Qodo caps its own engineers at $10,000 of AI tokens a month, and grades each dollar by benefits over costs

Most developers at Qodo do not use up the $10,000 monthly ceiling on AI tokens their employer gives them, and the cap exists so somebody has to answer which use of the money was worth it. CEO Itamar Friedman told The New Stack that Qodo, a $70 million Series B company whose product checks AI-generated code before it merges, treats every token the same way it treats every other line in the budget: as an amount you can measure the return on, or an amount you cannot explain.

Friedman calls the number "generous." He gave The New Stack the reasoning behind it too. The cap forces "visibility and efficiency" so the company can "scale without runaway costs." The question the ceiling makes somebody answer, in his words: "Which path of automation or usage will be the best use of our money?"

The equation

Friedman's own account of it, quoted by The New Stack, borrows a shape from The Phoenix Project, the 2013 DevOps novel that sorts software work into good buckets and bad buckets. Add up what AI produced, divide by what it cost:

AI benefits divided by AI costs equals AI ROI, where a larger number is better.

Friedman told the reporter he thinks a startup should pick six to eight terms to plug in. Qodo's own numbers, by its own account: roughly double the pull requests every couple of months, and a falling count of bugs and incidents. Plugged into the ratio, more shipped and less broken makes the top bigger and the bottom smaller, so the score goes up. Qodo's separate figure that its AI infrastructure spend is growing about 5x year over year, driven by customers rather than internal engineers, sits on the cost side of the same fraction.

How the same math gives a smaller number

Friedman's other example is the one that gives the article its title. Qodo tried to fully automate its email, then pulled back after finding that the model could not match tone or extract tasks from messages well enough. The company shifted from automation to enhancement on that job. Faster code generation on the top of the ratio, and slower code review sitting on the bottom, produces a smaller final number. That is how, as Alex Wilhelm summarises the risk in The New Stack, a team can spend a large amount on AI credits and produce less shipped software than before.

Two limits worth stating in Qodo's own terms. Friedman told The New Stack the numbers on the top and bottom are the company's own, not an independent audit, and that the whole thing is a simplification. His argument for using it anyway is that a rough calculation written down beats keeping the same information as an unwritten idea. "What you can't measure, you can't improve," he said.

The useful thing here is smaller than a formal ROI model and more honest than a token dashboard. Any team spending on AI can list what they actually get back and what they actually pay, decide not to count usage as its own reward, and check every couple of months whether the ratio is going up. When the answer is no, adding more credits will make the score worse.

Source

• Alex Wilhelm, The New Stack, AI spending can run negative. Qodo's CEO built an ROI equation to fix it., 23 September 2026.