What I Mean by AI-Native

“AI-native” has been diluted to the point where it means almost nothing. A company that uses ChatGPT to write email copy is not AI-native. Neither is one with a chatbot on the website, or a team that summarizes meeting notes with AI, or a fund that runs a ChatGPT prompt against its deal memos.

Those are AI-assisted operations. The humans are still doing the work. The org chart looks normal. Headcount is still the primary operating cost. AI shows up as a productivity multiplier on individual outputs — useful, often genuinely valuable, not structural.

AI-native is a different claim. It’s a claim about how the business is designed: agents handle the recurring operating layer, humans concentrate on the decisions that require actual judgment. The org chart is built around what agents can sustain. The unit economics are different from inception.

I’ve now built enough of these systems — across Uplift Capital, Pyrennial Farms, and a small portfolio of operating companies — that the distinction has become concrete rather than theoretical. Here’s where I’ve landed.

What agents handle well

The pattern across every system I’ve built is the same: agents work well when the problem is take structured inputs, apply deterministic logic, produce structured outputs — on a schedule, without forgetting, without variation.

Uplift OS ingests the GL detail exports from six properties, normalizes them across five property managers’ chart-of-accounts conventions, runs fee-anomaly detection against the management agreements, and produces the unit-economics dashboard I make decisions from. No analyst touches it between the raw export landing and the output appearing in the dashboard. The system runs. I review what it surfaces. I decide what to do about what it surfaces.

The bank screener pulls from five public databases on a quarterly cycle, applies a Graham/Dodd/Buffett-weighted scoring model across 4,000+ community banks, and produces a ranked shortlist of the forty or fifty names where the quantitative case is strong enough that qualitative diligence matters. It runs deterministically. The AI that pair-programmed it is not involved at runtime.

The grazing simulation runs every morning at Pyrennial Farms, pulls live weather data, models forage growth and soil moisture across four paddocks seven days forward, and tells me where to move cattle. Built with AI assistance. Runs without AI.

The common structure: AI was the contractor that built the machine. The machine runs without the contractor.

This framing — AI builds the infrastructure, the infrastructure runs the operation, the human operates at the top of the value stack — is what I mean by AI-native. Not AI doing the work. AI building the deterministic system that does the work.

What agents can’t do

This is where most AI-for-business content either overclaims or underclaims. Overclaiming: agents can do everything, unlimited leverage, replace the whole team. Underclaiming: agents are just autocomplete, none of this is real.

Neither is right. In my experience, agents break down reliably in four categories.

Relationship work. Investor relations, LP communications, seller negotiations, lender relationships — these are trust-dependent interactions where the quality of the output is inseparable from who is doing it and what they’ve actually built. I draft with AI assistance. The words that go out are mine. The relationship is mine. An LP who gets an email from “the AI” is not actually getting a letter from me, and they know the difference.

Judgment under genuine uncertainty. The bank screener surfaces which names look interesting. It cannot tell me whether a specific CEO’s capital allocation history reflects the kind of owner-operator discipline that produces long-run value. That requires reading proxy statements, listening to earnings calls, sometimes talking to people who know the person. The system defines what I should be asking. It can’t do the asking.

Material, irreversible decisions. Pulling the trigger on an acquisition, walking away from a deal that looks good on paper but feels structurally wrong, deciding to replace a property manager — these require bearing actual accountability for the outcome. Agents don’t bear accountability. Humans do. The accountability is not separable from the decision.

First-principles translation. Building Uplift OS required me to articulate what I believed about multifamily fund operations precisely enough that a machine could implement it. The bank screener required expressing a century of value investing theory in terms precise enough to operationalize. That work — translating judgment into systematic rules — is the hardest part of every build, and it cannot be delegated. The AI builds what you can specify. The specification is yours.

What this means when you’re building something new

The most useful application of this framework is not retrofitting it onto existing operations. Those retrofits are still worth doing — Uplift OS is a retrofit on a fund that was already operating — but the structural economics are better when the design is native.

When you’re building a new company and you ask honestly which workflows agents will handle and which require human judgment, the org chart looks different from the start. You don’t hire an operations manager to run reports that a pipeline could run. You don’t staff a research function to pull data that an ETL script can pull. You build the systems first, staff humans into the roles where their judgment is the actual product, and let the cost structure reflect that design.

The economics follow from the design. A business designed this way runs on a fraction of the fixed headcount a traditional version of the same business would require. The operating leverage is real because the leverage is structural — not productivity on top of headcount, but headcount concentrated at the top of the value stack where it belongs.

This is what I mean when I describe the operating companies I’m building as AI-native. Not that they use AI tools. That the business design starts with the question: what does the recurring operating layer look like if agents are handling it? The humans fill in the roles that genuinely can’t be filled any other way — which, on a well-designed system, turns out to be a much shorter list than the traditional org chart assumes.

The constraint isn’t AI capability. It’s the discipline to build the systems before you hire the people, and the honesty to know which judgment calls actually require a human in the seat.