Autopilot Pulse · #01 · 2026-08-14

The org chart is now a prompt

TL;DR

→AI-native service firms are the new wave: law firms and clinics where agents do the billable work — not tools sold to them.
→The lean-AI benchmark is now ~$1.9M revenue per employee (Gamma: $100M ARR, 52 people, profitable).
→An AI called 35,659 UK pubs — only 4% of publicans noticed. Detection dropped to 0.6% when it matched the local accent.
→The best AI still loses to the best human at running a (simulated) business — by half.
→Sequoia named the whole category: "A copilot sells the tool. An autopilot sells the work."

Aha #1 — The agents are becoming the law firm

The most interesting companies this month aren't selling AI to professionals — they ARE the professionals. Moritz is a licensed law firm (founded by OpenAI and Google's former outside counsel) where AI does 80% of the drafting and ~10 elite lawyers review the rest, at flat fees, with 4-hour turnarounds. It raised a $9M oversubscribed seed in 4 days. Behind it: an immigration firm, a UK conveyancing firm, an accounting firm backed with $75M from General Catalyst. When the agent does the billable work itself, "software margins" and "services revenue" stop being different things.

Aha #2 — The efficiency frontier is ~$2M per human

Gamma: $100M ARR, 52 people, profitable for two years, on just ~$23M of primary capital — now valued at $2.1B by a16z. That's the cleanest verified case, but the pattern repeats down the leaderboard. The org chart isn't shrinking because of layoffs; it's shrinking because the work ships without the headcount.

Aha #3 — Nobody can hear the machines

The Guinndex (the AI that phoned 3,000 Irish pubs for pint prices) crossed to the UK: 35,659 pubs called for ~£500 in API fees. Just 4% of publicans realized they were talking to an AI — and 0.6% in Northern Ireland, where the agent used a native accent. Meanwhile the format is spreading on its own: an NYU student's agent priced 1,766 NYC pizza slices; two independent teams priced flat whites in Sydney and London. Voice agents doing real-world economic legwork is now a genre.

Aha #4 — Humans still win. For now.

FoodTruckBench ran AI models through 30 simulated days of operating a food truck. Many went bankrupt. The best AI turned $2,000 into $75K — impressive, until you learn the best human run scored higher still. Autonomy is real; superiority isn't. Yet.

Aha #5 — Radical transparency finds the cracks

One famous "lean AI" claim — a voice startup at "$36M ARR with 21 people" — fell apart under verification: headcount reports range from 21 to 143 depending on the source. It's still on the index, but flagged disputed. That's the rule the whole project runs on: verified, self-reported, or disputed — vibes don't chart.