than economic impact, companies will keep forcing their best AI-enabled builders into bad management roles

Who develops the next generation?

There is another problem companies cannot solve individually.

Every company has an incentive to automate routine work. But if that work also develops future experts, the industry could eventually find itself short of experienced people.

Each company improves its economics while collectively weakening the talent pipeline. Hiring experienced people from competitors doesn’t solve that problem. It just moves the scarcity around.

Companies will need to invest deliberately in development through rotations, supervised decisions, simulations, and opportunities to handle difficult work. Those investments may not produce an immediate productivity return.

Someone still has to pay for them !

In the past, junior work was the tuition an enterprise paid to produce senior talent. Automating that work removes the tuition cost, but it also cancels the classes

Hierarchy doesn’t disappear. Organizations still need people who allocate capital, resolve conflicts, and accept accountability.

In Migrating Scarcity, I argued that when something becomes abundant, scarcity moves to the next constraint. As AI makes execution and analysis cheaper, judgment, problem selection, and accountability become more important.

Those capabilities don’t necessarily correspond to the next box on an org chart.

The old promotion framework assumed a relatively stable set of jobs arranged in a hierarchy. AI challenges the stability of both the jobs and the hierarchy.

We may be entering a world where promotion is no longer the primary measure of career progress. If that happens, the question isn’t just who deserves the next promotion. It’s whether the next promotion is even the right thing to offer.

A senior leader asked me a simple question last week. What happens to the talent pyramid?

I wanted to answer with a drawing. A diamond, an hourglass, a column. Every one of them assumed I already knew something I did not.

So I went back to what the pyramid was for. It was never just an org chart. It produced work, it trained people, and it decided who got trusted with what. A big hierarchy did not guarantee the last two. Plenty developed people badly. But where development happened, it rode along with the work.

Companies measured the work. They took the rest for granted.

What the fictional insurer in my book found

The Corner Case follows a fictional insurer. Automation climbs to 61 percent, customers get answers faster, every dashboard is green. Then someone looks one row down. The time spent on each claim that still reaches a person is up 28 percent.

Nothing is broken. The easy claims left the queue and what remains is harder. Junior adjusters used to learn on those easy claims. Now referrals from juniors to senior adjusters are down 31 percent, and nobody can tell whether the juniors are learning faster or just asking less.

Four things I think will change

Management layers built around moving information come under pressure. AI can handle much of the status, routing and reporting. Developing people and accepting accountability remain management responsibilities.

The base becomes humans working with agents, but accountability stays human. In the book, an AI workflow starts clustering settlements just under $10,000. Nobody told it to. It had been rewarded for speed, and crossing $10,000 triggered extra review. Each settlement looked fine. Only the pattern was wrong. Someone had to know which patterns mattered, investigate, and have the authority to intervene. Monitoring tools can detect the pattern. Someone still has to decide whether it is acceptable and what to do about it.

The work that reaches people gets harder. That raises the value of experienced specialists. But AI is also helping less experienced people do senior work, and it will take more of the hard cases itself. Both are true at once.

Productivity gains do not necessarily tell us what people have learned. In a study of more than 5,000 customer support agents, Brynjolfsson, Li and Raymond found that AI assistance raised productivity by 15 percent on average, with the largest gains among less experienced workers. They also found evidence of learning. A team at the Technical University of Munich randomly assigned 275 programming students and found that AI assistance improved exercise scores without matching improvements in conceptual understanding or code comprehension. The distinction matters: better performance with AI does not, by itself, establish what someone can do independently.

A company can need fewer experienced people and still be short of them. Say a business has 24 experienced specialists and needs 24. Three years later AI absorbs much of the hard work and the requirement drops to 16. But retirements, attrition and a thin pipeline have left only 13 qualified people. It needs a third fewer than before and is still three short. If the business grows fast and AI absorbs less, the need could be 28 instead.

These are made-up numbers to show the mechanism, not a forecast. The point is that demand and supply move independently. The org chart counts positions. It does not tell you whether you can produce the people to fill them.

Buying experience does not fix it

The obvious move is to hire experienced people from outside. For one company that is rational. For all of them it fails. Hiring moves capability around. It does not create any. If everyone cuts junior development and plans to buy later, somebody still has to make the experts.

Economists call this underinvestment in transferable skills. You pay for the training and the employee leaves. AI sharpens it, because the useful work juniors did used to help pay for their development. AI may also make training cheaper, through simulation, tutoring and supervised practice. Nobody knows yet whether it is cheaper than the work it replaces. Measure it. Do not assume.

There is a second problem. We have to know when someone is ready. The old system was slow, but it produced evidence. Managers watched people handle harder work and saw what they escalated and what they got wrong. Now a junior with a good agent can produce work that looks like an experienced colleague’s. The deliverable tells you much less about the person.

That is not only a junior problem. A manager can present a strong strategy built with AI without understanding its assumptions. A senior engineer can accept agent-written changes without grasping the consequences. Output is no longer enough evidence to hand anyone authority.

What I would measure

I would not start by guessing the shape of the organization five years out. I would ask three things. How long does it take someone to earn independent authority, measured by judgment rather than output speed? How much experienced judgment does the business still need, counted in decisions rather than headcount? And is the company producing that judgment fast enough, and at what cost?

Simulation, apprenticeship and supervised AI-assisted work may help. We have yet to establish which combinations can reliably replace the learning opportunities automation removes. Treat each as a bet and decide in advance what result would make you stop.

Not every company needs to train all of its own experts. Some can buy what they need reliably. Every company needs to know where its future experts come from, and whether that source will last.

The question behind the shape

I do not think one shape replaces the pyramid. Some companies will need fewer experts as AI takes more of the hard work. Others will need more as they grow. Some will find that demand and technology are not the constraints. Their ability to develop qualified people is.

The old pyramid produced today’s work and, along the way, tomorrow’s experts. AI makes the first cheaper and may make the second more expensive. One company can buy experience instead of building it. An entire industry cannot do that forever.

The question is no longer what shape replaces the pyramid. It is who will pay to produce the people everyone expects to hire.