Jack Flintoft

If AI agents replacing service businesses created billions of dollars in outcomes for the economy, autonomous robots replacing service-based businesses will create an order of magnitude more. Intelligent service sectors are sized in the billions whereas physical service sectors are sized in the trillions (!).

In the last few years, we’ve seen billions of dollars in outcomes come from AI-native service companies. Importantly, these aren’t SaaS-like tools used by an existing team, but rather they’re third-party firms that repackage themselves as the service itself, delivering an outcome faster and cheaper than a human firm. (For those unfamiliar with the AI native model, Sequoia makes it quite digestible here).

As I highlighted in my recent post about AI-tokenomics, companies like Sierra and Decagon replace the outsourced customer support team through autonomous agents that do the work. If you’re the clothing brand GAP, rather than paying millions for an outsourced customer support call center to manage customer disputes, you and plenty of other brands now hand that budget to AI-native companies like Sierra, who deliver the same outcome (of a closed customer support ticket) cheaper using their AI agents.

I want to suggest that what happened with AI-native services over the past few years is happening to robotic-native services today.

Companies like Terrafirma aren’t just selling autonomous construction machinery to subcontractors, but are actually becoming the subcontractors themselves — winning on the industry’s existing bid sheets as they’re faster, cheaper, and safer than the subcontractors they’re replacing. Durin is a similar story. Rather than selling autonomous drilling rigs, Durin operates as a subcontracted mineral discovery firm — selling on a services-based pricing model.

Both currently retrofit decades-old machinery with autonomy stacks (“retrofit kits”), which let Terrafirma, Durin, and many others operate these rigs via teleoperation (humans controlling the machines virtually with a joystick) and, increasingly, run them fully autonomously, with no human in the loop at all.

Keen readers may be noticing where this is heading … the same human-in-the-loop to autonomy story played out with Sierra’s AI agents (and other similar AI native service companies). Over time, they needed less human intervention as they did more jobs, studying user context and reasoning traces along the way (Sierra’s AI agents now handle 70-90% of support tickets fully autonomously, up from just a year ago). The economic outcome of this was much better margins than their human service based equivalents. We can analogize this to service robotics companies today who harvest similar data on the edge. The ratio of autonomous machines to humans overseeing them is trending higher over time, because these companies can retrain their existing autonomy stack (the robotics version of agent RL) on (a) teleoperation and egocentric data from human operators “stepping in”, and (b) completed jobs with real customer feedback.

And importantly, this robotics data is harder to come by than the equivalent in AI-native services (which is exactly what makes it valuable and compelling!). Agents are afforded the privilege of training on troves of web data and synthetic RL environments, but nothing like that exists for the physical domain^[1]. You’ve really just gotta ‘drill baby drill’ and harvest it on the edge.

Durin is quite literally the physical manifestation of this, where every hole it drills teaches its automation stack something foundational robotics labs don’t have access to (what physical movements result in more success). This is why I think vertical applications will really shine as they stay immune from the model and hardware commoditization happening upstream. If Chinese open-source labs in the LLM era tell us anything, it’s that open source models keep pace with, and sometimes even beat, the US frontier. My hunch is that world models go the same way. Couple this with falling hardware costs and we get a picture that these service-based robotics companies will buy their inputs cheaper and cheaper each year (as was the case with AI native services). It seems then that defensibility sits less in the tech stack itself, and more in the app layer above it.

So where should we look then? Well, in the near term, lights-out physical labor is pretty unlikely … autonomy-assisted labor, with a human in the loop, is the more compelling bet right now, which is why retrofittable industries will be compelling in the short term (mining, construction, etc. where great machines already exist). The more ambitious companies, though, sit further out, with services in net-new spaces where the work was unfeasible before. These command new autonomous machines which are non-retrofittable. Deep-sea mining. Nuclear waste removal. De-orbiting of data centers. The possibilities for new services-based labor are close to infinite, and we can only imagine the kinds of transformative businesses and sectors that spring up in this new white space …

For anyone tackling a business model like this, or for those just curious, please reach out!

1. I want to stress that world models could be the silver bullet here to the data problem, but that right now there is no consensus answer (JEPA’s divergent thesis from the VLA researchers’ show us that even the experts disagree!). I worry therefore about getting to generalized intelligence in the near term that allows for general models to one-shot vertical / niche applications. New research could change this, but even with perfect world models, there needs to be companies that go out and capture the producer surplus from actually deploying these robots (since world-model companies will struggle to fully capture this alone). ↩︎