A 9-Person Startup Shrank to 4

Paying Customers Held at 10,000 After Five Engineers Left

Reader, Butternut AI, which makes a no-code website builder, once had 9 employees. Today it has 4. At the end of 2025, its two co-founders let 5 engineers go. They concluded that AI agents could do their work—and do it 24 hours a day. In its 3rd year, the company counts around 10,000 paying customers, and co-founder Simar Singh says the product continues to grow.

The Wall Street Journal recently profiled several such companies: early-stage founders who ask whether AI can do the job before they ever post a job opening. Some have let existing staff go, while others cut ties with contractors and agencies. The goal is to stay small for as long as possible.

Stories of tiny teams generating massive revenue are no longer new. What makes this report worth a closer look comes down to two questions: which roles disappeared first, and how long the investors quoted believe these startups can actually stay this small.

The First Roles Cut Turned Requests into Deliverables

Lindy, which builds AI “teammates” that work inside Slack, has just over 20 employees and has raised $50 million. In January 2026, alongside a pivot in product direction, it let several marketing staff go. Its cumulative user base has since grown to 500,000, yet marketing is still handled by just two people and AI agents. The company continues to hire developers, but it never refilled those marketing positions.

Haneen Azhar, Lindy’s growth marketing manager, describes the shift this way: where one person used to build marketing data dashboards, “now that’s a single prompt.” The number of external creative agencies Lindy hired for content and video production dropped from five to zero. Videos are now produced using AI video generation.

AgentCollect, which uses AI agents to collect overdue accounts, shrank from a peak of 50 people to about 30. Roles that were largely automated include billing coordinators, customer call quality reviewers, and a product manager. CEO Jon Benner explains that this was not about cutting costs, but about a hiring principle: employees should work on tasks requiring creativity and judgment.

Lay them side by side, and a pattern emerges. These were jobs where someone stated a need, and a person turned it into a deliverable—a dashboard, a video, a billing outreach, a call scorecard. The roles that handled the handoff and polish between request and final output were the first to go.

Two caveats are worth keeping in mind, however. Lindy’s layoffs coincided with a product pivot, meaning the decision cannot be explained by AI alone. Furthermore, going from five agencies to zero does not show up on headcount charts. Looking only at internal rosters conceals the scale of work eliminated outside company walls. The traces of that vanished work remain with the contractors who used to do it. The story of freelance developer rates discussed in Issue 164 is the other side of this very same coin.

From 32.5 to 20.7 People: Interest Rates Are in That Number Too

The article also cites an analysis from LinkedIn. Grouping US startups with 5 or more employees in their first year by founding cohort, it compared average headcount in year 3. Companies founded in 2018 averaged 32.5 employees by year 3; those founded in 2022 averaged 20.7. The average headcount of US startups peaked in 2021 and has been trending downward ever since.

Reading this figure purely as an AI effect is a mistake. Year 3 for a 2018 cohort falls roughly in 2020–2021, the absolute peak of venture capital abundance. Startups founded in 2022 grew up after interest rates surged. The force driving leaner hiring reflects both changing tools and changing capital, and this comparison alone cannot separate the two.

The examples highlighted by Y Combinator¹ require similar caution. Emergent, which went through the program in 2024, reportedly reached $15 million in annualized revenue² with 15 people, while Retell AI, from the same batch, hit $60 million with around 40 people. These serve as proof that small teams can generate big numbers, but they are cherry-picked accelerator success stories. Startups that stayed tiny and quietly vanished are nowhere on that list.

Customers and Investors Set the Limit on Staying Small

So how long can these companies hold out as tiny teams? The answers from investors in the article point to the market rather than model capabilities.

Liz Wessel of First Round Capital argues that it depends on the company type, the product, and how fast revenue is growing. Companies experiencing explosive growth are still competing fiercely for top-tier engineers.

Cowboy Ventures founder Aileen Lee draws a more specific line. Today, two co-founders can build a prototype, get it into customers’ hands, and validate traction without hiring a soul. But every business eventually hits a ceiling where it cannot scale further without headcount—especially when selling to enterprise clients. When selling contracts worth $60,000 to $75,000 a year to mid-market or enterprise buyers, the purchasing side wants to buy from a company with real people.

Backers have their own benchmarks as well. Lee notes that venture capitalists expect $4 million to $5 million in first-year revenue to justify a Series A³, and hitting that number without people is exceedingly difficult.

The reality that enterprise clients demand human presence is visible from the other side too. As discussed in Issue 197, Microsoft, AWS, OpenAI, and Anthropic have all set up dedicated organizations this year to embed engineers directly inside client sites. Even the companies building the most advanced AI lead with humans when massive contracts are on the line.

Explaining It to a Junior Took as Much Work as Prompting Claude

The final figure in the article is Advait Paliwal, a solo founder who has not yet taken outside funding. He is building a consumer mobile app that interacts with multiple AI agents, and he treats OpenAI’s Codex as his chief technology officer. Instead of hiring employees, he runs 20 subscriptions across Claude and Codex.

Over the past year, he interviewed 4 software engineering candidates, including a 14-year-old applicant from Belgium. In the end, he hired none. The volume of instructions he had to provide to a human was identical to what he gave Claude, and Claude responded faster.

This comparison measures cost only in the present moment. A junior hire requires extensive explanation precisely because they are learning the craft. For a solo founder running a company alone, that explanation time is indeed an unbearable overhead. But if everyone runs the same calculus, the entry-level roles where people learn by being guided will vanish entirely. The “learning ladder” problem explored in Issue 154 reappears here at startup scale.

Oswarld’s Lens

Whenever I design enterprise deployments or consulting engagements, I operate on the premise that models will always improve. That is why founders who expand subscriptions instead of headcounts make sense to me. The work of turning requests into deliverables will increasingly shift to models as their capabilities grow.

Yet this article highlights a boundary that does not budge even as models improve: who a client expects to meet before signing a contract, and at what revenue threshold an investor begins to demand headcount. This line is drawn not by model performance, but by trust and convention between buyers and sellers.

That is why designing an organization around the single question “Can AI do this?” solves only half the problem. The question that must accompany it is: “Does the customer need to know who did the work?” Roles where the answer to the first question is yes will continue to shrink. But roles where the answer to the second question is yes may actually expand as the company scales.

Closing

Butternut’s 4 people, Lindy’s 2 marketers, and Paliwal’s 20 subscriptions prove that staying small is viable. The investors in the very same article make clear where that viability reaches its limits. A startup’s headcount reflects who it sells to and what it sells just as much as what AI can do.

---

💬 Among the tasks you do today, which have shifted to AI, and which do clients still insist on receiving from a human? Share your thoughts in the comments.

📨 If you know a founder or team lead rethinking their hiring roadmap, please share this issue with them.

---

References & Further Reading

• The Wall Street Journal, “These Tiny Startups Are Getting Even Smaller With Help From AI” (2026). The source for the Butternut AI, Lindy, and AgentCollect case studies, the LinkedIn analysis, Y Combinator examples, and investor commentary.
• INLEVEL9 Letter Issue 182, “No Solopreneur Unicorns, but Twice as Many Millionaires”
• INLEVEL9 Letter Issue 197, “Microsoft Sent 6,000 Employees into Other Companies”
• INLEVEL9 Letter Issue 164, “Why a Rooftop Developer Delivers Food Every Night”
• INLEVEL9 Letter Issue 154, “Ferrari Designers Still Start with a Pencil”

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Y Combinator: A US accelerator that invests in early-stage startups and runs an intensive multi-month cohort program. ↩

2. Annualized revenue: A short-term revenue figure, such as a single month, multiplied out to reflect a full year. It may differ from actual trailing twelve-month revenue. ↩

3. Series A: The first significant institutional venture capital round following a seed investment. ↩