Meta built an AI engineering division in March at up to 50 employees per manager. Six months later, it is asking some employees in that division to become managers again. If your 2027 headcount plan widens spans of control on the promise that AI absorbs manager work, size each span by the work AI did not absorb (code review, performance calibration, hiring, incident ownership, unblocking) and write down what would make you reverse course before you cut.
This is not an argument that wide spans fail. Plenty of teams run well at 15 or 20 direct reports, and the flattening wave at Amazon, Google and Uber removed layers that were adding little. It is an argument that the number has to come from measured manager load, not from the headcount line you want to hit. Meta has not said why it is reversing course, but it is a live example of how quickly an aggressive span gets revisited.
What Meta Built in March, and What It Is Undoing
Meta's Applied AI Engineering organization was designed at twice the span most org designers treat as the ceiling, and Meta is now recruiting managers back into it. According to an internal memo reported by The Wall Street Journal in early March, the group is led by Maher Saba, a vice president from Reality Labs, reports to CTO Andrew Bosworth, and runs "up to 50 employees per manager" across two teams: one on interfaces and tools, one on tasks, data collection and evaluations. Its job is to build the "data engine" that speeds up improvements to Meta's models. Fortune noted at the time that 50-to-1 is "double the 25-to-1 ratio that is usually seen as the outer limit of the so-called span-of-control scale."
Now the reversal. Business Insider reported, citing four people familiar with the matter, that Meta has begun asking individual contributors in Applied AI whether they would move back into manager roles as part of a recent reorganization, according to Fortune's follow-up. The move is reportedly voluntary. Roughly 7,000 employees were reassigned into the division this year, some of them former managers. Meta did not immediately respond to Fortune's request for comment, and no report includes a revised ratio or a stated reason. Treat any figure you see attached to the "why" with suspicion.
The ratio was also a break with Meta's own rule. In the 2023 "Year of Efficiency" post, Mark Zuckerberg wrote: "We still believe managing each person is very important, so in general we don't want managers to have more than 10 direct reports." Applied AI was built at five times that guideline, inside a company that was cutting hard. Meta's second-quarter results put headcount at 75,472 on June 30 and said that figure still "includes approximately 8,000 employees impacted by the May 2026 headcount reduction." We covered the logic of that cut in Why Meta Is Cutting 8,000 Jobs While Spending $135B on AI.
The Strongest Case for Wide Spans
The case for wide spans is strong, and most of it has nothing to do with AI. Zuckerberg's 2023 argument was that "every layer of a hierarchy adds latency and risk aversion in information flow and decision-making." Amazon set a target to "increase the ratio of individual contributors to managers by at least 15% by the end of Q1 2025." Google told staff in August 2025 that it had 35% fewer managers with small teams than a year earlier; a person familiar with the matter told NBC News that meant managers overseeing fewer than three people, and that many stayed on as individual contributors.
Uber went further on Sept. 2. It is cutting about 3,300 jobs, roughly 10% of staff, reducing managers by 20%, halving the number of one- and two-person teams and letting go of staff more than seven layers below the CEO. Dara Khosrowshahi's memo blamed "more layers, more coordination, more fragmented ownership," not AI.
AI made the pitch louder. Gartner predicted that "through 2026, 20% of organizations will use AI to flatten their structures, eliminating more than half of current middle management positions," as SHRM summarized it.
And the data do not say big teams fail by default. Gallup reported in January that the average number of people reporting to a manager rose from 10.9 in 2024 to 12.1 in 2025, while the median held at about five to six, and 13% of managers already oversee 25 or more. Gallup's finding is conditional: larger teams hold up when engagement is high, managers spend less than 40% of their time on individual-contributor work, the manager has real talent for the job, and people get meaningful feedback weekly, which Gallup says "nearly triples" the share of engaged employees.
Wide spans can work if the structure changes with them. James Stanier told LeadDev that at 50:1 "the manager can no longer be the communication hub," so accountability has to sit with Directly Responsible Individuals, communication has to be async-first, and the manager has to be "ruthless about going deep on the two or three most critical things rather than staying shallow across 50."
Not everyone bought it. André Spicer of Bayes Business School, quoted by Fortune in March: "It's going to end in tragedy is the bottom line," warning that junior employees get overlooked and line managers burn out. Meta's reversal does not prove him right, because nobody outside Meta knows why it changed course. But a company does not go looking for managers in a unit it thinks has enough of them.
The Manager Work AI Did Not Absorb
Span of control is the number of people who report directly to one manager, and the right number is set by how much manager-only work each of those people generates. Manager-only work is anything that needs a person with authority: approving a risky change, defending a rating in calibration, closing a hire, owning an incident, breaking a cross-team deadlock. In the best telemetry available, heavier AI use has come with more of at least one of those, not less.
Review got bigger, not smaller. On teams with high AI adoption, developers completed 21% more tasks and merged 98% more pull requests, but PR review time rose 91%, average PR size rose 154% and bugs per developer rose 9%, according to a 2025 study by Faros AI, which sells engineering-analytics software, drawn from telemetry on more than 10,000 developers on 1,255 teams and comparing periods of low and high AI adoption. It is correlational, not a controlled trial. Review is exactly the work that lands on tech leads and managers when a structure goes flat. It is the same pattern as agent teams hitting 65 PRs a week while nobody got time back.
Accountability did not move to the tool. When AI-written code breaks production, a named human still owns the outage and the postmortem.
People work is where flat structures failed last time. In 2002, Larry Page and Sergey Brin eliminated Google's engineering managers. The experiment lasted only a few months; the founders relented "when too many people went directly to Page with questions about expense reports, interpersonal conflicts, and other nitty-gritty issues." Today's tools can draft a status update or summarize a sprint. None of them sits in a calibration meeting defending a rating, mediates between two senior engineers, or accepts accountability for a missed launch.
Gallup adds the constraint most flattening plans skip: managers who spend more than 40% of their time on individual-contributor work have lower engagement, "and this gets worse as the number of workers they manage increases." A 50-person span leaves no room for a player-coach.
Use this as a starting template, not a benchmark:
| Manager work | What AI changed | Leading indicator | Example reversal trigger |
|---|---|---|---|
| Code and design review | More, larger PRs to review | Median time from PR open to first review | Doubles against pre-reorg baseline for two sprints |
| Performance calibration | Drafts write-ups; judgment unchanged | Days from cycle close to final ratings | Ratings ship late or need extra calibration rounds |
| Hiring | Scheduling and screening help | Interview loop completion time | Loops stall for lack of interviewers |
| Incident ownership | Faster triage; same accountability | Incidents closed without a named owner | Any Sev-1 or Sev-2 postmortem with no owner |
| Unblocking and escalation | Summaries of who is blocked | Escalations that skip straight to a director | Directors become the de facto managers |
| Career development | Little | Regretted attrition among senior ICs | Rises above your trailing 12-month rate |
How to Size a Span Without Guessing
Size the span from measured manager hours per report, not from the headcount you want to remove. The arithmetic is simple enough to run in a planning meeting.
Take a 40-hour week. If you hold managers under Gallup's 40% threshold for individual-contributor work, at least 24 hours are left for management. If each direct report needs 90 minutes a week of manager-only work (a 30-minute one-on-one, 45 minutes of review and unblocking, 15 minutes of amortized calibration and hiring), the span comes out at 16. To reach 50, per-report load has to fall below 29 minutes a week.
Those inputs are illustrative. Your own logs will produce different ones, and that is the point: measure, subtract only what tools removed in your own data, then divide. If the answer is 16 and the plan says 40, the gap is not a tooling problem. It means review and unblocking have to move to someone else, which is what Stanier's DRI model does, and those staff engineers and tech leads need that time budgeted in their own plans. Otherwise the load is still there. It just stops showing up on an org chart.
Some teams can go wide: senior engineers, well-specified work, low incident load, strong written ownership. Teams with new hires, junior engineers, heavy on-call or active migrations cannot, and those are the people Spicer warned get overlooked first.
Set the Reversal Trigger Before You Cut
A flattening is far cheaper to undo when the conditions for undoing it were written down before the reorg, with an owner and a review date. After the fact, the people you need are gone or reassigned. Meta is asking former managers to come back, reportedly voluntarily, which only works if enough of them want the job. A company that laid its managers off instead has to rehire, and many that cut for AI have already said they regret it.
There is a legal dimension for CHROs too. In July, 26 Meta employees sued the company over its AI-related layoff practices, Fortune reported. How you choose which managers become individual contributors, and what data you use to choose, is a decision you should expect to defend.
Pick three indicators from the table, record today's baseline, set a threshold for each, and name one person who calls the reversal. Review it after one full performance cycle, because calibration is where hidden manager load surfaces first.
What to Do Before Your 2027 Plan Locks
This Week:
- Pull the span distribution for engineering and flag every manager above 15 and above 25 direct reports. Gallup's 13% at 25-plus is a reference point, not a target.
- Export median time-to-first-review and median PR size for the last two quarters from your Git platform. That is the baseline your reversal trigger measures against.
This Month:
- Have five engineering managers at different spans log one week of manager-only work per report, using the six categories in the table above.
- For every team you plan to widen past 15 reports, name who owns review and unblocking, and put those hours in that person's plan.
- Write the reversal triggers into the reorg document: three indicators, three thresholds, one owner, one review date.
Before Headcount Lock:
- Pilot the wider span in one organization for a full performance cycle before rolling it out company-wide.
- Before moving managers into individual-contributor roles, ask which of them would come back if the structure fails. Meta is having that conversation after the fact.
The Bottom Line on Flattening With AI
Headcount is the easy number; manager load is the one that decides whether the structure holds. Google's no-manager experiment lasted a few months in 2002. Meta set a general limit of 10 direct reports in 2023, built a division at up to 50 in 2026, and six months later is recruiting managers back into it. The tools changed a great deal in between, and in the Faros data heavier use came with a bigger review queue, not a smaller one.
The flattening wave is not wrong. Some of those layers were pure latency. But the 2027 plans that survive will be sized on hours measured, not hours promised.
Cut the layer where the work left with it. Everywhere else, you have only hidden the work.