The shape of AI-native work

Where the market is, and what it’s getting wrong

In the last eighteen months, almost every productivity tool you use has grown an “Ask AI” button. Most of them are a category mistake. The button is not the product. Adding the button does not make the product AI-native; it makes it a product with an AI button.

A sharper way to put what most companies are doing: they are tacking another wheel onto the back of the carriage and calling it the new era’s car. The carriage gets a fifth wheel; it is still a carriage. The road has changed; the carriage has not. Pretending it has the shape of the road’s future is what most “AI strategy” amounts to today.

This essay is about the shape we think AI-native work actually wants — and why we believe the right answer is much further from the carriage than most of the market is willing to look.

Two forces drive the bolt-on pattern. Incumbents must protect distribution: their value proposition is the existing workflow, and tearing it down to absorb a new productive force is not an option their cap tables tolerate. They add the button. New entrants could in principle redesign from scratch, but most exploit the AI-faith premium instead — a ”+ AI” sticker raises pricing and makes a generic SaaS look special, without doing the harder work of asking what the product should be when the AI is real.

Both end up at the same place: AI as a feature inside a workflow that was shaped before AI was capable. The workflow is the legacy. The button is the lipstick.

What this gets wrong is not a question of polish. It is a category mistake about what AI is. AI is not a feature — it is a productive force. You cannot bolt a productive force sideways onto a workflow built for a previous productive force and expect the workflow’s shape to survive. The workflow itself is going to change.

If we are right about that, what does it mean to build AI-native? That is the real question, and we think it has two layers — not one.

The first layer: the workflow reshapes

When the productive force underneath an industry shifts hard enough, the work that gets done on top of it does not stay the same. It rearranges. Some tasks vanish. Some merge. New tasks emerge. AI-native, in this first sense, means designing for the shape of the work after the rearrangement, not retrofitting AI into the shape of the work before it.

Three concrete patterns:

Vanish. A growing fraction of meetings that exist today are state-synchronization meetings — standups, status updates, sync calls whose only function is to bring people up to date on what each other has been doing. They exist because state lives inside people’s heads, distributed across the team, and no shared surface keeps them aligned in the background. When the workspace itself holds the team’s state continuously, and each member’s AI client can read it on demand, the meeting becomes a re-sync ritual that has nothing to re-sync. The same is true for context-switching across AI client tabs. Today an engineer’s Claude knows things the designer’s Cursor does not; the salesperson’s ChatGPT does not know what either of them is shipping. The team operates as N parallel monologues with N AI clients, each fluent in their corner and ignorant elsewhere. Bring those monologues onto a shared surface and the friction vanishes — not because the AI got smarter, but because the workflow stopped requiring twelve translations.

Merge. Two tasks fuse when one AI loop can carry both. The classic pre-AI separation between “writing a spec” and “starting to build” is a workflow accommodation: humans wrote specs because humans then handed implementation to other humans, and the spec was the interface across the handoff. Once the AI is the handoff partner, the spec and the first-pass implementation become the same loop. The interface across humans-and-agents is the typed row in the workspace: one operation — delegate this — does what used to require writing a spec, opening a ticket, assigning the ticket, and explaining the context in a meeting. The action collapses; the relationship between intent and execution becomes one click.

Emerge. New work appears that did not exist before, because it serves operations that did not exist before. Convention-encoding is one — writing the team’s working playbook into the workspace itself, in plain language, so every AI client a member runs inherits the team’s standards on day one. Agent-output curation is another: when a team’s AI is producing more work than any one human can read closely, someone has to decide which runs land in canon and which are discarded — a craft that has no name yet but is becoming a real role. Context-shaping is a third: choosing what the agent gets fed, in what order, with what framing, is starting to look more like writing than like configuring software, and the people who are good at it are different from the people who are good at writing prompts.

These are not exotic predictions. They are already happening on small AI-native teams, just unevenly and without the workspace shape to make them stable. The point is that the unit of “what work is” has shifted, and the workflow shape has to shift with it. The button on the legacy product cannot do this work because the legacy product was not built for any of these tasks.

Workflow change is the visible part of AI-native. It is not the deepest part.

The second layer: the relations of work reshape

This is the part most of the market is not yet looking at, partly because the surface is moving too fast for anyone to look up.

There is a reading of history, going back to Marx, that says when the productive force of an era changes, it does not just change what gets made and how. It changes the relations through which it gets made — the social structures that organize labor around the new productive force. The industrial revolution did not just make things faster. It made the factory. It made the firm in something close to its modern form. It made the eight-hour day and the org chart and the foreman’s role and the wage relation as we know it. The deep change was not in the steam engine. It was in everything that organized humans around the steam engine.

We think AI is a productivity revolution of the same order. If that is right, the relational layer is going to change too. We are at the very beginning of seeing how, and we want to be honest: it is too early for anyone — including us — to call which way the change settles. What is not too early is to look up and ask the question, which is what most of the market is failing to do.

The connecting theme we keep returning to is this: the boundary of the firm is becoming permeable. Organizations exist in their current shape because, historically, the cost of coordinating work across people was high. The firm internalizes coordination so people can specialize and execute without renegotiating context every time they need to collaborate. This is not a moral or aesthetic claim about firms; it is a transaction-cost claim, the one Coase made in the 1930s. Firms exist because the alternative — pure markets, deal-by-deal — has too much friction to operate. AI may be lowering that friction in ways that, taken seriously, change what the firm is for. We do not know how far. But we know enough to want to think about it.

Three open questions we find interesting — not staked claims, just illustrations of the kind of question that becomes available when we take the relational layer seriously:

Cross-organization networking, brokered by agents. Today, the outside of the org is a thin surface — emails, intros, LinkedIn messages, the occasional sales meeting. The inside is rich (Slack, docs, dashboards) and the outside is starvation rations. What if AI agents could broker richer connections across org lines — collaboration arrangements, business transactions, info-sharing, even friendships and intellectual community — at the same density humans currently have inside their own org? The firm’s interior might not get less rich, but its boundary would become much more porous.

Deliverable-as-unit-of-work. Every deliverable, in the end, is a tuple: input context, requirements, output criteria. The reason we wrap deliverables in jobs and teams and organizations is that one human can only carry so many deliverables at once, and the coordination across deliverables is itself expensive. What if that wrapping is partly an artifact of human bandwidth that AI starts to lift? An Upwork-everywhere world — where every unit of work flows to whoever (or whatever) can do it best, brokered laterally rather than hierarchically — sounds like science fiction now. We are not so sure it is.

AI credit and tokens as personal capital. Today, capital in the AI-productivity stack is mostly mediated by employers — companies buy enterprise plans, allocate seats, control access. Individuals also accumulate AI credits (subscriptions, leftover usage, specialized models they have trained or fine-tuned). What if those become tradeable? An economy where productive resources are individually held and laterally exchangeable — rented, traded, lent — is a different economy than one where AI-as-productive-resource sits inside the firm.

We are not telling you which of these comes true. We are also not telling you these are the only candidates. Our point is narrower: the relational layer is going to move, and right now almost no one is building for the possibility that it will. Our work tries to hold both layers in mind: one eye on the workflow layer (concrete enough to design and ship against) and one eye on the relational layer (too early to ship against, but not too early to take seriously in how we shape the primitives we are building). It is too early for anyone — us included — to call this. That is the point. We invite the reader to think with us.

What is work even optimizing for?

If work is being reshaped, where is it going? To answer that, it is worth stepping back further than the last decade and asking what work has been doing across human history.

A short version of the trajectory: gathering → farming → factory work → knowledge work. Each transition meant something specific changed in the relationship between humans and the work itself. We do not think it is obvious what the underlying objective has been, and we want to be careful here, because this is exactly the kind of claim where it is easy to over-fit a single narrative.

Let us name the candidates explicitly:

Comfort. One reading: work has gradually reduced physical hardship. Gatherers walked all day and ate when they could. Farmers fixed shelter and predictability. Factories enclosed the body in shelter while extracting motion. Knowledge work mostly removes the body’s load entirely. There is real signal here — people consistently choose office work over physically demanding work even at lower compensation. A water-pipe technician or a delivery rider can earn more than a junior office worker in many cities, and the office worker still chooses the office.

Autonomy. Another reading: each transition expanded the worker’s control over time and task. The serf’s day was structured by the lord; the factory worker’s day by the line; the knowledge worker’s day, in principle, by the worker. This is also real. Some of the strongest preferences observed in modern labor markets are about flexibility and self-direction, and they matter independently of total compensation.

Cognitive engagement. A third reading: the trajectory has been a shift from bodily labor to mental labor, where mental labor is itself the reward. People prefer to use their minds rather than their bodies, all else equal — and “all else” includes pay, hours, and even physical comfort.

Meaning, or purpose. A fourth reading: people will accept lower comfort, lower pay, and harder work for meaning. The persistence of low-paid creative, mission-driven, and craft work — and the readiness with which talented people leave high-paid abstraction-heavy jobs for something they can describe to their parents — is hard to explain on a pure comfort axis.

Status. The most cited driver. We think it deserves an honest treatment, because it looks like a candidate but probably is not. People do choose office work partly because the office is socially encoded as higher-status than physical labor. But status is a byproduct of how a society distributes work, not an underlying objective the worker is optimizing for in some essential sense. If a society shifted its prestige hierarchy — paid more respect to skilled trades than to white-collar middle managers — preferences would shift accordingly. Status is descriptive of social ranking, not the thing being chosen for its own sake. We will set it aside.

The honest read of the four remaining axes is that they are all real and they have been moving together. Each historical transition reduced physical hardship and expanded autonomy and shifted toward mental engagement and opened a space for meaning that previous arrangements pinched off. The trajectory is composite, not single-axis. Comfort alone does not explain why someone leaves a cushy product manager job to start a hard creative business — autonomy and meaning do most of that work. Autonomy alone does not explain why creative gig workers still want healthcare and a desk — comfort matters too.

What matters for our argument is what comes next, and here we are stating a hypothesis, not a fact. Knowledge work has largely closed the physical-hardship gap. Ergonomic chairs, climate control, remote flexibility, and the absence of physical danger or strain have reached a point where additional gains are marginal for most people in this category of work. The remaining frontier is the experiential one — the bundle of autonomy, engagement, meaning, and aesthetics. We think the next era of work optimizes there.

If you accept that hypothesis, the next question is: what does experience-optimized work look like, concretely? Our answer is that it looks like a game.

What “game-like work” looks like

We mean “game-like” in a specific sense. A well-designed game is the most refined existing instance of an experience-optimized activity that keeps the player engaged voluntarily for hours. The properties that make it work are well understood, even if no one has tried very hard to apply them to most of contemporary work.

We say “game” knowing the word has been abused. Gamification — points and badges bolted onto otherwise unchanged tasks — is the carriage-with-a-fifth-wheel version of this idea, and we want nothing to do with it. We do not mean leaderboards. We mean that the conditions which make people naturally engaged with a well-designed game can be designed for in work, and that doing so does more for the experience of work than any number of motivational posters or quarterly off-sites.

Six properties in particular:

Clear objectives. A good game tells the player what success is, without ambiguity, at every nested scale. The matching mechanic in a workspace is typed-table rows: every unit of work has an explicit owner, an explicit status, and an explicit due date, sitting in a structured table inside the doc where the work is happening. The unit of work is not a paragraph in a chat thread. It is a row.

Effective feedback. A good game closes the loop between action and consequence fast and visibly. The matching mechanic is delegate-this-row: an operator points at a row, hands it to an agent, and the agent’s output posts back inline as an artifact in the same surface — comments, files, edits to the row itself. The loop closes where it started. The latency of the agent is the only friction; the path back is built in.

Fair and transparent rules. A good game has a model of authority that the player can trust. In a workspace built for this, the human is always the assignee of any task. The agent is recorded in delegatedTo metadata — executor, never assigner. The permissions are server-enforced, not advisory. This sounds like a small implementation detail; it is actually the foundation of the whole design. If the human cannot trust who is responsible for what, the game is not playable.

Quantifiable progress. A good game makes progress visible at a glance: experience bars, status indicators, level numbers. The workspace equivalents are status badges, owner avatars, dates, the shape of a kanban column. The state of the team’s work is scannable in seconds, not extracted from a half-hour of catching up on Slack.

No wasted mental load. A good game keeps the player inside the core loop. It does not make them parse the menu system between every turn. The matching design choices: canonical storage in plain markdown so agents read the workspace state natively; one surface for chat, document, and tracker so the operator does not have to switch contexts to follow a task. The medium is one surface that rewards looking at it.

Strong visual and interactive surface. A good game looks good and feels responsive. This is the property hardest to reduce to mechanics; it is also non-optional. Work surfaces shipped as engineering artifacts — designed for what is implementable, not for what feels good — leak players. A workspace built for these properties has dashboards, action cards, clean navigation, and grid-and-kanban views of typed tables — reference-grade visual design borrowed from products that have already proven the form. A surface that does not feel good to use is not viable as a daily home for someone’s working life.

These six are not a checklist; they are facets of the same underlying property — that the work surface should be voluntarily engaging in the way a well-designed game is voluntarily engaging. No workspace, including the one we are building, yet lands all six perfectly. Some of the items above are still aspirations more than they are present-day reality. The point is that the design constraints are clear enough to aim at, and that aiming at them produces a different kind of product than aiming at “AI features in a productivity tool.”

But this raises a question. If the work itself is increasingly absorbed by AI, what role does the human have inside the game?

The accountability trap, and learning as the answer

We want to name the dark version honestly, before we say what we are doing about it.

The default trajectory, if no one designs against it, is bad. As AI gets more capable and more deeply embedded in workflows, a steadily larger fraction of execution moves to the AI side of the loop. Humans get pushed toward the only thing that legally and organizationally cannot move to the AI: accountability. Approve or reject. Sign off. Take responsibility when things go wrong. Click the button. This is a real risk, not a strawman, and we have already seen early forms of it. People spend their days reviewing AI output without engaging deeply, gradually losing the skills that would let them engage deeply, and end up as second-class citizens of their own work — present in the system because someone has to be the one to say yes, but not actually doing anything that grows them.

If you let this run, the human becomes a load-bearing rubber stamp. The work might still get done, but the human is no longer becoming better at anything by doing it. They are becoming worse — atrophy by approval.

That outcome is not inevitable. It is what happens when no one is designing against it. Someone has to design against it, and we want our work to be part of that.

Our answer is that game-like work is not only an experience strategy; it is a learning strategy. A well-designed game is the most refined existing instance of an activity where the player gets better at the activity, naturally, as a side effect of playing — without setting aside time to study, without grinding through tutorials, without consciously trying to improve. Mastery emerges from engagement. The game’s mechanics and feedback loops do the teaching.

If the work surface has the same design properties, work itself becomes upskilling. Not a separate “training mode” bolted onto the side of the product. The work is the training, in the same way that playing chess is also how a chess player improves at chess.

There are two scales at which this plays out, and both matter:

At the team scale, as members use a workspace built around delegation loops, the team’s own playbook compounds. Patterns surface in everyday operation — which agents handle which kinds of tasks well, what context each one needs, which review checks catch the most defects. The collective-memory primitive captures these patterns as workspace skills, encoded once and inherited by every member’s AI client thereafter. Onboarding shifts from “shadow a senior for three months” to “your AI inherits the team’s playbook on day one.” The team’s accumulated taste becomes a compounding asset rather than something that lives in one person’s head and leaves when they do.

At the individual scale, reviewing AI output well is itself a craft. The operator who works seriously inside such a surface does not stay at the level of clicking approve or reject. They develop a feel for the model’s failure modes by encountering them in legible form, repeatedly. They learn what kinds of context produce what kinds of output. They learn where the model is reliable and where it is not. They learn how to break a complex task into a sequence of agent operations that close cleanly. None of this is taught in a curriculum. It accumulates as taste, the way a player of a strategy game develops intuition without ever reading a manual. The interface is designed so that paying close attention is the natural mode and paying close attention is also how skill grows.

This reframes the human role decisively. Not approver. Not bystander. Player whose mastery compounds over time. The person who works with AI inside such an environment in five years should be sharper, faster, and more capable than the person working with AI in it today, because the environment was designed to make them so. That is the answer to the second-class-citizen risk. There is no other durable answer.

What we are betting on

Most of the market is selling AI as a feature on top of a workflow that was designed before AI was capable. We think that is a category mistake. AI is a productive force, and a productive force of this scale reshapes both the workflow it sits inside and the relations that organize humans around the workflow. The first reshape is already underway, and is concrete enough to design for; the second is too early to call but too important to ignore. The trajectory of work, taken across the long arc, points toward an experience-optimized frontier, and the most refined shape we know of for experience-optimized activity is the well-designed game. The version of game-like work that survives the seriousness test is the one in which playing makes the player better — where the human’s role is not to approve the AI but to play alongside it, growing as they go.

That is the direction we are exploring: a workspace shaped for all four of these — workflow, relations, experience, and human growth — at the same time. We are at the beginning. The shape is right.