The hottest argument on X this morning is not whether companies should use AI. It is whether private AI windows are quietly breaking apprenticeship, visibility, and operational judgment inside the firms trying to scale with agents.
There is a bad habit in AI discourse: every week gets flattened into the same headline.
AI changes work.
Agents are coming.
Productivity goes up.
Jobs go down.
Rinse, repeat.
That is not what is actually happening on X this morning.
The live argument with real heat is narrower, sharper and much more useful:
Should AI at work live in private windows, or in public systems where everyone can see how the work gets done?
That sounds like a product design choice. It is not. It is a management choice. And increasingly it looks like a consequential one.
The trigger for today's debate is Tobi Lutke describing Shopify's internal AI agent, River. The headline figures alone are enough to wake operators up: 5,938 employees used River in the last 30 days, across 4,450 Slack channels. It opened 1,870 pull requests in the last week in Shopify's main monorepo. Roughly one in eight merged pull requests last week was authored by River and reviewed by humans.
Those are not “we tried a copilot” numbers. That is an operating model.
The more interesting part is the constraint. River only works in public Slack channels. It does not live in DMs. It does not sit inside a private window. It is public by default, searchable by default, interruptible by default and educational by default.
That is the real provocation.
Because most of the AI industry has trained people to think in private windows. ChatGPT is a private window. Claude is a private window. Cursor, for most teams, is effectively a private window. The user asks. The model responds. Some work gets done. Everyone else learns nothing.
Tobi's point is that this design quietly destroys apprenticeship.
He is right, and the implication is bigger than Shopify.
The optimistic case for private AI windows is obvious.
They are fast.
They are personal.
They are low-friction.
They let an employee move without broadcasting every half-formed thought to the company.
That is why they spread so quickly.
But private windows create a structural problem the AI boosters usually skip: they individualise the gain and socialise none of the learning.
One developer figures out how to scope a debugging prompt properly. Great. One analyst discovers how to get better warehouse queries from an assistant. Fine. One operator works out how to use an agent to draft a support workflow, review it and ship it. Useful.
But if all of that happens inside sealed chats, the firm does not really compound. The individual does.
That is a very different outcome.
It means the company pays for intelligence tooling, but gets only fragmented private advantage in return. The best patterns stay trapped in personal workflows. Junior people cannot watch senior people reason with the machine. New hires cannot see what “good” looks like in practice. Managers cannot distinguish serious AI operating force from theatre. And nobody builds shared institutional taste about where agents are useful, where they are dangerous and where they are simply producing expensive sludge.
That is the real cost of the private window.
Not privacy itself. Not even security, though that matters. The real cost is that it blocks visible learning.
Simon Willison clocked this immediately, comparing River's public-in-Slack design to early Midjourney in Discord, where people learned prompting by watching each other. That is not a cute analogy. It gets to the heart of how new craft standards actually spread.
People rarely learn a new medium from official documentation alone. They learn by seeing good work happen near them.
That is true for design.
It is true for code.
It is true for sales.
It is true for operations.
And now it is true for AI-mediated work.
A company that hides those interactions inside private windows is not building an AI-native culture. It is building a thousand tiny private apprenticeships with no school.
This is the contrarian read on the moment: the first serious organisational fight in AI was never going to be about which model wins.
It was always going to be about legibility.
Can managers see how work is being delegated?
Can peers improve each other's prompting and scoping?
Can domain knowledge move faster than the org chart?
Can an agent's actions be inspected in context instead of after the damage?
Most of the current tooling stack is much better at personal acceleration than organisational legibility.
That is fine if your goal is local productivity. It is a problem if your goal is company-wide capability.
This is why the River thread matters more than yet another benchmark chart. Benchmarks flatter labs. Operating constraints reveal how firms are actually going to absorb the technology.
And the firms that get this wrong will create a weird kind of brittleness.
They will look fast on paper because individuals are shipping more. They will still be fragile in practice because nobody has shared visibility into how those gains are produced.
That brittleness will show up everywhere.
In onboarding, because new people cannot inherit the real playbook.
In compliance, because actions happen without clear shared context.
In quality, because prompt patterns and review standards remain private folklore.
In succession, because when a high performer leaves, their AI workflow leaves with them.
In management, because leaders mistake tool access for capability development.
This is where a lot of executive AI strategy still looks embarrassingly shallow.
Buying licences is not the same as building a learning system.
The smarter reading of Tobi's post is not “Slack is better than chat”.
It is that Shopify is trying to preserve a visible craft culture while radically increasing machine participation in the work.
That is much harder than it sounds.
Most companies treat AI adoption as a procurement exercise with a comms layer attached. Roll out a tool. Encourage experimentation. Collect a few wins. Write a slide about productivity. Repeat.
Shopify is doing something much more opinionated.
First came the memo: reflexive AI usage is now a baseline expectation. Not a nice-to-have. Not a pilot. Not an innovation theatre side quest. A baseline expectation.
Now comes the next layer: if AI usage is fundamental, then the way people use it should itself be part of the company's learning environment.
That is the real move.
The phrase Tobi used, Lehrwerkstatt, matters here. A teaching workshop. A shop floor that teaches by making the work visible. Not through a training deck. Not through a quarterly offsite. Through proximity to the actual craft.
That is old-school management in the best sense. It assumes competence is not mainly transferred by policy. It is transferred by example, repetition and observation.
AI threatens that if it stays private. River is an attempt to stop the loss.
Will it work perfectly? Of course not.
Public channels create performance anxiety.
Not every task should be fully open.
Some work does need confidentiality.
Some people will self-censor.
Some experimentation will move more slowly in public.
Fine. Those are real trade-offs.
But the opposite trade-off is not neutral. The opposite trade-off is a company quietly losing its ability to teach itself.
That is worse.
The timing is useful. While operators were arguing over River, OpenAI was out talking about its Deployment Company: a formal attempt to help businesses actually build around intelligence rather than merely admire demos from a distance.
That matters because it points to the same underlying shift.
The market is moving beyond “how smart is the model?” and into “how do we make this usable inside real organisations?”
That problem is not solved by model IQ alone.
It needs deployment patterns.
It needs permissioning.
It needs review loops.
It needs accountability.
And, increasingly, it needs some answer to the visibility question.
Because once agents move from chat toy to production coworker, the company has to manage them like a serious operating component. Not in the sci-fi sense. In the very boring sense.
Who can invoke them?
Where do they work?
What do they touch?
Who reviews the output?
What is searchable later?
What gets learned once and reused by everyone else?
Those are management questions.
The AI market spent two years pretending the main bottleneck was model quality. In plenty of domains, it no longer is. The bottleneck is whether the organisation around the model is designed to learn, supervise and compound.
That is a much less glamorous story than AGI discourse. It is also the story that will decide who actually captures value.
This is the part many AI vendors do not want to hear.
Private windows are perfect for adoption curves and imperfect for institutional memory.
They feel magical because they create immediate personal payoff. Ask a question, get a result, move on. The loop is clean and satisfying. Product people love this because it is easy to sell. Users love it because it reduces friction.
But companies are not just collections of private wins. Or at least they should not be.
A company is a machine for turning repeated work into shared capability.
If the AI layer helps people individually but prevents the organisation from observing, refining and standardising the best patterns, then the company is outsourcing its own learning to whatever employees happen to remember.
That is not strategy. That is vibes with a software budget.
The deeper point is that AI is not just a productivity layer. It is becoming a layer of organisational behaviour. It changes how people ask for help, how they frame problems, how they review work, how they document decisions and how they escalate uncertainty.
If all of that remains locked in one-to-one interfaces, the company gets the outputs but not the operating wisdom.
That is a terrible trade for any business that claims to care about operating force.
Do not misread this as “ban private AI”.
That would be stupid.
The point is to stop treating the private window as the only natural interface.
Some work should stay private.
Some work should be public.
The real question is whether your organisation has made that boundary a conscious design choice.
A few blunt rules would already put most firms ahead of the pack.
Use public-by-default AI channels for repeatable operational work.
Keep agent interactions searchable where possible.
Make strong examples easy to discover and reuse.
Treat prompt patterns and task framing as teachable craft, not personal magic.
Review not just outputs, but how the agent was scoped and supervised.
Decide explicitly which workflows must stay private and why.
That is the beginning of an AI operating model, not just an AI procurement model.
The companies that figure this out early will build something far more valuable than a tool habit. They will build shared machine literacy.
That becomes a cultural moat.
Not because nobody else can buy the same model. Because most firms will keep using the same model badly, privately and invisibly.
Once companies accept that private AI windows are not enough, the next debate will arrive immediately.
If work with agents becomes public by default inside the firm, who governs the public layer?
Who decides what agents can do in a shared channel?
Who owns audit, rollback and approval?
Who decides when visibility helps and when it becomes noise?
How do you stop a public agent from becoming a public nuisance?
Those are good problems to have.
At least they are real problems of adoption in production, not fantasy arguments about whether anybody should be using AI at all.
That fight is already over in serious companies.
The live fight now is whether they will build AI usage as a visible craft or tolerate it as a private habit.
One of those compounds. The other just looks busy.
A new Tobi Lutke thread lit up operator discussion this morning because it offered rare concrete numbers and a sharp design opinion at the same time: Shopify is not just using AI heavily, it is forcing its internal agent into public channels so the company can learn in the open. That cuts directly against the default private-window design of most mainstream AI products, which is why the debate has heat.
The operating test is deliberately plain. In the next 90 days, the serious teams will know which part of this shift changes revenue, cost, risk or customer control. They will assign an owner, define the workflow, measure the before-and-after state, and decide what should become permanent. The unserious teams will collect examples, forward threads, buy tools, and still leave the underlying system unchanged. That difference sounds small until the market moves. Then it becomes the gap between a company that learned and a company that merely watched.
That is the useful discipline here: separate novelty from control. Novelty gets attention, but control decides who captures the economics. Once a new interface becomes normal, the winners are rarely the people who merely noticed it first. They are the people who rebuilt the surrounding system before everybody else accepted that the old one had already stopped working.
The management problem is not that employees use AI. The management problem is that most companies cannot see where work now happens. Important judgement is being moved into private model sessions, browser tabs, copied documents and personal toolchains. Some of that is useful. Some of it is reckless. Almost none of it is governed well.
That creates a new split between companies with visible operating systems and companies running on hidden improvisation. The first group will turn AI use into shared workflows, reusable prompts, permissioned data access, review thresholds, audit trails and measurable outcomes. The second group will get a burst of apparent productivity followed by confusion over quality, provenance, responsibility and customer risk.
The fix is not to ban private AI use or pretend every action needs a committee. That would miss the point. The fix is to decide which work should become official, which tools are safe for which data, where human approval is required, and how good discoveries from individual workers become organisational capability instead of disappearing into chat history.
This is where the AI employee story becomes infrastructure rather than metaphor. Companies need shared context, policy, monitoring and rollback. They need ways to distinguish a harmless draft from a customer-impacting action. They need managers who understand workflow design, not just software procurement. And they need to accept that the old boundary between productivity tooling and operational systems is breaking.
The companies that handle this well will look calmer, not flashier. Their AI use will be less theatrical because it will be embedded into ordinary work. The companies that handle it badly will produce more activity and less trust. That is a brutal trade, because once trust is gone, every automation becomes suspicious and every output needs another human check.
