We're used to thinking of AI as a tool that helps us work better. Something more interesting is happening. AI is on the verge of doing the whole job. Think of it like the evolution of self-driving cars. Today's AI is a very smart cruise control. Tomorrow's is the car that drives itself.
This isn't a new story. We are about to see in knowledge work what already happened in agriculture.
Farm jobs as a share of all US jobs, 1790–2000
Table
| Year | Farm jobs, % of all jobs |
|---|---|
| 1790 | 90% |
| 1800 | 83% |
| 1810 | 81% |
| 1820 | 79% |
| 1830 | 71% |
| 1840 | 69% |
| 1850 | 64% |
| 1860 | 59% |
| 1870 | 53% |
| 1880 | 49% |
| 1890 | 43% |
| 1900 | 40% |
| 1910 | 31% |
| 1920 | 27% |
| 1930 | 21% |
| 1940 | 17% |
| 1950 | 12% |
| 1960 | 8.3% |
| 1970 | 4.6% |
| 1980 | 3.4% |
| 1990 | 2.6% |
| 2000 | 2.6% |
In 1900, about 40% of American jobs were on farms. By 2000 it was 2.6%.
Two things happened at once. The land consolidated into larger operations, and the acreage a single farm works multiplied. Cropland on farms of 2,000 acres or more went from 15% of the US total in 1987 to 36% in 2012. Over the same twenty-five years, the midpoint acreage for major field crops roughly doubled or tripled depending on the crop: corn went from 200 acres to 630, wheat from 400 to 1,000.
Cropland has shifted to larger farms
Share of US cropland by farm size, 1987–2012
Table
| Farm size | 1987 | 1992 | 1997 | 2002 | 2007 | 2012 |
|---|---|---|---|---|---|---|
| Under 100 acres | 8% | 7% | 6% | 6% | 5% | 5% |
| 100 to 249 | 17% | 15% | 13% | 12% | 10% | 9% |
| 250 to 499 | 21% | 19% | 17% | 15% | 13% | 12% |
| 500 to 999 | 23% | 22% | 21% | 19% | 18% | 16% |
| 1,000 to 1,999 | 16% | 18% | 19% | 20% | 22% | 22% |
| 2,000 acres or more | 15% | 19% | 24% | 28% | 32% | 36% |
Midpoint acreage per farm, 1987 vs 2012
Half of all harvested acres are on farms bigger than the midpoint.
Table
| Crop | 1987 | 2012 | Change |
|---|---|---|---|
| Wheat | 400 | 1,000 | +600 |
| Cotton | 450 | 970 | +520 |
| Rice | 290 | 800 | +510 |
| Corn | 200 | 630 | +430 |
| Soybeans | 240 | 560 | +320 |
Fewer people, working much more land each, producing far more. That is the shape of the thing.
The transformation of knowledge work will look like the transformation of agriculture on a compressed timeline. Within a few years we'll see companies where AI agents outnumber human employees a hundred to one. These will be companies built from the ground up to be run by AI, with humans as managers and directors rather than doers.
Why Now?
The autonomous company has been the promise of AI for decades. So why now? A few breakthroughs made it possible.
Large Language Models
Starting in earnest around GPT-2 and GPT-3, large language models are the core technology behind the AI tools we use today. They're the result of better algorithms, vast amounts of data, and powerful GPUs converging at once. These models are so large and power hungry that AI data centers are expected to account for 8 to 9% of US power consumption by 2030.
AI labs have condensed the internet's knowledge into a single model that can understand and respond in natural language. What that unlocks is the ability for computers to interface with the world in a human way. Before LLMs, software needed a structured API to talk to other software. Now AI can communicate through the same channels we do: email, text, fax, and phone calls.
Since the 1950s, computer scientists have been trying to build a computer that could convince a human they were talking to another human. They called it the Turing Test. For all practical purposes, the Turing Test for text has been passed. When you talk to ChatGPT you feel like you're talking to a person. Everything we're doing today is only possible because of that.
Voice AI
Voice AI builds on LLMs and adds spoken conversation. It required solving several problems at once.
- Text to speech. A voice that sounds natural and not robotic.
- Speech to text. Accurate, fast transcription.
- Latency. Responding quickly enough to keep a conversation flowing.
- Turn detection. Knowing when the user is done talking.
We're close to passing the Turing Test for voice, and that unlocks another fundamental mode of human communication for AI.
Generative Video
Generative video is the next step, and it makes a video call with an agent possible. That takes a convincing voice plus a realistic avatar with synchronized lips. The hard parts were generating video fast enough and getting the lips to match the words.
The implications are significant. Picture our agents with their own boxes in a Zoom call, talked to like any other coworker. Patients and caregivers on a video call with an agent. It matters for marketing too, where synthetic video means quality content at scale.
Generative video splits into realtime and offline. Realtime is generated on the fly for a live call. Offline tends to be better quality because there's time to work, coming back with a finished video in five to fifteen minutes.
Agentic AI
Agentic AI is what gives LLMs the ability to reason and plan. It's the scaffolding that lets several modes work toward a goal, figuring out what steps need doing in what order and using tools to accomplish them. Over time, the complexity and duration of the tasks that can be handled will keep growing. I wrote about how to grade these systems in Levels of Agentic Behavior.
Vision Language Models
Vision language models give AI the ability to see. The immediate use is reasoning about computer screens, and the long-term implications are much broader. When AI can see, it can use software tools the same way we do.
The Autonomous Future
What this paints is a company that can coordinate and administer care for a huge number of patients, at a quality no non-autonomous company can match. It lets us expand into many service lines and stay fully integrated. The operational complexity that made that impossible becomes manageable.
We'll be able to build a company that's nimble, one that can enter a new business line or absorb a rule change easily. Retraining agents is orders of magnitude easier than retraining thousands of employees.
A core consequence is that AI compresses time. Things get done ten times faster. It's the difference between picking crops by hand and driving a combine. As we build, that requires an almost unnatural mindset: for every process, ask how we do this ten times faster. This will be hard. We have never experienced a productivity jump of this size in our working lives, so the instinct toward incremental improvement will be strong. We have to fight it.
This means a team of no more than a hundred people managing healthcare at home for millions of seniors across the country. We'll be like Uber's headquarters, but for healthcare. We won't provide the hands-on care, that stays with caregivers, nurses and other professionals, but we will coordinate all of it with a very small team.
The Reality Check
You might be thinking this is all well and dandy, but didn't MIT just say 95% of agentic AI deployments fail? You're right to be skeptical. They do fail, and not for the reason you'd think. It isn't that the AI isn't smart enough. It's that companies are chaotic and undocumented.
In most companies knowledge isn't written down, it lives in people's heads. That's why a new hire takes two or three months to become productive. They have to learn the unwritten rules.
- Why did we make this decision?
- Who made it?
- What's the goal of the team we're collaborating with?
- How should I do X?
- Am I allowed to do Y?
There's a concept in software engineering called the bus factor: how many people on a team would have to get hit by a bus before the team could no longer operate. I saw that play out literally at Microsoft. Don't worry, nobody died. After the Bing annual all-hands at a convention center in Seattle, the entire Bing team walked back to the office and crossed a street downtown at the same time. As we're crossing, my boss says, "see guys, this is why bus factor matters."
When I say companies aren't well documented, I don't mean they're badly run. Before this wave of AI, the cost of documenting everything and keeping it current was simply too high. The overhead didn't justify the benefit. Imagine going back into every document after every meeting to update whatever just went stale. That used to take hours. Today an AI does it from the meeting notes.
AI changes the calculus. It brings the cost of documentation down and raises the benefit. With everything written down, an agent has the context to do its job correctly. Your prompts get shorter because you aren't constantly re-supplying background. The agent behaves like an employee who has been here for months rather than one who started today.
What AI Agents Need to Be Productive
Three things.
1. Knowledge and Context
All the background from meetings, conversations, support tickets, pull requests, and decisions. AI has a real advantage here. The moment something is documented, every agent knows it. There's no bandwidth limit and no delay, which is the opposite of humans, who need sync meetings and knowledge transfer sessions.
2. Standard Operating Procedures
The rules of the game at this company. How does it actually operate? What are the constraints and procedures that define the way we work? Every company runs differently, and the SOPs are the written version of ours.
3. Actuation
Context tells the agent what's true and the SOPs tell it what to do. Then it has to be able to act.
- Integrations with our tools to pull information
- Scripts or integrations that push data: text a caregiver, submit a payroll run, correct a timesheet in the scheduling system
- Computer vision, so it can drive software the same way a person does
Historically software could only act through APIs. Today's AI can use the software itself.
This documentation-heavy style suits a remote company. In an in-person company a lot of communication happens in hallways and on whiteboards and is never captured. Remote, everything runs through a tool, which makes it far easier to capture for humans and agents alike.
Building It From Day One
An autonomous company is a different kind of company. The interplay between human and AI employees means rethinking the organization from first principles. For an existing company, retroactively documenting its context and SOPs is a massive undertaking, with a huge coordination problem: writing everything down while still running the business. It's rebuilding a car while driving it.
We don't have that baggage. We get to build with autonomy in mind from the start, where the context is written down, the SOPs are defined and maintained, and the actions can be taken by an agent.
Look at electric vehicles. The best ones come from Tesla, Rivian, and Lucid, companies that built the car as an EV from the ground up instead of dropping an electric motor into an existing platform. The result is a fundamentally better product that actually uses what the new technology is good at.
That's our advantage. We're an engineering company building for healthcare, not a healthcare company trying to figure out AI. The product is the service we provide and the autonomous organization we build to provide it. One that can serve millions of patients with a hundred times fewer employees.
The Frontier
AI isn't fully there yet. As I covered in Levels of Agentic Behavior, we're not at a stage where we can trust it with every task. For now humans fill the roles AI can't. We should always be building for the future where it can.
We'll make progress on two fronts.
Tasks moving from humans to AI
Individual tasks start with a human and slowly move over. Today the only way to get a job done might be for a person to do it. Next year we'll likely find agents can do most of them.
Take billing or payroll. Today those are human tasks because they're sensitive. Soon AI will make fewer errors than we do, and it will be reckless not to have those functions run by AI. For the analogy, imagine calculating payroll by hand with no computer. Once you have a spreadsheet, doing it by hand is the reckless option.
Business lines becoming feasible
Entire business lines become automatable as agents get reliable. There are lines today where we have no advantage, because AI wouldn't give us a big enough lift.
Consider expanding into external caregivers. We'd become a two-sided marketplace that has to recruit caregivers as well as find patients, and then manage the logistics of getting the right caregiver to the right patient, accounting for last-minute cancellations, staffing gaps, sudden increases in patients, and complex scheduling.
That's an enormous coordination problem. Tackling it today would take a huge number of employees, and we'd look like any other home care agency. In a year or two, AI will be able to run functions like recruiting or scheduling, which lets us enter and automate entirely new lines.
There are dozens of business lines to tackle after family caregiver compensation. We need to be ready the moment the technology is.
Building for what's coming
The simple question to ask is this: if GPT-6 comes out tomorrow, will I be excited or upset? We should be planning against a world of increased capability. Assume costs come down 10x next year and capability goes up. The work we do today should not be work we throw away when the next generation ships.
And the real product is the autonomous company we're building. Whatever we do manually today should be set up so that if today's AI were slightly smarter, it could take over.
The Right Time
There's an analogy to Google. A big part of why Google succeeded is that it was founded at exactly the right moment. A few years earlier and the online advertising market wouldn't have been mature enough to support them, because there weren't enough people online. A few years later and the internet would have been too big to index affordably.
We're in a similar position. If we'd started this company even a year earlier, we'd be building a more traditional home care agency, because the technology wasn't good enough to do much of the work autonomously. If we wait a couple of years, the market will be full of autonomous healthcare companies with years of practice.
We're starting at the moment where the technology is mature enough to build something real and the field is still open enough to lead it.
This is the most exciting time in history to be building a technology company that provides a real-world service. We'll look back on the pre-AI era of knowledge work the way we look at farmers picking crops by hand. What it asks of us is a constant mindset of automation and compressing time.