What six people can now attempt

The build started as a stress test on team size. The hypothesis was modest: a small group, using current agentic tooling, could ship a feature set that would have required a mid-size engineering team two years ago. The hypothesis was wrong in one direction.

The team was not just faster. It attempted a different class of problem.

The default reading is that AI made engineers more productive. The numbers support a stronger claim. Anthropic’s August 2025 internal study of 132 engineers and researchers found Claude was used for roughly 60 percent of their work, with engineers reporting about 50 percent more productivity than a year prior, and 27 percent of AI-assisted tasks rated as work that would not have been done at all otherwise (Anthropic, December 2025). At GitHub’s controlled study, developers completed coding tasks 55 percent faster with Copilot (arXiv, 2023, replicated in 2025 enterprise studies). Pull request cycle time fell from 9.6 days to 2.4 days in deployment studies (GitHub, 2025).

Pure speed is the small finding. The larger finding is scope. When the marginal cost of writing, reviewing, and refactoring code falls, the constraint on what a small team will attempt also falls. The work itself reshapes. At Anthropic, code design and architectural planning rose from 1 percent to 10 percent of engineer time, and implementing new features rose from 14 percent to 37 percent (Anthropic, December 2025). The team did not just do the same work faster. It moved up the stack.

The market evidence runs the same way. Cursor reached over 500 million dollars in annualised revenue by mid-2025 and crossed one billion in November 2025, on a team that was roughly 50 in early 2025 and grew through the year (SaaStr, 2025; jobsbyculture, 2026). Lovable crossed 400 million dollars in annual recurring revenue in February 2026 with 146 full-time employees, a ratio of 2.77 million dollars in ARR per employee that already exceeds Gartner’s 2030 unicorn benchmark of 2 million dollars per head (TechCrunch, March 2026). Midjourney reached around 500 million in revenue with roughly 100 to 160 staff, self-funded, no traditional marketing (Sacra, 2025; getlatka, 2025). Cognition’s revenue run rate grew from 37 million in May 2025 to 492 million by mid-2026 with a team of 49 at end of 2025 (getlatka, 2025; TheNextWeb, 2026). At scale, the gap holds. AI-native companies are running at roughly 6x the revenue per employee of legacy SaaS, with the median traditional SaaS company at around 200,000 dollars per head (Deepstar, 2025; web-strategist, 2025).

This changes three things at once.

It changes who can attempt what. The old read on B2B software was that scope plus enterprise distribution required tens of millions of capital and three-figure headcount before serious problems were tractable. That floor is lower. A six- to twenty-person team can now credibly ship the surface area of what a thirty- to sixty-person team shipped in 2022. Tomasz Tunguz noted the obvious downstream effect: ARR per employee measured in the millions is now a category, not an outlier (Tunguz, 2025).

It changes underwriting. Seed economics are bifurcating. Rounds between 200,000 dollars and 5 million dollars fell from 70 percent of seed funding in 2018 to 26 percent in 2025, while sub-100 million seed rounds for elite AI teams are now common (Crunchbase, 2025; Institutional Investor, 2025). The underwriting question is no longer how many engineers you can hire with the round. It is whether the team can hold a small, dense centre against an incumbent’s distribution and an open-source clone.

It changes the resource advantage that protected incumbents. The classic SaaS defence ran on distribution scale, services revenue, and the implementation burden a buyer would not absorb twice. AI does not erase those. It does narrow the cost asymmetry that justified them. When the orchestration layer above an incumbent system becomes the workflow the customer actually uses, the system of record becomes plumbing (Bain, 2025; a16z, 2025). The fight moves from features inside the seat to control of the workflow above it. A copilot is not a moat. Workflow ownership might be.

What the build did not change is the harder half. Taste did not get cheaper. Enterprise distribution did not get easier. The buying committee did not shrink. Compliance, SOC 2, data residency, integration into systems of record, change management inside large customers: none of this collapsed because the code got faster. Vanta posted 400 percent growth in 2025 precisely because the compliance surface enterprises require has expanded, not contracted (Sprinto, 2026). Sixty-five percent of AI tools inside enterprises are now operating without IT approval, which is a near-term tailwind for bottom-up adoption and a near-term ceiling on contracted revenue (NICE, 2026). A small team can build the product. A small team cannot wave away the integration burden, the security review, or the renewal conversation.

The pattern, as an early read rather than a law: AI does not just make small teams faster. It widens the band of problems a small team will rationally attempt. The constraint shifts from engineering capacity to taste, distribution, and the willingness of an enterprise buyer to trust a thin organisation with a thick workflow. The right question for a CEO is no longer “how many engineers do we need” but “what scope can a dense core hold, and where does the organisation have to thicken to defend it.” The right question for a board is whether existing services revenue and headcount were ever moats or only scaffolding.

The next test is sharper. Hold the engineering team at its current size for two more product cycles. Spend the freed capital on workflow integration, security posture, and one specialist hire in the function that owns the buyer. Then read the renewal cohort at twelve months. If renewal quality holds, the small-team model has crossed from build economics into commercial defensibility. If it does not, the model is a build-stage advantage that resets at the enterprise gate.

Sources

1. How AI is transforming work at Anthropic, Anthropic, December 2, 2025
2. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, arXiv, 2023
3. GitHub Copilot Results Show 55% Faster Development with AI, Blesssphere, 2025
4. Cursor Hit $1B ARR in 17 Months, SaaStr, 2025
5. Working at Cursor (Anysphere) in 2026, jobsbyculture, 2026
6. Lovable says it added $100M in revenue last month alone, with just 146 employees, TechCrunch, March 11, 2026
7. Midjourney Revenue 2025: $500M ARR, getlatka, 2025
8. How Cognition AI hit $80M revenue with a 49 person team in 2025, getlatka, 2025
9. Cognition raised $1B at $26B valuation, The Next Web, 2026
10. AI-Native Startups Growing 100%+ While Traditional SaaS Stalls at 23%, Deepstar Strategic, 2025
11. How Much More Efficient Should a SaaS Startup Be When Using AI?, Tomasz Tunguz, 2025
12. Seed Funding Is Bigger Than Ever And Harder To Get, Crunchbase, 2025
13. Will Agentic AI Disrupt SaaS?, Bain & Company, 2025
14. Good news: AI Will Eat Application Software, a16z, 2025
15. 5 AI Compliance Companies You Must Know In 2026, Sprinto, 2026