A German startup called Microagi just raised $55 million to clean apartments in New York. The cleaning is free. The condition is that everything gets recorded. The footage becomes training data, some of it licensed to frontier AI labs, some of it kept in-house to improve the company's own robot brains. It was, per the newsletter that surfaced it, the largest seed round in German history.

I keep coming back to that trade. Free labor in exchange for a consented data pipeline. The customer gets a clean apartment. The company gets a proprietary sensor network pointed at a physical world that no open dataset covers. It sounds like a stunt. It is actually a business model, and it is aimed at the part of the AI stack that most marketing leaders have not yet noticed is where the fight has moved.

While Microagi was quietly instrumenting kitchens in Brooklyn, Moonshot released Kimi K3 as an open-weight model, reportedly distilled from the closed frontier. Running the top version still needs a serious GPU cluster, so this is not the week Nvidia gets displaced. But the direction is clear enough. Model quality is drifting toward table stakes. Anyone with capital and compute can rent a very good brain by dinner. That is the ceiling, and the ceiling keeps dropping.

Which is why the Microagi story is not really a robotics story. It is a preview of the next competitive shape of AI in business. When the model is a commodity, the scarce thing is the interface that meets the customer first, and the consented data loop that makes that interface better every time it is used.

Call it the distribution default.

What a distribution default actually is

Here is what I mean, in a customer journey rather than a slide. A shopper asks an agent to reorder laundry detergent. The default brand is not the one with the best benchmark score or the smartest chatbot. It is the one whose agent completes checkout in three steps, remembers the household size from the last order, learns the preferred delivery window, and cleanly hands off returns to a service surface that carries the same context. The customer picks the interface that finishes the job. The brand behind that interface gets the next order, and the order after that, and a proprietary log of what actually happened at the point of decision.

That log is the moat. Not the model. The log.

Because the log tells you which prompts led to abandonment, which substitutions were accepted, which reorder cadence tracks with lifetime value, which second question the shopper asked when the first answer was wrong. None of that is inside a foundation model. All of it is inside the surface that captured the intent and closed the loop.

Microagi is doing the physical-world version of this. Marketing teams are about to do the digital version. The mechanic is the same. Offer the customer something they actually want. Instrument the exchange with consent. Feed the resulting data back into the interface that produced it. Watch the interface get faster, more accurate, and more sticky than anything a competitor can rent from a model provider.

Why this reframes the AI budget conversation

Most marketing AI budgets I see are still weighted toward model access, tool subscriptions, and generic content acceleration. That made sense in 2024 when the question was whether the model could write a passable draft. It makes less sense now. If the model is a commodity, the marginal dollar spent on a better model returns less than the marginal dollar spent on the interface that captures customer intent and the pipeline that turns intent into context.

My read is that the next twelve months will separate marketing organizations into two groups. One group will keep evaluating models. The other group will start evaluating capture velocity, which is a plainer way of asking: how quickly does an interaction with our brand's agent become usable context for the next interaction? The gap between those groups will not look dramatic at first. It will look like one team's agent gets slightly better every week and the other team's agent stays exactly as smart as the underlying model.

Compounding does the rest.

I wrote a piece earlier about why agentic commerce needs channels, not just conversations, and the argument there is adjacent to this one. Channels carry commitment state across sessions. Distribution defaults extend that logic outward. The channel is how the agent remembers what it promised. The default is how the agent gets a chance to make the promise in the first place.

Consent is a speed tax on this model, not the story. Microagi's cleaners record the apartment because the customer agreed to the exchange. A marketing agent captures preferences because the customer traded them for a better outcome. The teams that will move fastest are the ones who design the consent layer as a product surface rather than a compliance afterthought, so that the ask is clean, the value is legible, and the customer opts in without friction. Get that right and consent stops being a brake. Get it wrong and the pipeline never fills.

That is a design problem. It is solvable. It is not the frontier.

What is actually within your control

There is a quiet discipline underneath all of this. Model benchmarks are not within your control. Open-weight release cadences are not within your control. Which lab distills which lab next quarter is not within your control. The interface your brand puts in front of a customer is entirely within your control. The consent design is within your control. The data pipeline behind the interface is within your control. What you do with the resulting context is within your control.

That is the useful list. Marketing leaders who spend their attention on that list will compound. Leaders who spend their attention on the leaderboard will look busy and stay flat.

What this makes possible

A few things become newly available once you stop treating the model as the product.

Brands can invest in agent interfaces that own the first action, not just the first answer. Reorders, renewals, returns, upgrades, service escalations, appointment changes. These are the moments where intent is highest and where a well-designed agent can finish the job in the same session it was raised. Owning that moment is the modern equivalent of owning shelf placement, and the shelf is now conversational.

Product teams can design consent-native capture into normal customer workflows so that context accumulates as a byproduct of usefulness rather than as a separate data project. Microagi's free cleaning is the extreme version. A subtle version is a support agent that asks one clarifying question and remembers the answer forever.

Go-to-market teams can start treating agent routing partnerships as distribution deals, not integrations. When a shopper's assistant of choice has to pick a merchant agent to complete a task, being the default is worth more than being available. That is where the real negotiation will happen over the next year, and most teams are not yet staffed to conduct it.

The uncomfortable part

None of this is the fun work. Instrumenting a customer journey for consented capture is less glamorous than fine-tuning. Negotiating for default placement inside someone else's assistant is less clean than launching your own. Redesigning the interface to actually finish the job takes longer than swapping a model behind an existing chatbot. The teams doing this work will not have the loudest AI announcements this year. They will have the quietest customer data by next year.

That asymmetry is the opportunity.

The race for the best model is a race almost no marketing organization can win, and increasingly does not need to. The race for the interface that becomes the default, and the pipeline that makes the interface smarter with every interaction, is a race any serious operator can enter today with the budget already in hand.

The question is not which model your team should adopt next quarter. The question is which customer action your brand's agent will be the default way to finish.

Sources and further reading