SaaStr Runs $20M on 3 Humans and 20 Agents. Revenue per Human Hides the 13-Hour Review Day.

Jason Lemkin published the revenue-per-employee table yesterday that every board will be shown this quarter, with SaaStr's own number in it. This is a Ship Story: what SaaStr decided, what the number holds, and the two lines I would add under it.

The short version

• Lemkin's November 9 post, as he reports it from ICONIQ, Anthropic's S-1, and company filings, puts the old B2B median at $237K of ARR per employee, ICONIQ's top-growth companies at $655K, Anthropic over $6.6M, Palantir about $2.0M, and Cloudflare from about $496K to about $690K after cutting 20% in May while growth rose from 34% to 39%.
• SaaStr itself is on track for about $20M in revenue with 3 humans and about 20 agents, about $6.7M per human. The same post says the humans still handle every real reply and the daily check on the agents, and that a recent 13-hour day went entirely to reviewing and fixing agent work.
• Revenue per employee was a planning ratio for a world where the next $237K needed another person's hours, and Lemkin's new planning line, a person per $1.75M to $9M, keeps the unit and shrinks the denominator.
• The hours did not leave the business; they moved from doing to reviewing, where a head count cannot see them. When I audited the 39 agents I had appointed by April, 13 were orphans nothing outside the fleet read and 13 fired at the hour the reader sits down.
• Keep the ratio on the slide and put two lines under it: reviewed hours per week by agent with the reader's name, and errors that reached a customer by agent, which for SaaStr's quote-to-cash agent is one.

What Lemkin shipped

Two decisions, in two posts a day apart, both with numbers attached, which is rarer than it should be.

The first is the company. SaaStr, per the November 9 post, is on track for about $20M in revenue this year with 3 humans and about 20 agents in production. Six or so agents do most of the daily work. AI SDRs handle about 90% of outbound. A quote-to-cash agent "has run on its own since its fourth deal, with one wrong invoice since." Agents handle first touch, follow-up, every renewal, invoicing, collections, and commissions. Humans handle every real reply, non-standard pricing, the biggest relationships, and the daily check on the agents. The November 8 post puts the fleet at 21 agents and adds that he turned off Canva after 12 years because the agents cannot log into it, and that moving SaaStr Connect, the matching and recruiting app, to Haiku 5.5 and a cheaper model mix cut total costs 72% for the same workflow, the first time since early 2025 that anything he ran agentically got cheaper per use on his own bill.

The second is the table, and it is the one that will travel. ICONIQ's State of Software 2025 puts the typical B2B company at $237K of ARR per FTE, and its 2026 Pacesetter Index puts the median top-growth company at $100M+ ARR at $655K per employee while growing 115%, with GTM teams 21% to 43% smaller at the same ARR than they used to be. Anthropic's S-1 reports trailing revenue per employee over $6.6M at September 30, 2026, up from $1.4M a year earlier, with headcount tripled. Palantir's Q3 annualizes to about $2.0M on 4,634 people, up 205 people in a year while quarterly revenue grew 96%. Cloudflare cut about 20% in May, growth went from 34% to 36% to 39%, and revenue per employee went from about $496K to about $690K, with internal AI tool usage up more than 600% in three months. Atlassian cut about 1,600 in March, more than 900 in R&D, and cloud revenue grew 31% against 23% guidance. AppLovin does $1.92B a quarter on 876 full-time employees, about $8.8M each.

His planning line for 2027: under $300K per employee and growing under 30% will draw board questions; $650K is where the best private companies at $100M+ already are; $1.5M is where Palantir, Bending Spoons, Anthropic, and AppLovin sit. "The median company will keep hiring a person for every $237K. The companies on the list above add a person for every $1.75M to $9M of revenue."

Lemkin says which figures he annualized himself and which companies report the number directly, and I did not open the filings. Everything above is his.

What the ratio holds

Revenue per employee was a planning ratio. For fifteen years, as Lemkin says, every $200K to $300K of new revenue needed another person, and boards approved hiring plans against it. The ratio worked because a person's hours were the thing that produced the revenue. Add revenue, add hours, add a head.

His new line keeps the ratio and changes the number: a person per $1.75M to $9M. Same unit. Smaller denominator.

Now read the SaaStr paragraph again. The humans handle every real reply, non-standard pricing, the biggest relationships, "and the daily check on the agents." Then: "A recent 13-hour day went entirely to reviewing and fixing agent work. The headcount is small, and a big share of the humans' time goes to reviewing the agents."

The hours did not leave the business. They moved from doing the work to reviewing it. A ratio that counts heads cannot see that move, because the head is the same head. It just spends its day differently, and on some days it spends 13 hours on the agents and none on the revenue.

This is the SaaS to AI Business Models thesis seen from the cost side. Software priced per seat was priced against a labor cost that AI removes, and the margin structure moves with the pricing. What the revenue-per-employee table shows is the labor cost coming out of the seat count. What it hides is the labor cost going into the review queue, which is a cost of revenue now, not an overhead. I wrote in Vercel Runs 'Our Internal Sierra' for $150K a Year. Write the Invoice Nobody Sends. that the $150K was tokens and infrastructure with the three engineers outside it. The $6.7M per human is the same shape one level up: revenue over heads, with the hours outside it.

$6.7M per human

I did not count them either

I am not writing this from the outside. By April I had appointed 39 agents, and the audit I wrote up in 44 Appointed Agents, One Attention Budget: Three Lines on the Charter found two things I had not planned. Thirteen of the 39 were orphans, agents whose output nothing outside the fleet ever read. And 13 of the 39 fired between 7 and 9 a.m. without anyone choosing that window, because that is when the reader sits down and the fleet had drifted toward the only person who would open the output. The fleet is 44 today and 15 of them still name a 7, 8, or 9 a.m. time.

Nobody had counted the review hours because nobody had counted the readers. The fix I put on the charter, before any agent is appointed, is three lines: a slot, a budget line, and a reader's name. Blank reader, no appointment.

Twenty agents and three reviewers is the same fleet shape with better numbers. Lemkin's humans are the readers, and a 13-hour day is what the reader's line looks like when it is a person instead of a line.

The unit is in his post

The right unit is in Lemkin's own paragraph, one sentence after the ratio. The quote-to-cash agent "has run on its own since its fourth deal, with one wrong invoice since."

That is the number. Not heads, and not hours alone. Outcomes per reviewed hour, with the error that got through next to it. An agent that invoices every deal and has produced one wrong invoice since deal four is an agent whose review can be a sample, and the hours it needs are small and known. An agent that needs a human on every real reply is an agent whose review hours are the cost of each outcome, and they belong on the per-outcome line I described in Three Token Bills, One Unit, next to the tokens.

So two lines under the ratio, on the same slide the board sees.

Reviewed hours per week, by agent, with the reader's name. Not estimated. Counted, the way the 13-hour day was counted, because it was lived.

Errors that reached a customer, by agent. For SaaStr's quote-to-cash agent that is a known number: one. For most fleets it is unknown, which is a different thing from zero.

Keep revenue per human above them. Lemkin is right that boards will ask for it and right that under $300K will draw questions. The two lines are what turn the question into a conversation about cost of goods. A company that reports $6.7M per human and not the review hours is reporting gross margin before cost of goods, and Cloudflare's 600% jump in internal AI tool use alongside its 40% jump in revenue per employee, as Lemkin reports both, is the same hours moving to the same place at a company with 4,700 people.

The 72% is worth one more sentence. It changes the token line, and Lemkin's lesson, re-benchmark your model mix every time a small model ships, is right. It does not touch the hours line. Tokens re-benchmark themselves every quarter now. Reviewed hours only fall when an agent's error count earns it a sample instead of a reader, and that is a decision somebody has to make per agent, with the count in front of them.

This week

If you run agents and you have shown anyone a revenue-per-head number: pick the three agents with the most output and write down, for last week, who read it and for how long. Then write down how many of their outputs reached a customer wrong. Put both under the ratio. If the hours surprise you, that is the line the ratio was hiding. If the errors are zero, check whether anyone would know.

Related answer: What should an AI-native company report next to revenue per employee?

Sources: Jason Lemkin, "The New Normal in Revenue Per Employee: $237K Is the Old Median, $655K Is the New Top Tier, and Six B2B Leaders Now Clear $1.5M," SaaStr, November 9, 2026, read in full, and "Jason's Takes on This Week's 20VC: A 72% Cost Drop on Our Own App, the $100M Acquisition Bar, and DoorDash's Third Competitor," SaaStr, November 8, 2026, read in full. Every company figure is as Lemkin reports it from ICONIQ, Anthropic's S-1, and the filings he cites, which I did not open; the ICONIQ, Palantir, Cloudflare, Atlassian, and AppLovin numbers are his, including the ones he says he annualized. The 39, 13, 13, 44, and 15 are my own fleet counts, from the post linked above.