And Vijay Says…

Most conversations about AI transformation ask whether the technology can do the work. We debate accuracy, cost, reliability and how quickly companies can move from pilots to production. I expect the technology to keep improving. I’ve become more interested in what happens when the answer is yes.

In Migrating Scarcity, I argued that when technology makes one thing abundant, the constraint moves somewhere else. My new book follows one of those moves into the workplace.

Consider an insurance company where AI now resolves 61 percent of claims, up from 43 percent six months earlier. Customers get answers faster, and cost per claim is down. Then a manager looks one row lower in the operating report. For the claims that still reach a person, average handling time is up 28 percent.

The employees haven’t suddenly become less productive. The work reaching them has changed. AI is taking the straightforward claims, leaving humans with a greater concentration of difficult ones. The 28 percent increase is real, but comparing it with the old number can lead management to the wrong conclusion.

There is a less obvious problem. Those easy claims were doing more than generating transactions. They were how new adjusters learned the business, how managers developed confidence in someone’s judgment, and how experienced people noticed patterns no single file revealed. Remove that work and you change how the organization develops expertise, not just what it costs to process a claim.

We’ve heard plenty about AI eliminating entry-level jobs. I’m interested in what happens to the careers that remain. We assume people will move to higher-value work, but that assumes they already have the judgment to do it. If AI removes the work that used to teach them, where does that judgment come from?

These questions became the basis for my new book, The Corner Case: Where Human Value Moves When AI Does the Easy Work. A corner case is the situation that doesn’t fit the pattern. As AI absorbs the patterns, more of the human role becomes handling the cases that don’t, and that is the work people now have fewer chances to learn.

I chose to explore the questions through a fictional insurance company called Northstar. Maya Rao runs its claims operation. Daniel Cho is a veteran adjuster whose experience has made him indispensable. Leah Bennett is earlier in her career, trying to establish herself in a business where the path that produced Daniel’s expertise is changing underneath her.

The company is succeeding at automation. Its leaders are competent, and the technology is delivering much of what was promised. But success creates decisions that are harder than the original technology choices. Maya, Daniel and Leah see those decisions differently, and each has good reasons for doing so.

I won’t say much more about what happens at Northstar. I’d rather readers encounter the choices themselves. The book combines the story with research and practical playbooks for leaders and individuals. It also makes a claim that can be tested instead of asking readers to accept a fictional company as evidence. If AI removes work that helped people learn, and nothing replaces that learning, comparable new groups should take longer to earn independent authority, perform worse on unfamiliar cases, lean harder on senior colleagues, or need more rescuing when the AI is uncertain or wrong. Much routine work teaches almost nothing, and if that is true in your organization, the effect won’t appear. The point is to look before deciding.

I come to this from two perspectives. I spend much of my professional life helping companies make these bets on AI, which means I have a commercial interest in these questions being taken seriously. I’m also the father of a daughter preparing to enter a labor market that may look very different from the one I entered. I believe strongly in the potential of this technology, but I don’t think enthusiasm should exempt us from examining what happens when it succeeds.

The Corner Case is not another attempt to predict how many jobs AI will eliminate or create. I wanted to examine something that gets much less attention: what happens to expertise, responsibility, ambition and trust when the work around which they developed starts to disappear.

I suspect executives and people earlier in their careers will read the same story quite differently. If you’ve led a transformation, I’d like to know whether Maya’s choices ring true. If you’re earlier in your career, I’d like to know whether Leah’s do. Reply and tell me where you disagree, or send this to someone on the other side of your org chart and compare notes.

The Corner Case: Where Human Value Moves When AI Does the Easy Work is now available on Amazon.

Everyone seems to agree on the same story: agents write the code, and the tools we already use stick around. Jira orchestrates the agents, GitHub verifies their work, Terraform enforces policy.

I buy the first half. The second half assumes that because a function survives, the vendor performing it keeps getting paid. That doesn’t follow. Teams will keep needing coordination, version control and infrastructure management. Whether they keep paying today’s prices for them is a separate question.

Start with Jira. It was built to organize human work, and it’s priced that way, by the seat. If an agent finds a bug, fixes it, runs the tests and prepares the release, a ticket may still exist, but fewer people touch it. The workflow survives, but the relationship between work performed and seats sold starts to break down.

Atlassian sees this. Rovo now meters AI activity and access to organizational context through credits, with additional usage billing scheduled to begin in December 2026. GitHub is moving toward a mix of seats and usage too. That’s adaptation, not collapse.

I still think Jira’s traditional economics are more exposed than GitHub’s, and the reason is margin. A traditional software seat costs relatively little to serve. Agent usage carries inference costs that someone has to pay. Usage revenue can replace seat revenue and still leave a worse business, so margin is the number to watch, not just revenue growth.

The best argument against me is that every agent becomes a heavy consumer of Atlassian’s organizational context. If that consumption carries attractive margins, AI could strengthen Jira. My bet is that inference-heavy usage will struggle to match the margins of traditional seats. Atlassian could prove me wrong if its context services command enough pricing power to offset those costs.

GitHub sits closer to the artifact. Repositories, history and provenance matter whoever writes the code. Its problem is different: passing tests isn’t the same as proving the change is correct.

An agent can write the code and the tests and still be wrong about the requirement. A second agent reviewing the work may share the first one’s assumptions.

Imagine a claims system. An agent finds that removing a validation step speeds processing by 35%. The tests pass and the security scan is clean, but that validation enforced an underwriting rule. The agent saw an inefficiency where the business had a constraint.

More tests derived from the same incomplete specification won’t fix that. The underwriting constraint has to be represented independently of the agent’s optimization objective.

Infrastructure has a version of the same problem. Terraform was declarative long before AI. What changes is volume. Ten changes that each pass policy can still add up to a capacity or resilience problem no single check sees.

I expect basic configuration generation and routine policy checks to lose differentiation, with pricing power moving toward whoever manages infrastructure state and cross-system consequences. More agents could also mean more managed resources, which helps incumbents. I’m predicting where pricing power concentrates, not who disappears.

In Migrating Scarcity, I argued that when something scarce gets cheap, the constraint moves to whatever it depended on.

Implementation was the scarce thing in software. As it gets cheaper, deciding what’s worth building and which risks to accept become relatively more important. Some of that work will be automated too.

Authority is the part I think holds, because someone has to answer for it. Regulators, insurers and boards can ask who allowed a consequential change. An agent can enforce a policy and propose a new one, but it shouldn’t be able to give itself permission to remove the constraint it was meant to respect.

Permission to execute is not permission to redefine success.

Scarce isn’t the same as profitable, though. Companies can meet accountability obligations through internal policies, contracts and existing controls. Audit logs, identity systems and attestations already cover part of the problem.

My bet is that a market still emerges around enforcing authority across systems. A repository can prove what changed. An infrastructure tool can prove what was deployed. Neither necessarily knows whether the change broke a business rule maintained in a third system.

I’d guess the winners are platforms that already control permissions: identity and security vendors, and GitHub where authority attaches to code. Atlassian has valuable context, but I suspect context will be easier to bundle than control. Cloud platforms are the wild card.

I’m wrong if agent consumption of Atlassian’s context grows enough to offset lost seats at attractive margins. I’m wrong if companies can delegate consequential decisions through existing automated policies without paying for another governance layer. And I’m wrong about the business, though not necessarily the constraint, if incumbents bundle all of this and charge nothing extra.

One more worry. If agents do the routine work, how do junior engineers build the judgment to own the big decisions later? Apprenticeship meant doing ordinary work first. A company could ship far more software while gradually losing the ability to govern it.

That would be a strange productivity victory. So the question isn’t which tools survive. It’s which constraints stay scarce enough that companies will pay to manage them.

I came across a paper on harness learning that got me thinking about something we don’t discuss enough in enterprise AI. We spend a lot of time worrying about what agents are allowed to do. What happens when they can also change how they do it?

The researchers train a model, called a proposer, to modify the software surrounding another model. This software, usually called the harness, determines how the model uses tools, handles information and carries out tasks. Rather than improving the underlying model, the proposer experiments with the harness and uses execution feedback to find changes that improve performance.

The interesting finding is that the proposer learns something transferable. It can improve harnesses for tasks it hasn’t previously encountered, without changing the underlying model’s weights.

I can see why people are excited about this. It suggests there may be more room to improve AI outcomes without constantly upgrading models. But it also raises a problem for enterprises. What if we eventually become much better at generating improvements than we are at evaluating and authorizing them?

When better performance isn’t enough

Consider an insurance company that uses an agent to settle straightforward claims below $10,000. The workflow includes policy verification, document checks, fraud screening and escalation rules. Some of those checks rarely affect the final settlement decision, but the company performs them because it needs evidence that the correct process was followed.

Now suppose a proposer discovers that removing one of those checks improves throughput by 20% without materially changing settlement accuracy. Its evaluation says the modified harness is better.

The difficulty is that the check may exist for reasons the evaluation never measured. An insurer may need those records to defend a disputed claim years later. A regulator may require them. Or the check may catch a rare class of errors that the test data barely represents.

The proposer has found a way to improve the metric it was given. It doesn’t need to be trying to circumvent a control. The control may simply look like unnecessary work when judged against an incomplete objective.

A change can improve measured performance while remaining unauthorized.

This example is hypothetical. The paper demonstrates improvements on reasoning and question-answering benchmarks, not autonomous modification of regulated insurance workflows. But the distinction it raises matters well beyond insurance.

Permission to execute versus permission to change

Harness learning makes an old distinction considerably more important: permission to perform an action is not the same as permission to change the system that authorizes and performs it.

In my book, Migrating Scarcity, I argue that when technology makes one capability abundant, the constraint often moves somewhere else. One idea I develop there is permission architecture: the authority, evidence and controls that determine whether an AI system may take an action. Harness learning extends that idea one level up.

An agent might have authority to settle a claim for $8,000. That doesn’t give it authority to eliminate a fraud check, change the escalation criteria or rewrite the process by which settlement decisions are approved.

Of course, enterprises already have change management. Banks have model risk management. Technology organizations have release approvals, separation of duties and production controls. None of this was invented because AI came along.

What changes is the rate at which modifications can be proposed, and the possibility that they affect the very controls governing the system. The existing principles still apply. The challenge is enforcing them when software can repeatedly propose changes to other software, potentially much faster than people can understand the consequences.

What should the proposer be allowed to change?

I think the architecture needs to separate the parts of the harness that can be optimized from the controls that govern its operation.

In the insurance example, settlement limits, mandatory documentation, fraud checks and escalation requirements would be identified as protected controls. Business and risk owners would decide which controls belong on that list, while technology would enforce the restrictions. The list would need regular review, because a control nobody remembered to protect is still a control.

The proposer could experiment with retrieval methods, context assembly and other parts of the workflow. Before any revision entered production, an independent deployment process would examine the changes, test their effects and determine whether they fell within previously delegated authority.

If the proposer removes a protected document check, the release gets blocked. It can still recommend the change and provide evidence that the existing requirement is too expensive or that another control would work better. But it cannot grant itself authority to implement the recommendation.

A change that never directly touches a protected control may still undermine it by altering the information it receives or the order in which steps execute. Code comparisons alone won’t be sufficient. The company needs behavioral testing, independent validation, versioned releases and a way to reconstruct which harness and policies governed a particular transaction.

There is also a problem with the human reviewer. A proposer that becomes good at finding improvements may also become good at explaining why those improvements deserve approval. Over time, risk owners could become rubber stamps for recommendations they cannot independently assess. We need reviewers with evidence they can verify themselves, enough time to examine consequential changes and the authority to say no.

Audit trails are necessary, but they don’t solve authorization. Suppose an unauthorized change removes a legally required document check and 50,000 claims are processed before anyone notices. Perfect logs would help reconstruct the incident and identify every affected transaction. They wouldn’t make the change authorized or the missing checks acceptable.

Human over the loop

We often talk about human in the loop, where someone approves individual decisions, and human on the loop, where people monitor automated execution and intervene when something goes wrong.

Harness learning introduces a different responsibility. If the process itself can change, someone must govern what those changes are allowed to be. That’s what I mean by human over the loop.

This requires more than an operations manager watching an exception dashboard. The people governing these changes need to understand technology, business processes, risk and how a seemingly local optimization might affect the broader operation. It is more demanding work than reviewing transactions, and enterprises will need to staff and organize for it.

That doesn’t mean every harness modification needs a committee meeting. Enterprises can preauthorize categories of changes that meet defined conditions and tests. But changes affecting regulatory obligations, financial authority or customer rights deserve a different level of scrutiny.

The economics of permission

If harness learning advances, the cost of discovering possible workflow improvements could fall substantially. The cost of validating those improvements, understanding their broader consequences and authorizing them may not fall nearly as quickly.

This is how scarcity migrates. If generating improvements becomes cheap and fast, the ability to evaluate and authorize them may become the binding constraint. Enterprises could discover that their capacity to govern change, rather than their capacity to generate it, determines how much AI productivity reaches production.

We are beginning to build systems that can propose changes to the way they operate. The research is still early, but the architecture for authorizing those changes can’t wait for it to mature.

AI may learn how to improve the system. The enterprise still has to decide what the system is allowed to become.