Your AI productivity metrics are rising. Token spend, PR volume, deployment frequency: all up, all appearing on executive dashboards, all being used by people with authority to make consequential decisions based on them.
The industry has been here before. The numbers are different. The mistake is the same.
The KLOC playbook
In the 1970s and 1980s, software productivity was measured in KLOC (thousands of lines of code). The logic seemed sound: more code means more work done. So teams targeted it, managers tracked it, and the metric went dutifully up.
The problems showed up quickly. KLOC incentivised verbosity. Refactoring, which improves a system by deleting code, registered as negative productivity. Teams learned to write bloated implementations when a clean solution would have been shorter. The metric was easy to game because it bore no relationship to what anyone actually cared about.
By the 1990s, KLOC was widely recognised as a mistake. The industry moved on to velocity: story points completed per sprint. Then to PRs merged, deployment frequency, cycle time. Each felt more sophisticated than the last. Each had the same structural flaw: output metrics measure what teams produce, and say nothing about whether what was produced was worth producing. And in each case, teams learned to optimise for the number rather than the outcome behind it.
I’ve argued before that this is not a new tendency. Teams have been building paperclip machines for decades: optimising powerful systems for the wrong objective and wondering why the output doesn’t translate into what they actually wanted. Goodhart’s Law is the formal name: when a measure becomes a target, it ceases to be a good measure.
What tokenmaxxing actually measures
Tokenmaxxing is what you call it when engineers compete on token consumption: using as much AI as possible to score high on internal productivity leaderboards. Axios reported this year that some tech companies have built exactly these leaderboards. Who burns the most AI gets the highest performance rating.
The name is new. The dynamic is not. Competing on token consumption is competing on activity rather than result, which is exactly what teams did with story points when velocity became the thing managers watched. The tool changed. The incentive structure did not.
Itamar Gilad writes: “Commits, pull requests, and tokens suffer from the exact same problem as lines of code did in the 1980s — they say nothing about the real value the software creates.”
The evidence behind that claim is specific. A/B tests consistently show a success rate of around 33%: most things that get built don’t deliver the value that was expected. If AI doubles throughput while the hit rate stays the same, it doubles the waste as well as the wins. Gilad simulated three teams starting from the same output rate: one optimising for throughput (10% faster), one optimising for outcomes via product discovery (10% slower), one unchanged. Over two years, the outcomes-driven team generated nine times more business value than the throughput-maximising team, despite shipping more slowly.
The productivity is real. The value may not be.
When the metric becomes a weapon
This would matter if it were only a team-level miscalibration. It is not.
John Cutler writes: “McKinsey, BCG, Deloitte, EY, etc. are all pitching that we can cut our workforce by 30% in 2 years. And uninformed execs without informed reps are buying the story.” Output is up per engineer, therefore you need fewer engineers. It is the KLOC fallacy, repackaged as an AI efficiency dividend and sold to executives who don’t have the context to question it.
Someone is about to get fired because their token spend was low. Someone else is about to receive a positive performance review for generating outputs that nobody validated and nobody used. Both outcomes follow directly from measuring the wrong thing.
Outcome, not output
The alternative is not a secret. It is hard to operationalise and easy to defer.
What it looks like in practice: tracking whether what got built was adopted, whether it changed behaviour, whether it produced commercial value. Not did we ship on time but did shipping this matter. Gilad’s framing is useful here: if the cost of building is falling while the hit rate stays flat, the highest-leverage thing is making better decisions before building: not making it faster to build the wrong thing.
The engineering teams getting the most from AI are not the ones burning the most tokens. They are the ones with the clearest sense of what good output looks like, and the discipline to stop generating when they have it.
That has never been about the tool.
The KLOC lesson took twenty years to land. The story points lesson is still landing. The next version is being built right now.
Leave a Reply