Pay for output, not hours: Tenex, a 75-person AI firm, says each engineer now does about twice the work
Oct 29, 2026Original Latent Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC ↗
In 3 lines
- Tenex, a 75-person firm, pays engineers for output (story points) instead of hours.
- AI roughly doubled output per engineer, yet the top-to-bottom pay gap is about 4x, not 10x.
- Less coding means more time on customer context, and how to pay for that work is still unsolved.
"If AI makes an engineer faster, an hourly rate can make the engineer less money for the same job." That is why Tenex, a 75-person AI consulting and development firm, ties pay to story points, or delivered output. In a podcast Latent Space published on Oct. 21, cofounder Alex Lieberman and engineering director CJ Hess walked through what 18 months of the experiment has shown and where it strains. This piece summarizes their conversation.
When AI gets faster, hourly pay becomes a penalty
Tenex pays AI engineers for what they deliver, not the hours they spend. As cofounder Alex Lieberman puts it, if AI makes an engineer faster, an hourly rate pays less for the same job. So the company flipped the incentive, and even at 75 people story points remain an important part of compensation.
"The feared side effects weren't the big problem"
The classic worries about output-based pay are that people inflate story points or ship sloppy code. Lieberman said neither turned out to be a real problem, because Tenex's AI strategists, who are responsible for customer retention and growth, are incentivized on the opposite side of the engineers.
The pay gap in numbers: 4x, not 10x
Lieberman said the highest-paid engineer earns roughly 4x the lowest-paid. Output-based pay invites talk of '10x engineers' making millions, but the actual spread was narrower. He estimated that each engineer's output has roughly doubled since about December.
Less coding time, more customer context
CJ Hess, Tenex's first hire and director of engineering, said he used to code a lot and now codes less. Given the right context, models can get you to about 90% of a solution, so the value shifts to sitting with a customer, hearing their brain dump and handing it to the model. That makes it hard to reward work that never shows up as a ticket, such as architecture and client meetings.
The project's memory lives in the repo, not in heads
Hess described the team's MetaHarness: shared skills, structured documents and a CLI that agents drive. Plans and decisions are checked into Git next to the code, so an agent returning to a task can read not just what was built but why. As models improve, the team is actually reducing the structure it enforces.
What is one story point worth?
The team tracks average story points per engineer per month. Whether a point is worth the same as a year ago is still an open question; they recalibrate at least once a quarter and call it "a tricky equation." Tenex's own write-up also admits the structure can be overkill for a small team on one project.
Enterprises are still early
Of roughly 100 enterprises Lieberman talks to, about 90% have rolled out Claude or Codex to employees, about 60% have trained them, about 50% have put coding agents in product and engineering, and about 5% run agents in production for non-engineering work. All are his estimates. Permissioning, he said, has become "an absolute nightmare" for CIOs and CISOs.
When building gets cheap, what is left?
The hosts described Kill My SaaS, an experiment to replace a $43,000-a-year speaker-management tool; the team spent about $1,000 on a replacement. Yet the human-judged winner among 72 submissions ranked 19th on the AI judge's leaderboard, a reminder that passing scenarios and being something people want to use are different jobs.
Building isn't the end; owning it is the cost
Lieberman noted that ripping out incumbent software gets harder the further upmarket you go, and that any tool you build carries ongoing maintenance. The host countered that the upside includes a better product experience and faster iteration, not just savings. Lieberman is generally bearish on vertical AI, while Hess separates products that are mainly an agent from tools that improve the workflow around the task.
