This essay is about the AI businesses that will eat the world, and how to build them.

the future of work

There is no question that a large majority of white collar work will be automated, what is at stake is how, and what the journey will look like getting there.

Join @xAI and help build a purely AI software company called Macrohard. It’s a tongue-in-cheek name, but the project is very real!

In principle, given that software companies like Microsoft do not themselves manufacture any physical hardware, it should be possible to simulate…

— Elon Musk (@elonmusk) August 22, 2025

Pretty much every big AI lab is building some version of Macrohard, with the goal of getting AI to replicate the roles that a human does. Currently, this fails in very non-obvious nebulous ways (even after you get past the context/memory bottleneck).[0]

In this essay, we’ll focus on verticalized products that can achieve traction and usability today.

the enumeration problem

I think, in general, it doesn’t make sense to build a product without talking to customer. One day you can’t just show up and say “we’re going to replace your entire workforce”, you have to build trust over time.

Even doing Macrohard is much more complicated than meets the eye, as every employee on every team interfaces with so much software and so many systems and in so many ways, the mere enumeration of an individual employee’s stack is near impossible.

Nobody, not even the employees themselves, can fully articulate what they do. You need to observe employees working over multiple days like a hawk. You need to know their job front to back and back to front.

This is perhaps the hardest part of starting a “vertical AI” (or any other) startup, figuring out what the customer (or their employees, more aptly) actually does on a discrete/atomic level. This is often a mix of proprietary information and proprietary systems that you effectively have to recreate in your own environment, and is different for every org in your market.[1]

A simple solution is to sell a small product and put a Palantir-style FDE/support eng in the org to gather intel for building your next product.

enduring through super-intelligence

Starting an AI startup really is a bet on the path to super-intelligence. Let’s be very precise when dealing with this notion of a singularity or of super-intelligence.

There probably won’t be some discontinuity in AI intelligence anytime soon, but it is also true we are on an exponential (fast takeoff). I think the distribution of exceptional humans is quite compact, meaning the right tail of human intelligence is very short compared to the right tail of possible intelligence.

The point being, the path to super-intelligence is continuous, therefore your company-building strategy must be continuous too.[2]

building for the excluded middle

So what does the continuous path to a super-intelligence powered enterprise look like? For this, we can look to the current killer use case for AI, software engineering.

In the beginning there were plugins that were a small layer on top of existing tools (GitHub Copilot), then there were existing tools forked with some tweaks (Cursor), then new tools (Claude Code), and finally pseudo-employees (Devin).

Though many enterprises will get software that skips some of these stages because of market inefficiencies. The incumbent software companies will maybe get to Copilot or Cursor, but building a new product like Claude Code or Devin will be extremely difficult for them.

Also, the models currently can’t really support Devin’s ideal usecase, some more innovations need to happen. But Claude Code works really well. In addition, because plugging into all the surfaces an employee does is really hard for any given industry, it makes sense that in order for Devin for X industry to become tractable and get useful data for it, Claude Code for X would have to be deployed first.

So my playbook for building enterprise software then, is to build a killer Claude Code for X tool (agents that do tasks) then after PMF, transition to a Devin type of model with what you’ve learned.

It is also much easier to get adoption on a low level with people whose jobs are at risk of being replaced if you sell them on a tool that helps them do their job easier instead of simply replacing them. This is in the same vein of the continuity I was talking about earlier.

the genealogy

The platonic ideal (that super-intelligence would eventually use) for this sort of thing would be a genealogy (or as the austere technological community would say, “system of record”) of all the data that comprises the business and the surface of all possible actions to be taken within the business.

This is the bridge between your Claude Code phase and your Devin phase. Every task your Claude Code shaped tool executes is a data point mapping the company’s business processes. Over time, you’re building the API the company has to itself.

This data is essentially what the business is, which is why Palantir calls it an “ontology”. The ideal AI implementation is predicated on prior deployment.

The core hallmark of a sticky AI native company is that it gets better as models improve, that it takes change to be constant. The business is built from day one with this assumption.

This is why similar reasoning doesn’t work with the incumbents, they are slow to change, and can only come up with very linear versions of product updates.

thesis/narrative engineering

Critically, at each stage, your narrative, your thesis that is the engine of company should not be tied to any specific product. It should be much grander. You sell tools to engineers/white collar workers. To VCs, execs, managers etc. you should sell your thesis.

A good way to test this distinction is, your customer should never have your product in a table with other software vendors as columns. They should never compare features. You shouldn’t be competing on having better features. You should be competing on flipping the paradigm on its head, or more concretely, selling a story, a future of the customer’s company.

Your thesis should essentially be this essay verticalized, how will you enable/underpin the continuous ascent to a super-intelligence run business inside your vertical.

As a founder, you must get good at this narrative engineering.

what if the models plateau?

I don’t think this will happen, but in any case, you are essentially 0-beta against the broader AI ecosystem. Because of the continuity of business strategy, you continuously are interfacing with reality and creating value.

At the Claude Code stage you have a tool business, at the Devin stage you have a staffing/automation business, at the genealogy stage you have the system of record.

Anywhere you land, you have a killer, de-risked business.

closing

So now we have a rough path for starting enduring enterprise AI businesses.

1. Come up with a specific view on the way certain companies should be run, something close to “Level 5 autonomy for X”
2. Figure out what the people do on a low level in those companies
3. Make Claude Code for them
4. Use the data you gather to make Devin for their managers
5. Build the genealogy and rails super-intelligence will use to fully automate the business end to end

Interface with reality (customers often), be continuous in growth and narrative, but not linear in product.

[0] Essay on frontier AI failure modes soon!

[1] There is an argument to be made for “AI native” businesses that are direct competitors to incumbents from day 1. The environment needed to properly evaluate AI agents in the long term approaches the totality of the business built around them. But it seems the models are not good enough to build a business like that yet; there is some difficulty in faithfully creating such systems to be functional.

[2] Even if you believe in a discontinuity, in that future with super-intelligence, the super-intelligence would aim to be efficient and use pre-existing rails as much as possible.