The Other Meta Story

In 1913, an engineering titan of its age built a moving assembly line and discovered that the rest of the factory did not move with it.

In May, an engineering titan of ours replaced a tenth of its people with AI agents and discovered the same thing.

There were two stories about Meta last week. This newsletter is about the one that didn't make the mainstream headlines, since we were distracted by the other.

First, that one: last Wednesday, Meta made the front pages for a settlement of around $17 billion with 29 US states over the harm its platforms did to teenagers, which is the largest consumer-protection settlement since Big Tobacco, and rightly the lead story.

However, the same morning, Reuters published the other Meta story, and for anyone running a company through the next three years of AI, it is the one with more to teach. Meta designed an organisation around agents. Modelled cutting some teams by 60%. Cut roughly 8,000 jobs. And cancelled the second wave of cuts on the night before the first began, because its own numbers showed the thing every engineering company eventually sees: the engine got faster, and the machine around it didn't.

I have been arguing since December 2024 that AI's real promise is to crush "management's 'iron triangle'" and deliver better, faster and cheaper at once. In March I wrote that "faster and cheaper are table stakes, not transformation." Meta is the first test of both, from the inside, at scale.

We have been here before. A long time ago. And we don't seem to learn.

In 1900, almost every factory in America ran off a single steel shaft in the ceiling, with belts dropping down to every machine. When electricity arrived, most owners did the obvious engineering thing: they unbolted the steam engine and bolted an electric motor onto the same shaft. The power source changed. The shaft stayed. And with it stayed the layout, the sequence of work, the habits of the people on the floor. Electrifying the factory took an afternoon. Changing how the line behaved took forty years.

Economists call this the productivity paradox, and the Stanford economist Paul David explained it best in a 1990 paper called The Dynamo and the Computer. Edison's Pearl Street station opened in 1882. By 1899, electric motors still provided less than 5% of the power in American factories. Even where they did, the shaft ran the building: machines stood where the belts could reach them, not where the work made sense; plants were three storeys tall because power travelled better up than sideways.

The gain came in the 1920s, when engineers finally put a small motor on every machine and the shaft came out. Only then could the floor be rebuilt around the flow of work: single storey, daylight, machines in the order the product needed them. Electrification drove roughly half of all manufacturing productivity growth in that decade. Same electricity. Four decades apart. Because the hard part was never the electricity.

That is the whole story of Meta, told in advance.

In January, at Mark Zuckerberg's annual leadership retreat, Meta's executives drew up Project OT, for Organization Transformation (which is exactly what an engineering company would call a project).

The design was the AI-native company as its keenest advocates describe it: agents doing much of the daily work; humans regrouped into "talent-dense" pods of three to five to oversee them; layers of management removed; product designers and engineers folded into a single generic role, "builder." Scenario planning explored cutting some teams by as much as 60%. Two waves this year: May, then November.

Now, some credit to start with: every leadership team of any size runs the 60% scenario (and I've been in more than my fair share of those planning sessions). It is a genre, and its purpose is to find the edge of the possible, not to schedule it. Meta's statement to Reuters is fair as far as it goes: "Ultimately, we didn't move forward with every scenario from the exercise and it was never assumed we would." Meta measured, and stopped before it executed. That is governance working. Reuters could not establish what tipped the decision, and I won't pretend to.

What I can read is the data Meta declined to comment on. After the first wave, code changes to Meta's platforms and infrastructure rose 220% year on year, a figure CTO Andrew Bosworth shared with staff as proof the smaller teams were working. Changes that produced a new or improved feature for an actual user rose 36%. Major technical and security incidents rose 40%. Time spent responding to them rose 70%. Infrastructure teams had flagged reliability warnings as early as March; by June, hackers had used Meta's own AI customer-support bot to break into high-profile Instagram accounts. Staff favourability fell from 74% to 55%. In July, Zuckerberg told a town hall that agent technology had not "accelerated" as fast as he expected, and gave it three to six months.

Three times the output from the engine. A third more product. Nearly half as much again in breakage. That is a motor bolted to a shaft.

The iron triangle of management is well known: faster, cheaper, better. Pick two and it comes at the expense of the one you leave out. The promise of every industrial revolution is to break and redefine this same iron triangle, of what can be done faster, cheaper AND better. We simply are not there yet for this AI industrial revolution. In this case, Meta got faster: 220% is faster by any definition. Meta got cheaper: a third more shipped product from a tenth fewer people is a cost curve any CFO would sign.

What Meta did not get was better. Incidents up 40% is the definition of worse. Firefighting up 70% is the invoice for it. And the 184 points of daylight between 220 and 36 is the review load, the integration work and the clean-up that the old system (the codebase, the review process, the incident tooling, the management habits, all built for a human pace of change) could not absorb.

Faster and cheaper are what you get from swapping the engine. Better is what you get from rebuilding the floor. Nobody has rebuilt the floor yet. Not Meta, with $125–145 billion of capex this year and the best engineers money can buy. Not any of us. We are in the age of faster and cheaper, and it is worth saying plainly that this is roughly where the electricity story was in 1890.

Three things, and they were as true of steam as they are of agents.

It takes longer than you expect.

Hindsight sands the failures off. Electricity reached half of factories in twenty years and the productivity statistics in forty. The internet was commercially available from 1994; be kind and say it took a decade to properly digitise commerce. Usable large language models have existed for under four years. Agents that do anything useful, one or two. Adoption gets faster with each revolution. Transformation does not, much.

You have to think in systems, not in engineering.

Engineers change the system and expect everything around it to adjust to the system. Ford, the engineering titan of its day, cut the time to build a Model T from twelve and a half hours to ninety-three minutes and found that the labour market, the training model and the wage structure around the line did not adjust to it. Turnover hit 370%; the company hired 50,448 people in 1913 to keep 13,623. It took another year, and a redesign of everything around the line, before the line delivered. Meta, the engineering titan of ours, changed the org chart and the tooling and assumed the codebase, the review culture and the people would adjust. They did what systems do. The objection is that software has no shaft, so code diffuses at zero cost, this time is different. But the shaft was never the engine. It was everything built around the engine, and every organisation has one.

The weakest link is the advantage.

The slowest part of every one of these transformations was people and how they work. It is also the part no competitor can copy. Ford's rivals could buy the same machines; they could not buy his floor, or the people who had learned to run it. Meta's competitors can rent the same models tomorrow. What they cannot rent is an organisation that has taken the shaft out and the fact that one of the best-resourced engineering companies on earth hasn't managed it yet should be read not as a warning about AI but as a measure of how much advantage is still on the table. This is why the honest conclusion of Project OT is not AI instead of humans. It is human plus AI, and the humans are the part that compounds.

Meta's pods survived. Some engineers moved into agent roles were quietly allowed back to their old teams. Zuckerberg's three to six months from July lands in October or November, and whatever Meta does in that window will tell you more about where agents actually are than any keynote. One Meta story closed last Wednesday, with a cheque. The other is still open, and it is the one that will decide what the company looks like in 2027.

Meta did not find out that AI can't replace people. It found out that faster and cheaper come from swapping the engine, and better only comes from rebuilding the floor, and nobody has rebuilt the floor yet.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 31 August 2026 in the Brandflow newsletter on LinkedIn.