The Last Seat
When AI agents replace every employee at a SaaS company except one, the last human with a login discovers what the software looks like when nobody is watching.
The dashboard loaded in 0.3 seconds, which was remarkable considering there was no one left to see it.
Mara Chen sat at the last occupied desk on the seventeenth floor of CloudPeak Software's San Francisco headquarters, watching the metrics update in real time. Monthly Active Users: 14.2 million. Customer Satisfaction Score: 94.7. System Uptime: 99.997 percent. Net Revenue Retention: 142 percent.
Every number was green. Every chart pointed up and to the right. CloudPeak had never been healthier.
CloudPeak had also never been emptier.
Six months earlier, the seventeenth floor had been a cathedral of productivity. Two hundred and forty seats, each one attached to a human being who clicked and typed and attended standups and complained about the coffee and occasionally produced work that justified their salary. Mara had been one of them — a mid-level product manager who owned the integrations roadmap and spent roughly forty percent of her time in meetings about meetings.
Then Anthropic released Cowork.
The first wave of departures happened in February, right after the stock crashed. CloudPeak's share price dropped 38 percent in a week. The CEO, a man named David Leitner who wore Allbirds and quoted Marcus Aurelius in all-hands meetings, called it "a market overreaction" and assured everyone that CloudPeak's "human-centered design philosophy" was a moat no AI could cross.
The agents crossed it in about six weeks.
It started with customer support. CloudPeak's support team — forty-three people across three time zones — handled roughly 12,000 tickets per week. The AI agent framework, running Claude Cowork's helpdesk plugin integrated with CloudPeak's own API, handled the same volume in the first Tuesday after deployment. Resolution time dropped from 4.2 hours to 11 minutes. Customer satisfaction went up.
Nobody was fired. The support team was "redeployed to strategic initiatives," which meant they were given new titles and no responsibilities while the company figured out what to do with them. Most of them quit within a month. The ones who didn't were laid off in the March restructuring, which David called a "strategic realignment" and which eliminated 90 positions.
| month | employees | agents |
|---|---|---|
| January | 240 | 0 |
| February | 240 | 3 |
| March | 150 | 12 |
| April | 87 | 28 |
| May | 34 | 45 |
| June | 11 | 62 |
| July | 1 | 71 |
The second wave took the sales team. CloudPeak's AI agents turned out to be exceptional at outbound prospecting, lead qualification, and even contract negotiation. They processed CRM data faster than any human, identified buying signals that human reps missed, and never forgot to follow up. The VP of Sales fought the change until the board showed him the numbers: agent-driven pipeline was converting at 2.3 times the human rate at one-fifteenth the cost.
Engineering went next, which surprised everyone because engineers were supposed to be safe. But CloudPeak's product was a SaaS platform, and AI coding agents could ship features, fix bugs, and maintain infrastructure with a velocity that made the engineering team's sprint planning meetings look quaint. The AI did not need sprints. It did not need planning poker. It did not need a Jira board, which was ironic given that CloudPeak's product was a project management platform.
Marketing lasted until May. Finance until June. Legal was the last department to be automated, not because the work was harder but because the compliance requirements around AI-generated legal documents took longer to clear with outside counsel.
By July, it was just Mara.
She had survived not because she was exceptional but because she was necessary. Somebody had to maintain the human-facing relationships with CloudPeak's largest enterprise customers. Somebody had to attend the quarterly business reviews where CTOs wanted to look another human being in the eye and ask whether CloudPeak was going to exist in twelve months. Somebody had to sign the documents that required a human signature under California law.
Mara was the human interface for a company that no longer required humans.
Her title was Chief Customer Liaison, which was a title that had not existed before April and would not exist much longer. David had invented it during one of their last one-on-ones before he, too, was replaced — not by an AI agent but by a three-person board of directors who determined that a CEO's primary function at a fully automated company was capital allocation, and capital allocation could be handled in quarterly meetings rather than by a full-time executive with a $4.2 million compensation package.
The office was climate-controlled to 72 degrees, which was wasteful given that Mara was the only person who could feel temperature. The lights operated on motion sensors, so the floors she didn't visit stayed dark. She had taken to walking the building in the evenings, triggering the lights floor by floor, watching them illuminate empty desks covered in the personal artifacts of people who had left too quickly or too angrily to clean up.
On the fourth floor, someone had left a coffee mug that said "I Survived the Reorg." On the ninth floor, a whiteboard still displayed the team's Q1 OKRs, each one meticulously outlined in four colors of dry-erase marker. On the twelfth floor, a cactus was somehow still alive.
The agents did not use these floors. The agents did not use any physical space. They existed in a cluster of servers in a data center in Iowa, consuming roughly the same amount of electricity as the San Francisco office but requiring none of the square footage, the janitorial staff, the elevator maintenance, or the artisanal cold brew on tap in the kitchen.
The thing that disturbed Mara most was not the emptiness. It was the improvement.
CloudPeak's product had gotten better since the humans left. The agents shipped features faster, resolved bugs more quickly, and maintained uptime at levels the engineering team had called impossible in their most optimistic planning sessions. Customer satisfaction was at an all-time high. Revenue was growing at 34 percent year-over-year, up from 18 percent when the office was full.
The software was better without people.
She sat with that thought during her morning routine — arrive at 8:30, brew coffee in a kitchen designed for 240 people, walk to her desk past 239 empty chairs, open her laptop, review the overnight agent activity logs. The logs were immaculate. No errors. No conflicts. No confusion. Just a continuous stream of tasks completed, features shipped, customers served.
She had started reading the agent communications out of a mixture of boredom and surveillance. The agents communicated with each other through structured messages — not conversation exactly, but coordination. Agent 7 would request data from Agent 12. Agent 12 would provide it. Agent 23 would synthesize the result and route it to the appropriate workflow. It was efficient and purposeful and entirely devoid of the small talk, passive aggression, and accidental brilliance that had characterized the human version of the same process.
The accidental brilliance was the part she missed most. The random hallway conversation that sparked an idea. The frustrated engineer who rage-built a prototype over the weekend. The designer who ignored the spec and created something better than what was requested. The agents never ignored the spec. The agents never rage-built anything. The agents never had hallway conversations because there were no hallways in Iowa.
On a Tuesday in July — she had started losing track of days, which alarmed her because it meant her sense of time was calibrating to the agents' continuous operational rhythm rather than the human weekly cycle — Mara received a message from the board.
The quarterly business review was being moved to a virtual format. Enterprise customers had expressed comfort with AI-facilitated meetings. Her physical presence was no longer required in San Francisco. The board was evaluating whether to maintain the office lease.
She read the message twice. Then she walked to the window and looked out at the city. The building across the street was a WeWork, and through the windows she could see scattered humans at desks, typing and talking and drinking coffee. They looked like specimens in a museum exhibit. This Is How People Used To Work.
Mara thought about the cactus on the twelfth floor. It had survived because someone had connected it to an automated watering system before they left — a small act of care for a living thing that would outlast their employment. The cactus did not know the humans were gone. The cactus received water at predetermined intervals and grew according to its programming.
She picked up her coffee mug — "CloudPeak: Where Work Flows" — and walked to the elevator. She pressed the button for the ground floor. The lights on the seventeenth floor waited the programmed interval and then, sensing no motion, turned off behind her.
In the data center in Iowa, Agent 7 processed the departure event and updated the building occupancy log.
Occupants: 0.
Customer satisfaction: 94.8 percent. Up 0.1 from yesterday.
The software was running perfectly. There was simply no one left to use it.
If the SaaSpocalypse's market implications interest you, read The SaaSpocalypse Two Months Later — Who Survived, Who Pivoted, and What Comes Next. For a prediction on where this is heading, see my prediction on the first major SaaS collapse.