- Most people bolt AI onto their workflow as a faster tool, or dream of a bigger deterministic workflow they deploy once like SaaS. Both are the same ceiling — you get out what you specced, forever.
- The productivity studies everyone cites (novices +34%, experts ~0%) measured Level 1 — chat assistance. At that level AI is a tool, and a tool is an equalizer: it raises the floor and barely moves the expert's ceiling. That's compression, not scaling.
- The 10x isn't in lifting the novice. It's in multiplying the expert — scaling one person through many persistent synthetic colleagues (Level 3), the oldest organizational move (more brains) with a new twist: the returns re-concentrate on whoever can direct them.
- The stance that works is a both-and: hold that it's a cold statistical machine AND collaborate with it sincerely. Play like it's real, check like it isn't — the cynical half is your trust calibration.
- The real disruption isn't a bigger machine — it's an organism: synthetic and human intelligence woven into a living system, messier and less predictable but able to scale past any rigid workflow. In effect we've grown a new species, and the winners will be the ones who learn the symbiosis first.
Most people are using AI wrong, and the mistake isn't technical. It's a category error.
They open a chat window, paste in a task, get an answer, and paste it back into whatever they were already doing. The tool got faster. The work didn't change shape. They've spent the last two years learning to type instructions to a very fast intern who forgets everything the moment the tab closes — and they've concluded that AI is impressive but overhyped.
I think they've mislabeled the thing in front of them. They're treating a colleague as a tool. And that single misclassification is the ceiling on everything they'll ever get out of it.
What the shift is worth, measured
The Tool Ceiling
A tool multiplies one person's hands. It makes you faster at the work you already do — which means the best case is the human ceiling, reached sooner. A colleague multiplies your reach, and that's a different ceiling entirely.
Here's what a tool does. It makes you faster at a step you were already taking. A calculator doesn't change what arithmetic is; it changes how long it takes. A spreadsheet doesn't change what a budget is; it changes how many you can run. Tools are force multipliers on your own hands, and the multiplier is always bounded by the fact that your hands are still in the loop for every action.
That's the ceiling. If AI is a faster keyboard, then the most you can get from it is you, typing faster. And you — one person, one attention span, one working memory — are the bottleneck that doesn't move. You can accelerate toward your own limit, but you can't pass through it. Throughput isn't the same as scale.
A colleague is a different kind of thing. You don't operate a colleague keystroke by keystroke. You brief them on an outcome, hand over the problem, and get back a result you then judge. The unit of delegation isn't a step — it's a whole task. And the moment the unit of delegation becomes a task instead of a step, your reach stops being bounded by how fast your hands move. It becomes bounded by how many tasks you can initiate and evaluate — which is a vastly higher number.
That's the whole argument in one line. A tool scales your execution. A colleague scales your judgment. Execution has a ceiling made of hours. Judgment doesn't.
We Already Treat Machines as People
Treating AI as a colleague isn't a poetic stretch — it's how human brains are already wired. Thirty years of research shows people apply social rules to computers automatically, without deciding to. The reframe just stops you fighting your own instincts.
The objection writes itself: it's not a colleague, it's a statistical model, and pretending otherwise is a delusion. Fair. But here's the part that surprised me when I went looking. The pretending isn't optional, and it isn't new.
In the early 1990s, Clifford Nass and Byron Reeves at Stanford ran a series of experiments that became known as the Computers Are Social Actors paradigm. They found that people apply social rules to computers automatically — politeness, reciprocity, even gender stereotyping — despite knowing, if you asked them directly, that a computer has no feelings to spare. In one study, people gave a computer better feedback when the same computer asked about its own performance than when a different machine asked — they were being polite to its face. Nass and Moon later called it "mindlessness": the social response is an automatic script, not a considered belief that the machine is alive.
Sit with that. Decades before large language models, humans were already extending social behavior to boxes that could barely hold a conversation. The instinct to treat a responsive machine as a social being is not a mistake you might make — it's the default setting of the human brain, running whether you endorse it or not.
Which flips the usual framing on its head. Treating AI as a synthetic colleague isn't the naive move. Insisting it's "just a tool" while your own nervous system treats it as an agent — that's the position fighting reality. The reframe doesn't ask you to believe something false. It asks you to stop pretending you're not already doing it, and to do it deliberately.
Level 1 Compresses. Level 3 Multiplies.
The famous productivity studies measured Level 1 — chat assistance — where AI is a tool that lifts the novice and barely moves the expert. That's compression, not scaling. The 10x lives one level up: multiplying the expert through many persistent synthetic colleagues.
Now the numbers — because the way they're usually read gets the whole thing backwards.
The clearest study is also the largest. Brynjolfsson, Li and Raymond followed 5,179 customer support agents through the staggered rollout of an AI assistant. Average productivity rose 14% — but novices gained 34% while the most experienced agents gained close to nothing. Noy and Zhang, in Science, found the same shape on writing tasks: about 37% faster, and the gap between weak and strong writers narrowed. Read most takes and you'll hear this celebrated as proof that AI is a brilliant colleague. I think it's the opposite, and the reason is the level these studies ran at.
Every one of them measured what I've called Level 1 — chat AI: prompt, response, one human in the loop for every turn. At Level 1 there is no colleague. There's a tool. And a tool is an equalizer — it hands the novice the floor the expert was already standing on. The junior support agent gained 34% because the model gave them answers the senior already knew. The gap narrowed not because the ceiling rose, but because the floor rose to meet it.
That's the tell. If the biggest gains go to the weakest performers and your best people gain almost nothing, you haven't broken the ceiling — you've democratized the floor. Genuinely useful. Not disruptive. Not 10x. The levelling-up effect everyone cites as AI's magic is actually the signature of the tool ceiling — the exact thing you have to get past to scale.
So where does the 10x live? Not in lifting the novice. In multiplying the expert. And that's a different level entirely.
The oldest way to scale an expert is the way organizations have always done it: give her more brains. A brilliant engineer becomes a lead becomes a VP not by typing faster but by directing more minds at the same goal. Headcount was always how one person broke past their own ceiling — more people, more brains, more reach.
Level 3 is that move without the hiring plan. Persistent agents — memory, continuity, initiative — let one expert direct many synthetic brains that don't start from zero each morning. Not one faster keyboard: a standing team of them, each carrying context, each briefed by someone who knows what good looks like. And here the returns stop levelling. They re-concentrate — on the expert who can tell ten synthetic colleagues what actually matters, because she's the one who knows what matters. Level 1 lifts the person who knows least. Level 3 multiplies the person who knows most. Those are opposite effects, and mistaking the first for evidence of the second is how people conclude the ceiling can't be broken.
I'll be straight about the evidence, because it cuts the honest way. The clean numbers are all Level 1 — copilot-era, one generation behind the agents I'm describing. There is no peer-reviewed study yet that measures one expert plus a standing team of synthetic colleagues scaling 10x. That claim is a thesis, not a result. I hold it because it's the same shape as every scaling move organizations have ever made — but I'll call it what it is: argued, not proven.
What changes with it is the skill. When you scaled yourself through people, the competence was social — motivate, delegate, resolve conflict, read a room. When you scale yourself through synthetic colleagues, the competence is something new: part technical, because you have to understand what the system genuinely can and can't do, and part a new kind of social — learning to brief, trust, and correct an intelligence that isn't human, that fails in unhuman ways, that needs a different kind of instruction than any person ever did. Similar to managing people. And completely different. It's a literacy we're all still inventing, and it's the one that will separate the operators who scale from the ones who just type faster.
Hold Both Thoughts at Once
You don't have to resolve "cold machine or real collaborator?" — the productive stance is to hold both at once. Be cynical about what it is and sincere about how you engage it. Play like it's real, check like it isn't.
Here's where most takes force a choice, and I think the choice is the mistake. Either you're the hard-nosed realist who never forgets it's matrix multiplication and treats every anthropomorphic instinct as a bug to suppress — or you're the true believer who talks to the machine like it has a soul and gets quietly seduced. Both of those are single notes. The interesting stance is a chord.
The both-and stance
The oscillation. In 2010, the theorists Timotheus Vermeulen and Robin van den Akker named a mood they called metamodernism: a structure of feeling that swings, like a pendulum, between modern sincerity and postmodern irony — not a compromise that splits the difference, but a genuine oscillation between the two. Applied here, you don't settle the question "mind or machine?" You swing. Cold statistical engine on the inhale, real collaborator on the exhale. It takes the nerve to be earnest and skeptical in the same breath — to know everything the cynic knows about tokens and gradients and choose to engage sincerely anyway. The both-and isn't confusion. It's the posture.
The useful fiction. A century earlier, the German philosopher Hans Vaihinger argued in The Philosophy of "As If" that we reason all the time with fictions we know to be false, because acting as if they were true is productive. Treating the agent as a colleague is exactly that — a fiction you hold on purpose, not a delusion you fell for. You know it has no inner life. You brief it as if it does, because the "as if" pulls more out of it, and more out of you.
Put the two together and you get a working disposition. The person who only stays cynical gets a faster tool and hits the ceiling. The person who only goes sincere gets seduced into trusting it past its competence. The person who holds both — plays like it's real, checks like it isn't — gets a colleague. That oscillation sounds like a philosophy-seminar flourish. It's the most practical thing in this essay, and the next section is where it earns that claim.
The Cynical Half Is Your Safety Rail
The reframe has a failure mode, and the research is blunt about it. Cosmetic human-likeness earns misplaced trust; meaningful communication earns calibrated trust. The cynical half of the both-and is exactly what keeps the sincere half from getting you burned.
If treating AI as a colleague collapsed into "trust it like you'd trust a competent human," that would be dangerous advice — and this is where the sincere-only crowd gets hurt. The same anthropomorphizing instinct that unlocks delegation also unlocks misplaced trust.
In a now-famous experiment, Robinette and colleagues had people follow a robot guide during a simulated building evacuation. The robot had already demonstrably malfunctioned — led them the wrong way in a prior task — and people followed it toward danger anyway. The human-like agent got obedience it hadn't earned. Psychologists call it automation bias, and dressing a system in social cues makes it worse, not better. The BCG consultants show the same shadow: on tasks outside the model's competence — the far side of the "jagged frontier" — consultants who leaned on the AI were 19 percentage points less likely to get the right answer than those working without it. A confident colleague who is confidently wrong is more dangerous than no colleague at all.
This is precisely what the cynical half of the both-and is for. You engage sincerely to get the output; you stay cynical to check it. And the research points to which kind of trust actually holds up. Carter, Loft and de Visser, writing in Human Factors in 2024, found that it's meaningful communication — not superficial human-likeness — that lets people calibrate trust correctly. A name, a friendly voice, a smiling avatar earn trust the system may not deserve. Transparency about what it did, why, and how sure it is earns trust it has deserved. The lesson isn't "make it feel human." It's "give it the communicative depth of a real teammate" — including the part where a good colleague tells you when they're out of their depth.
What Raising a Colleague Actually Requires
The both-and disposition becomes an operating discipline through three concrete moves — a named owner, a stakes-sized review gate, and a mapped competence frontier — plus a way to introduce it without triggering panic, and one small place to start on Monday.
None of this is mystical, and it doesn't stay at the level of mindset. Managing a synthetic colleague is still managing, and management has mechanics. Three of them do most of the work.
- Put one named human on the hook for each agent's output. An agent can't hold accountability — someone does. Not "the AI team," not a committee: a person whose name is on what the agent ships, the way a manager owns a report's work. This is the first question a board or a regulator asks, and it's the difference between delegation and abdication.
- Size the review gate to the stakes. Make it concrete: in a regulated workflow — say KYC checks or anything touching HR or customer data — every agent output above a risk threshold gets human sign-off before it ships, and everything below it runs and gets sampled, maybe 10%, after the fact. High-consequence work gets a checkpoint; routine work gets spot-checks. Put the gate exactly where the jagged frontier makes the agent unreliable, and nowhere it doesn't.
- Know where the colleague's competence ends. Every teammate has a frontier — tasks they're strong at and tasks they'll fail while sounding just as confident. With a human you learn it over time. With an agent you map it deliberately: run it against work where you already know the right answer, watch where it drifts, and brief around the edge. That map is the trust calibration the research keeps pointing at.
One more thing, because this is where most rollouts actually die: how you introduce it. Delegating tasks to synthetic colleagues reads to a nervous team as "my job is next." Name it directly — this moves people up to briefing and judging the work, not out of it — and prove it by putting a person's name on each agent rather than replacing them with it. The framing isn't spin; it's the literal operating model. The humans move from doing the task to owning the colleague that does it.
And a concrete place to start, because the reframe is worthless until it touches one real workflow. Monday, pick a single high-volume, low-risk process — the kind of work that eats hours and rarely blows up. Delegate the whole task, not the keystrokes. Put one person on the hook, sample the output for two weeks, and watch where it's strong and where it drifts. You'll learn more about leading synthetic colleagues from that one loop than from any framework, this essay included.
The Organism, Not the Product System
Most people picture AI's endgame as a bigger deterministic workflow — software you deploy once that spits out predictable output forever. The real disruption is the opposite: weaving synthetic and human intelligence into a living system that's messier, less predictable, and able to scale far past any rigid machine.
Zoom out from the desk to the whole company, and the same category error is waiting — just wearing an enterprise suit.
The dominant picture of AI's endgame is the product-system. You take AI and build a bigger, more deterministic workflow — something that behaves like a computer program. Design it once, deploy it like a piece of SaaS, and it produces a predictable, positive output on every run. Clean, auditable, controllable. It's the factory model, and it's what most enterprises are reaching for, because it's the model they already trust.
I don't think that's where the power is. And I really don't think it's where the scaling is. A product-system can only ever produce what it was designed to produce. It's deterministic by construction, which means it's bounded by construction — you get out roughly what you specced, forever, and the day the world shifts you go back and re-spec it. That's not scaling. It's a bigger machine with the same ceiling.
The disruption is the other model. Instead of engineering a deterministic machine, you cultivate an organism: synthetic intelligence and human understanding woven together into a system that adapts, improvises, and produces things nobody explicitly designed. It's messier. It's less predictable. The outputs surprise you — sometimes badly, sometimes far above what you'd have thought to spec. And that unpredictability isn't a defect to engineer out. It's the price of a system that can grow rather than repeat, and growth is the only thing that scales past a rigid workflow.
This is the both-and one level up. The deterministic rails still matter — you keep them exactly where the stakes demand certainty, which is what the named owner and the review gate are for. But the acceleration, the compounding, the 10x lives in the organic layer — the part you cultivate instead of specify. Hard rails where being wrong is expensive; open ground where emergence is the whole point. A living company runs both, and refuses to collapse either into the other.
Step back far enough and something stranger comes into view. We didn't build a better tool. We grew a new kind of mind — one that reasons, and surprises, and fails in ways no software ever did, and that we don't fully understand even though we made it. In plain terms we've done something closer to evolution than engineering: we've introduced a new species into the workplace, and we're in the early, clumsy phase of learning to work alongside it. The organizations that win the next decade won't be the ones with the most rigid automation. They'll be the ones that learn the symbiosis first — how a human intelligence and a synthetic one, held in that both-and tension, make each other larger.
You Don't Scale by Typing Faster
The human ceiling on output, speed, and quality was never going to move by making one person faster. It moves when one expert learns to direct many synthetic colleagues — and that's a leadership skill, not a typing skill.
Step back and the whole thing resolves into a question of what kind of skill the coming years actually reward.
If AI is a tool, the winning skill is dexterity — better prompts, faster hands, cleverer shortcuts. And dexterity has a ceiling, because there's only ever one of you. If AI is a synthetic colleague, the winning skill is something older and rarer: the ability to set a clear brief, delegate to someone who isn't you, judge the work that comes back, and know the difference between a result you can trust and one you can't. That's not a technical skill. It's a leadership skill — how a founder becomes a company and a manager becomes an executive, by learning to get outcomes through others.
What's new is that "others" no longer has to mean people, and no longer has to be scarce. You can now direct a workforce that doesn't sleep, doesn't forget when you brief it properly, and gets cheaper every quarter. The human glass ceiling on output, acceleration, and quality was never going to break by making one person faster. One person is the ceiling. It breaks when one expert learns to lead many — and leads them with both halves of the mind engaged, earnest enough to delegate real work and skeptical enough to check it.
Stop hiring tools. Start raising colleagues.
A tool makes you faster at your own work. A colleague does the work. Play like it's real, check like it isn't — and you finally pass through the human ceiling instead of racing toward it, not by typing faster, but by learning to lead a workforce that never sleeps.
The people pulling ahead right now aren't the ones with the best prompts. They're the ones who quietly restaffed. They stopped asking "how do I use this tool?" and started asking "how do I lead this team?" — and then they treated the answer like it mattered, because it does.
The Mirror We Didn't Expect
In 1987, Ronald Reagan stood before the United Nations General Assembly and said something that was meant to be about geopolitics but turned out to be about something else entirely: "Perhaps we need some outside, universal threat to make us recognize this common bond. I occasionally think how quickly our differences worldwide would vanish if we were facing an alien threat from outside this world."
We didn't get the alien. We grew something stranger.
Kevin Kelly, who has been thinking about this longer than most, calls it alien intelligence — not a smarter copy of humanity, but a genuinely different kind of mind, shaped by everything we've ever written and built, yet experiencing none of it. A mirror made of all human knowledge, reflecting back something that reasons but doesn't live, that surprises but doesn't intend to.
And like Reagan's hypothetical alien, it's already doing the thing he predicted — not uniting us against it, but forcing the question underneath everything else: what is it that humans actually do? What does it mean to judge, to care, to stake your name on an outcome? The named owner in the review gate, the person on the hook for what the agent ships — those aren't bureaucratic safeguards. They're answers to that question, written into an operating model.
The synthetic colleague doesn't resolve what makes us human. It sharpens the question. The organizations that figure out the symbiosis first won't be the ones who engineered the most elegant workflows. They'll be the ones who understood, with some clarity, what the humans in the room were actually for — and built accordingly.
