Two People Could Replicate Your Best Line in 90 Days

Here is the question to ask about your own company before anyone else asks it for you: is there a high-margin line of your business that two people with off-the-shelf AI tools could rebuild in 60 to 90 days? Not the whole company. One line. The profitable one. The one that funds everything else and that nobody on the leadership team worries about because it has always just worked.

If the answer is yes, the honest follow-up is uncomfortable: someone, somewhere, has already started.

This framing comes from Peter H. Diamandis’s conversation on the organizational singularity, where he puts it bluntly — “a high margin line of your business that two guys with off-the-shelf tools could replicate in 60 to 90 days.” It is a good provocation. But a provocation is not a forecast, so it is worth pulling apart with actual numbers, the cases where it held, and the cases where it broke.

The mechanism is real, and it is not magic #

The reason two people can now do what used to take a department is not that the AI is brilliant. It is that the marginal cost of adding a “function” to a tiny team has collapsed. A World Economic Forum piece from August 2024 argued that AI brings what it called corporate-scale capability to small teams — covering legal, financial, marketing, and customer-service work that previously demanded dedicated headcount (WEF, 2024). I could not independently re-verify the exact phrasing — the article returned an access error on fetch, so treat the specific wording as paraphrase rather than quote — but the underlying economics are visible in harder data.

The hardest data point is a controlled field experiment. Harang Ju and Sinan Aral at MIT ran a study with 2,234 participants and found that human-AI teams produced roughly 50% more output per worker than human-only teams on creative production tasks; workers delegated about 17% more of their tasks to AI agents and cut their own direct execution effort by 62% (arXiv 2503.18238). Read that last number again. Execution effort down 62%. When execution drops that far, the binding constraint on a small team stops being labor and starts being judgment — which is exactly the regime where two people can punch far above their headcount.

The case that actually happened: Cursor vs. Copilot #

The cleanest real-world instance is Cursor, the AI-native code editor built by Anysphere. GitHub Copilot launched in 2021 with Microsoft’s full weight behind it — distribution through the world’s largest code host, an enterprise sales org, Azure infrastructure, a large engineering team. Cursor, founded in 2022, shipped its IDE in 2023 with a fraction of the people.

By the end of 2024, Cursor had gone from $1M to $100M ARR in twelve months — reaching that milestone faster than Wiz, Deel, or Ramp, each of which had hundreds of employees at the same revenue stage (Sacra). Sacra continues to track the trajectory past $1B ARR into 2025. The widely reported color around the company — a team of roughly twenty people, effectively zero marketing spend, a freemium conversion rate in the mid-30s — is repeated across coverage but does not appear in the Sacra research itself, so I am flagging it as reported-elsewhere rather than confirmed by that source.

What matters is the mechanism, and on the mechanism the sources agree. Cursor did not win on a secret model; both products call the same frontier APIs. It won on architecture. Developer-productivity tooling is a high-margin line that needs no proprietary-data moat, no regulatory approval, and no field sales force to reach its first hundred million in revenue. That is precisely the kind of clean-slate target a tiny, AI-native team can take (Sacra; AI Strategy Decoded). If your high-margin line shares those properties — software-deliverable, no compliance gate, sold without a human — you should assume it is reproducible.

The honest counter-case #

Now the part the provocation glosses over, because it is the part that decides whether you actually need to panic.

Task speed is not workflow output. This is the single most important caveat. A longitudinal study of 400 companies by DX found that a 65% increase in AI-tool usage translated into only a 7.76% increase in pull-request throughput (DX). And in a randomized controlled trial, METR found that experienced open-source developers were 19% slower on real tasks when allowed to use AI tools like Cursor Pro — even though they believed they were about 20% faster (METR; arXiv 2507.09089). The reason is mundane: when the bottleneck is planning, alignment, and context-switching rather than typing code, AI adds coordination overhead instead of removing it. “Replace fifty tasks” and “rebuild a business line” are different claims, and the gap between them is where most of the disappointment lives.

The replicable part is the routine part. Klarna is the cautionary tale. In February 2024 its AI assistant handled 2.3 million chats in a month — the stated workload of 700 full-time agents — for an estimated $40M in annual savings (Twig). Then, by 2025, the company reintroduced human capacity for complex cases, because quality degraded on the roughly 5% of conversations that were emotionally or commercially fraught. Two people with AI tools can absolutely cover the routine margin of a support operation. They cannot yet cover the exception-handling that protects the brand when a customer is angry, the case is compliance-sensitive, and the wrong answer is expensive.

Some moats are genuinely hard to copy in 90 days. McKinsey’s framing of AI-era advantage — durable moats like network effects, proprietary datasets, and embedded compliance versus non-durable ones like workflow efficiency and application-layer lock-in — is consistent with everything above, though I’ll caveat that the page was not directly reachable to confirm the exact taxonomy (McKinsey QuantumBlack). A paywalled Morningstar analysis reportedly makes the complementary point that incumbents in regulated industries already hold capital buffers, legal infrastructure, and governance that an AI-native entrant must build from scratch (Morningstar, paywalled) — I could not open it, so treat it as secondhand. The signal across both: if your margin rests on workflow convenience and habit, you are exposed; if it rests on data, regulation, or trust relationships, the 90-day clock does not start.

Building an MVP is not building a business. The 90-day window is real for standing up software that handles the core case. It is not real for accumulating the customer relationships, operational resilience, and trust that make a line durable. Most small AI teams do not survive; long-term viability runs in the rough range of one in ten to one in five. And the competition is not only the legacy incumbent — Cursor and Copilot are both compounding, which means the disruption window can slam shut faster than the framing implies once the first mover takes the beachhead.

What to do with this #

The provocation earns its keep as a triage tool, not a prophecy. Walk your P&L and sort your high-margin lines into two piles. Pile one: software-deliverable, no compliance gate, sold without a human, margin resting on convenience. Assume a two-person team is already prototyping it; your defense is to out-learn them, not to out-staff them. Pile two: gated by regulation, proprietary data, or hard-won trust, where the failure cost of the 5% edge case is severe. Those have real runway — but only until someone captures the data or clears the regulatory path, at which point they migrate to pile one.

The number that should stay with you is not 73x or $100M. It is 62% — the measured drop in human execution effort when AI agents enter the loop (arXiv 2503.18238). That is the structural shift underneath the whole thesis. It does not mean two people can replicate everything. It means the things they can replicate just got a lot cheaper to attempt, and the only question left is whether they are pointed at your best line yet.