bounded vs. boundless: what AI building itself means for people

identifying where you add the most value — and moving toward it — has never been more important

key findings
As AI transforms knowledge work, how do you future-proof your career? The key lies in distinguishing between bounded work — tasks with a natural ceiling — and boundless work, where value is infinite. This article provides an analytical grid to help you shift your role away from vulnerable, routine tasks and toward high-impact strategic areas. By moving toward "the frontier" of innovation and complex problem-solving, people can leverage AI as a tool rather than viewing it as a threat. Redesign your value proposition for the modern workforce.

Anthropic recently published a piece on recursive self-improvement, When AI Builds Itself, and I believe it's genuinely worth your attention. The real headline is in the numbers: Their engineers now ship roughly eight times as much code per quarter as they did a few years ago, and more than 80% of the code merged into their codebase is now written by Claude, not by people.

It's tempting to read that and conclude the same multiplier is coming for all knowledge work. Make that leap, and "AI writes 80% of our code" becomes "AI will do 80% of everyone's job."

But it’s important to slow down before making that leap. Not because the research is wrong — it's careful and largely convincing — but because software engineering is one of the most AI-friendly kinds of work that exists. The findings are real. The generalization is where it gets dangerous.

So let me do something the original piece doesn't set out to do: Take its findings seriously, then ask honestly what this all means for the rest of us. What does it mean for the recruiters, the analysts, the customer service teams, the marketers, the operations people — the knowledge workers who make up most of the economy and don't typically write a line of code.

why coding is the best case, not the typical one

The reason Claude got so good, so fast, at writing software isn't mysterious. Code has a property almost no other knowledge work has: You can check whether it's right instantly, cheaply and objectively.

Code works or it doesn't. Tests pass or they fail. When Anthropic describes Claude going from a 3x to a 52x speedup on a single optimization task in under a year, the decisive detail is sitting right in its own description: "the goal and the success metrics are fixed in advance." The model can write a version, run it, measure it and try again thousands of times, with an honest scoreboard every single time.

When AI builds code, that pass-or-fail scoreboard is the whole engine. It's what lets an AI grind its way to superhuman performance through sheer iteration, without a human needed in the loop.

Now picture your own work without that objective scoreboard. Did the sourcing strategy work? Was that the right hire? Did the campaign land? Was the analysis sound? Those questions don't resolve in milliseconds. They resolve in quarters, if they resolve at all — tangled up with a dozen variables you can't isolate. You can't run 800 versions of how you delivered hard news to a client and keep the best one. You don't get a green checkmark when you make a good judgment call.

Three other things make code unusually easy to hand off, and they're worth naming precisely because they're what most knowledge work is missing:

1. Coding itself lives entirely in the medium AI operates in.
   
   Code is text, written and tested inside a terminal. There's no customer to phone, no legal team to wait on, no physical thing to move, no meeting where the real decision actually gets made. Anthropic's line that "the doing now costs almost nothing" is true largely because, in software, the doing is text in and text out inside a sandbox.
   
   

2. Anthropic is the single most favorable environment on earth for this result.
   
   It built the coding tool for itself. Its engineers are, by its own description, among the best in the world. Its entire workflow is AI-native from the ground up. Generalizing its internal velocity to the average organization — with messy data, legacy systems, compliance gates and change-averse middle managers — is like generalizing a Formula One pit crew's tire change to changing a flat tire on the side of the road.  Same activity. Unrelated numbers.
   
   

3. Software breaks into clean pieces.
   
   Functions, modules, tickets, APIs. The work slices into well-specified units you can hand to an agent. A great deal of knowledge work is the opposite — the value lives in the integration, the context that never got written down, the relationship. The part that resists being turned into a ticket is usually the part that matters most.
   
   

So the honest headline isn't "AI is coming for everything at 8x." It's this: We just got a remarkable result from the most AI-friendly domain that exists. Treating the easiest case as the representative case is the mistake to avoid.

the counterargument we should take seriously

I want to be fair to the other side, because there is a strong case for AI conducting other knowledge work too.

You could argue that the scoreboard advantage of one verifiable truth is quietly eroding. Look closely at the Anthropic research and you'll notice something: a "Claude judge" decides whether a coding session succeeded. A separate Claude then judges whether the next step in an investigation was the right one. Automated reviewers caught roughly one-third of the bugs behind past production incidents. AI is increasingly manufacturing its own feedback — building synthetic scoreboards where none existed before.

If a model can generate a good-enough rubric for subjective, hard-to-measure work, then the protection I just described starts to weaken. Not because the work became verifiable, but because the verifier became artificial.

It’s why we can’t rest the human case on, "AI can't judge yet." That reasoning has an expiration date.  The enduring argument has to sit somewhere else, and I'll get to that. But first, let’s look at the distinction that actually predicts who feels this soonest.

bounded and boundless

Here's a question that matters more than "Can AI do this task?" and gets asked far less: When you make a worker dramatically more productive, where do those productivity gains go?

The answer depends on whether the demand for that work is boundless or bounded, or infinite or finite, if you prefer. One creates ongoing opportunities to add value; the other has a natural stopping point.

Boundless work has a backlog that never empties. There is always more software to write, more to refactor, more features to ship, more technical debt to clean up. So when Anthropic made its engineers eight times more productive, it didn't lay off seven-eighths of its engineering workforce. It shipped eight times as much, finally tackled the long-deferred cleanup and built exploratory tooling that simply wouldn't have existed otherwise. The work expanded to absorb the capacity. Sales behaves the same way — more capacity becomes more pipeline, more markets, more touches. You can always do more.

Bounded work has a ceiling you don't control. There are only so many support tickets in the queue, so many invoices to process, so many claims to adjudicate, so many appointments to schedule. The volume is set by outside demand, not by your appetite to produce. Make each worker eight times more productive here and you don't get eight times the output — the work isn't there to be done. You get the same output with a fraction of the people.

This is the part the comfortable narrative misses. We keep hearing that AI "augments" knowledge workers and makes them more valuable. In boundless work, where there’s always more value to create, that's accurate and exciting. However, in bounded work, where demand has a stopping point, augmentation and displacement can happen simultaneously. The productivity gain may have nowhere to show up except in the headcount line.

And that line cuts uncomfortably across the workforce. Bounded, high-volume processing work clusters at the entry level and at lower wages — precisely the roles that have always been the entry point into a profession. When those roles compress first, organizations lose more than capacity. They weaken the ability to build the experience and expertise people need to become tomorrow’s senior talent. 

pressure-testing the idea

A clean distinction is worth nothing if it can't survive its own edge cases, so let’s pressure-test a few.

is sales actually boundless? 

At the level of a single territory or a fixed addressable market, sales is not boundless;  there are only so many qualified buyers. But aggregate sales capacity keeps finding new segments, new markets, more frequent contact. It behaves more like boundless work at the firm level, even when any given territory is capped. This is worth saying out loud because the broader distinction shouldn’t be dismissed on a technicality.

is customer service actually bounded? 

This is a more interesting failure, and it generates real insight. The volume of inquiries is bounded, but quality is an unbounded dimension hiding inside the same role. Freed capacity can go into proactive outreach, retention and turning support from a cost center into a loyalty and revenue function. This means bounded work can sometimes be converted into boundless work — by redefining the job around the open-ended dimension (the relationship, the prevention, the experience) instead of the capped one (tickets closed per hour).

a map of consequences, not capability

That's the generalization worth carrying out of this: Almost every job contains a bounded dimension and a boundless one. AI compresses the bounded dimension toward zero cost. Whether that's a catastrophe or a liberation depends on whether the role has a boundless dimension a person can move into, and whether the organization actually lets them move into that work. That last clause does a lot of work, and I'll come back to it.

There is an honest limitation of the whole framework: Stated plainly, the distinction between bounded and boundless work is about demand. It predicts whether productivity gains show up as more output or as fewer people. It also predicts the order in which roles are likely to feel that pressure first. It does not tell you whether AI is capable of the work in the first place. It's a map of consequences, not of capability. Useful, but only half the picture. The other half is determining what work can be delegated to AI, and whether AI is the right fit for that work.

the grid

Pair bounded-versus-boundless with a second question — how hard the work is to specify and check — and you get a more honest map than "routine versus non-routine" has ever given.

Call the second axis “dimensionality.” Low-dimensional work is easy to define, easy to verify, easy to hand to a system. High-dimensional work is ambiguous, context-heavy and resistant to a clean rubric. 

The two axes come apart cleanly: Cold outreach is low-dimensional but boundless; complex litigation is high-dimensional but bounded (there are only so many cases). Cross them, and four very distinct kinds of work appear.

the pressure zone: low-dimensional and bounded
This zone includes tier-one support, data entry, basic claims processing and routine scheduling. It is easy for AI to do, and there will be no backlog to absorb the freed capacity. This is where displacement lands first and hardest, and it's disproportionately lower-wage, on-ramp work.

volume expansion: low-dimensional and boundless
This zone includes cold outreach, high-volume content, first-pass screening. AI handles it, and the work expands to soak up the capacity, so headcount holds better. The real risk here isn't unemployment; it's a race to the bottom on quality — an ocean of cheap output that nobody asked for.

capped but protected: high-dimensional and bounded
This zone includes specialist advisory, M&A work and complex litigation. It is difficult for AI to fully do, but the demand is finite. Expect fewer people doing this work over time, with each individual far more leveraged.

the frontier: high-dimensional and boundless
This zone includes strategy, original research, complex consultative selling, design and the genuinely hard parts of recruiting. It is hard to automate and effectively unlimited in scope. This is where freed human capacity should flow, and where the people who can get there will be worth more, not less.

The grid won't tell you your fate, but it will tell you which quadrant a given slice of your work sits in. That's most of what you need to know to make a decent bet about the next few years.

where humans actually fit

The Anthropic piece is admirably direct about where humans still hold the edge in their own work: setting direction, choosing which problems are worth solving, knowing which results to trust and seeing the bigger picture beyond the immediate task. That's a real answer, but it requires an uncomfortable footnote that’s half-acknowledged: The frontier is a retreating line, not an impenetrable fortress.

AI has already taken execution. It's now starting to take methods. It can now solve problems that aren't fully specified and choose its own approach. It's beginning to nibble at judgment. Anthropic's data showed the model's choice of next step beating a human's 51% of the time last November, and that’s grown to 64% just a few months later. AI fluency buys time, but not total immunity.

So the place humans fit can't be the smart parts AI hasn't reached yet. That argument may eventually expire. It has to rest on things that aren't capability questions at all:

• Accountability
  Someone has to be answerable. AI has no skin in the game — no career to risk, no license to lose, no reputation to burn, no one to look in the eye when it's wrong. In high-stakes work especially, the human is there to be responsible for the outcome, not to produce it. That role doesn't dissolve as the model gets smarter.
  
  

• Problem selection
  The research shows AI is excellent once the problem and the rubric are handed to it. But deciding which problems are worth solving in the first place is influenced by what we collectively choose to care about. Even if AI could choose which goals should be prioritized, we may decide otherwise. Whether AI is capable of the work or suitable for the work are different questions, yet that distinction is made far less often than it should be.
  
  

• Legitimacy
  The customer wants a person on their challenging case. The board wants someone accountable. The patient wants to connect with a doctor to understand their diagnosis and next steps. None of that is a capability gap a better model can close — it's about who we're willing to accept an answer from.
  
  

The bottleneck always moves. Anthropic invokes Amdahl's law, and it's the right frame; speed up one part of a process and the constraint just relocates to whatever didn't speed up (A great read on the Theory of Constraints is The Goal). In knowledge work, the parts that don't speed up are trust, making sense of ambiguity, navigating the organization and deciding what matters. Humans cluster at the new bottleneck. The job, increasingly, is to be a bottleneck worth being.

two readings, two different people

There's an optimistic reading of all this and a grim one, and it would be a mistake to treat them as a debate to settle. They're both true; they just describe different people.

the optimistic reading

One person is steering a pyramid of agents, doing the work of ten: This is a senior person's view. It quietly assumes you already have the expertise to steer, the judgment to know when the output is wrong and the standing to be accountable for it. For those people, this era is a multiplier.

the grim reading 

This reading lands on whoever's entire role was the bounded, low-dimensional execution layer — including the junior talent who would have built their expertise in this layer. You can't develop the judgment senior talent has, and that your business will continue to need, if the work that teaches it is entirely automated. The consequence is a weaker talent pipeline, with fewer opportunities for people to build the judgment and expertise organizations will continue to need. 

The reabsorption everyone is counting on — the boundless dimension quietly soaking up displaced capacity — is a choice, not a law. The economics permit it, but they don't guarantee it. Most organizations will be slow, and many will simply pocket the savings. The Jevons Paradox describes an equilibrium that arrives over years. 

The right response is neither panic nor the usual reassurance. It's to be clear-eyed about which dimension of your own work is bounded and which is boundless — and to start migrating, deliberately, toward the part of the job that has no ceiling.

The real question you need to think about, and not just rhetorically, is: Which parts of your work are bounded — a fixed amount of something that simply has to get done — and which parts are boundless, where there's always more value to create if you had the room to create it?

The first kind is being compressed as we speak, and the second kind is where people are going to have to live.