Thirty Percent of What?
Meta's code changes were up 220% year over year. Changes that actually reached users were up 36%.
Reuters published the investigation yesterday, and those two numbers are what I'd put in front of anyone about to restructure a team around AI.
Meta reportedly ran Project OT, an AI-native restructuring exercise that envisioned smaller teams working with autonomous agents. Internal scenarios explored cuts of up to 60% in some teams, and the planned second wave was later cancelled.
What I keep coming back to is the contrast between those two percentages, because that is where the argument about AI and headcount is actually happening, and it's the part I rarely see measured properly.
I've seen the same shape on go-lives that had nothing to do with AI. The status pack is green, every workstream reports on schedule, the activity numbers look impressive, and then somebody asks what has landed in production, and it is a fraction of what the dashboard implied.
On a railway portfolio, we built much better visibility across the system so we could spot problems before they became incidents. Incidents fell by more than 60%. The improvement came from seeing what was happening across the chain early enough to act, rather than from any one team working faster.
Activity is easy to move. Throughput is not, because between the faster task and the business outcome there is still a queue of handoffs, exceptions, approvals, rework, institutional knowledge that lives in two people's heads, and now the occasional AI-generated mess somebody cleans up on a Friday afternoon.
None of that disappears because an individual got quicker at their part.
So before anyone redraws an org chart around an assumed productivity gain, I'd want a fairly boring sequence of evidence.
Prove the capability in production, not in a pilot. Redesign the workflow around what the technology can genuinely do. Measure end-to-end throughput, quality, exceptions and rework. Work out where usable capacity has appeared, and in whose week.
Only then make the organizational decision.
That last gate is the one that gets skipped, because "AI saved us 30%" sounds precise right up until somebody asks: thirty percent of what, and where does that saved capacity show up in the operating result?
The version that works is genuinely unglamorous. Somebody can name the queue that got shorter. Somebody knows which handoff stopped happening. Somebody can point at the week that freed up. Finance can see what got cheaper.
If nobody can do that, the capacity is still theoretical, and restructuring around a theory is an expensive way to find out you were wrong.
What's the AI productivity number floating around your organization right now, and does anybody know what it's a percentage of?
Reuters investigation:
Comments