AI Coding Kills Middle Managers First, Not Juniors or Seniors
There's a through-line in the last two years of big tech reorgs that most of the commentary missed: the layoffs didn't land evenly. They landed hardest on the middle management layer.
- Meta kicked off its "Year of Efficiency" in 2023, and Zuckerberg kept using the word flatter in public comms. The goal was to thin out management levels.
- Google publicly acknowledged cutting a batch of VP-level and middle-management roles in 2024, with the official line being "fewer layers, faster decisions."
- Amazon cut over a thousand manager slots in the same window, and Andy Jassy said flat out in an internal memo that management growing faster than the people doing the work is a problem.
All of this happened before AI Coding landed at scale. Which means "cut the middle" was already the structural move big tech was heading toward. AI just turned it from optional into mandatory.
Why the middle? Because a middle manager's value never came from code. It came from routing information and coordinating decisions — the two things AI has replaced most efficiently in the last 18 months.
Two parts below: first, how unevenly the AI dividend is distributed across senior IC / junior / middle manager, then the three specific mechanisms squeezing the middle.
The AI Dividend Is Wildly Uneven Across the Three Layers
Senior ICs: AI is a multiplier on output.
NBER w31161 (Brynjolfsson, Li, Raymond 2023), the Microsoft Copilot RCT (Peng et al. 2023), the GitHub Copilot field experiment (Cui et al. 2024), METR's 2025 study on open-source maintainers — these don't all point the same way, but there's one clear intersection: people with strong judgment, who can decompose a task and review the output, capture significantly more from AI than anyone else. Staff+ engineers in public interviews basically all say their throughput multiplied. Goldman's public comms around deploying Devin say the same thing in different words.
Junior engineers: growth curve deformed, but the ticket in still exists.
Mid-level engineers reclaimed the time they used to spend mentoring, and juniors' ramp to independent output stretched out. That's a chronic injury, not an acute one — it hasn't shown up as mass junior layoffs yet.
The TL/EM layer in the middle: all three core job functions pulled out at once.
Task allocation, review, cross-team alignment. For the last decade those three were a TL's whole reason to exist. In 18 months AI moved every one of them from "must go through the TL" to "can route around the TL."
Mechanism 1: ICs Route Around Task Allocation Entirely
The TL used to be the strongest coder in the group. A junior with a request went to the TL first to get the work broken down.
The reality now: a senior IC forks three agent sessions, PRD in, PRs out, and the TL is never in the loop at all.
The damage here isn't "we saved one step." It's that the TL loses their sense of position in the workflow. Once an IC can take a PRD, spin up agents, and ship PRs alone, the TL either becomes the person who gets notified afterward or squeezes themselves back into the IC track voluntarily. Both endings dilute the role.
Mechanism 2: Review Authority Drowns in the PR Flood
GitClear's 2025 AI Copilot Code Quality report is blunt about the direction: duplicated code blocks up sharply year-over-year, short-term churn up, refactoring share down. DORA's 2024 Accelerate State of DevOps says it too — delivery stability for AI-adopting teams is negative, while the same respondents self-report a slight improvement in review speed.
Read those two side by side and one thing falls out: AI massively amplified PR volume, and human review capacity didn't expand at all.
A TL facing that flood has three options, and every one of them drains the role's authority:
- Read them all personally: backlog stretches to weeks, downstream ICs complain.
- Approve without comments: the signature stops meaning anything, and the ICs know it perfectly well.
- Hand it to an AI reviewer (Greptile / Bito / CodeRabbit): the role itself just got replaced by a SaaS subscription.
The quieter collapse: juniors stopped asking the TL anything. Getting stuck used to mean interrupting the TL — that was the main entry point for a TL's informal authority. Now a junior opens Claude, solves it in five minutes, and never interrupts anyone. The TL loses the "person who knows this code best" label, and with it the doorway to being trusted.
Mechanism 3: Alignment Got Replaced by Specs
Cross-team alignment was the TL's core move for a decade: call the meeting, present the approach, extract the commitment.
Spec-driven AI Coding compressed all of that into a document — AGENTS.md, an 8-line spec, shared context. Once the spec is clear, anyone can read it and anyone can run it. The alignment meeting stops being necessary as a ritual.
The cost to a TL: weekly syncs and cross-team standups, all those meetings named after "alignment," used to be part of what the salary paid for. When the marginal cost of alignment drops to roughly zero, that part of the compensation stops being justified.
Two Paths Left for the Middle
If you're sitting at TL/EM right now, AI leaves you two choices.
Path 1: Collapse back to IC — senior IC commanding multiple agents.
This is the dignified turn. The judgment you built over 5-10 years — which design will blow up, which abstraction becomes a liability, when to rewrite instead of patch — is exactly what AI doesn't have. Make it explicit, wire it into specs plus agents, and push your own throughput to several times a staff IC's.
Path 2: Expand span upward — from 6-person TL to 15-20 person EM.
Move toward being an organizational amplifier. Less allocation / review / alignment, more talent identification, direction tradeoffs, and pulling resources across departments — the things that need organizational trust and that AI can't take over anytime soon.
The people standing still — a TL micromanaging six people, personally reviewing PRs, personally running alignment meetings — are the first thing merged away in the next round of org optimization.
A Self-Check for Every TL
Take last week apart and do the arithmetic:
How much time went to coordination, judgment, and direction that AI can't replace?
Under 30% and your role is depreciating in real time. Under 15% and you'll probably show up on next quarter's optimization list.
And if you just got promoted to TL — congratulations, you got handed the chair with the highest odds of being washed away.