The short version
AI coding tools work. The org chart doesn't. Cursor and Claude Code boosted developer output but org-level velocity stayed flat because the gains get swallowed by review queues and coordination meetings. CodeRabbit's data: AI PRs are 18% larger and take 3.6x longer to review. Logilica: 21% more tasks shipped but PR review time up 91%. Fix three layers in order. One: get every engineer across the AI-adoption line in 90 days. Two: rebuild review for AI speed (move humans from line-by-line review to specification and verification). Three: flatten the coordination layer (the relay middle managers exist to do is exactly what agents replace). The subscription works. The org chart is the cost center.
Your company spent six figures on AI coding tools this year. Developer output went up. Nothing else got faster.
That's the pattern I keep seeing. The tools work. Cursor, Claude Code, GitHub Copilot. They do what they promise. Engineers produce more code, faster. But the org around them hasn't caught up, and the gains are getting swallowed by the same meetings, reviews, and approval chains that existed before AI showed up.
Shopify CEO Tobi Lütke put it bluntly in his April 2025 memo: before you hire a single new person, prove that AI can't do the job first. That's not a cost-cutting play. That's a structural bet. He's saying the org chart itself is the liability.
He's right. Here are three layers to fix, in order.
Layer 1: Get your engineers across the line
This is the easiest layer and most companies are still fumbling it.
Coinbase gave their engineering team one week to adopt Cursor in August 2025. CEO Brian Armstrong bought licenses for every developer and made adoption non-optional. People who refused were let go. By February, they hit 100% adoption.
That timeline was aggressive. A quarter is more reasonable. But the principle is correct: AI-assisted development is not optional enrichment. It's the new baseline.
Here's the trap, though. METR's 2025 study measured experienced open-source developers on real tasks and found they were 19% slower with AI tools, while simultaneously believing they were 20% faster. The 2025 Stack Overflow Developer Survey backs this up. Developers report positive sentiment about AI tools, but measured productivity gains are inconsistent and depend heavily on task type and experience level.
The gap between "I tried it once and it felt slow" and "I can't work without it" is roughly three weeks of focused practice. Those weeks need to be protected time, not wedged between sprint commitments and on-call rotations.
What to do this week:
- Set a 90-day deadline for AI-assisted development adoption across your engineering team. No exceptions.
- Protect transition time. Block two hours daily for the first three weeks. Let people struggle with the tools before measuring output.
- Measure after the quarter, not after one pairing session. Track PRs shipped per engineer, not vibes.
- Make it a hiring criterion going forward. If Shopify requires AI proficiency as a "fundamental expectation," you should too.
Layer 2: Rebuild your process for AI speed
This is where most of the value is stuck.
CodeRabbit's December 2025 analysis of 470 GitHub repositories tells the story: AI-generated pull requests are 18% larger, contain 1.7x more issues per PR, and take 3.6x longer to review (4.3 minutes vs 1.2 minutes for human-written code). Logilica's research across thousands of teams found that while developers complete 21% more tasks with AI, PR review time jumps 91%.
Read that again. Your agents ship code in 20 minutes. Then it sits in a review queue for two days because a human is checking every line as if another human wrote it.
The review process was designed for a world where writing code was the slow part. Writing is now the fast part. Review is the bottleneck.
The fix: move humans from line-by-line code review to specification (defining what to build precisely) and verification (confirming it works correctly). Agoda's engineering team calls this the grey box model. Fred Brooks identified the same insight in 1986: the hard part of software was never typing it. It was specifying it correctly and verifying it works.
I wrote about a version of this shift in The PM-as-Translator Is Dead. The old model of reformatting customer input into decks for engineers doesn't survive when an agent can go from spec to working PR in minutes. The human value moves upstream to judgment, not downstream to translation.
Standups need the same rethink. When an agent ships three PRs overnight, "what did you do yesterday?" is the wrong question. "What decision does the agent need from you today?" is the right one. I covered how to automate the status-reporting side of this in Stakeholder Updates on Autopilot. The agents handle the reporting. Humans handle the decisions.
What to do this week:
- Audit your average PR review time. If it's over 24 hours, your process is the bottleneck, not your developers.
- Split code review into two tracks: automated checks (linting, tests, type safety) and human judgment (architecture decisions, security implications, product correctness). AI PRs get the automated track by default.
- Rewrite your standup format. Replace "what did you do?" with "what's blocked?" and "what decision is needed?" Try it for two weeks.
- Kill one recurring meeting this week that exists only to transfer status. Replace it with an agent that pulls from Jira, GitHub, and Slack and posts a summary.
Layer 3: Flatten the coordination layer
This is the expensive one. And it's already happening around you.
The data is piling up fast:
- Gartner predicts 20% of organizations will use AI to eliminate more than half their middle management by 2026.
- Amazon CEO Andy Jassy wrote in his 2025 shareholder letter that AI enables "significantly fewer managers and layers." Amazon cut 14,000 corporate roles targeting the IC-to-manager ratio.
- Google cut 35% of small-team managers in 12 months.
- Meta's "Flatter is Faster" initiative capped direct reports at 10 and converted managers back to individual contributors.
- Jack Dorsey and Roelof Botha published From Hierarchy to Intelligence via Sequoia Capital, arguing that 2,000 years of hierarchical management is ending. Block laid off 40% of staff in February 2026.
These are not isolated headcount reductions. They are the same structural conclusion reached independently: the coordination layer between vision and execution is the most expensive, slowest, and most lossy part of the system. AI just made it optional.
Think about why middle management exists. Information doesn't flow well in large organizations. Managers are human routers. They collect status from below, filter it, reformat it, pass it up. They take direction from above, translate it, add context, push it down. Every hop adds delay and loses signal.
An agent that reads your project tracker, your repo, your team chat, and your analytics dashboard knows where things stand in real time. It doesn't need a weekly status meeting. It doesn't need a manager to ask "are we on track?" It can answer that question with data, at any time, to anyone who asks.
Research backs this up: knowledge workers spend 35-50% of their time on coordination, not creation. Engineers average 10.9 hours per week in meetings. For managers, it's 18 hours. That's not productivity. That's overhead.
What to do this quarter:
- Map your information flow. For every recurring meeting, ask: "Does this meeting exist to make a decision, or to transfer information?" If it's information transfer, automate it.
- Identify your human routers. Which managers spend more than half their time aggregating and reformatting status? That's the role AI replaces first.
- Build an agent dashboard. Connect Jira, GitHub, and Slack to an AI that answers "are we on track?" in real time. I run this with my own agent fleet. The Product Ops Agent and Executive Report Agent are good starting points.
- Push decision authority down. When the coordination layer thins, the people closest to the customer and the code need the authority to act. Small teams of three to five with clear outcome ownership, supported by agents, not managed through layers.
This is not "fire everyone and buy ChatGPT"
Klarna tried wholesale replacement in 2024, cutting their workforce from 7,400 to 3,000 and deploying an AI chatbot equivalent to 800 customer service agents. Customer complaints forced them to reverse course in 2025. The institutional knowledge, the judgment calls, the relationship nuance, those don't transfer to a model.
The distinction matters: remove the relay layer, not the judgment layer.
In the old org, a senior engineer's insight filtered through a team lead, then a manager, then a director before reaching the VP who could act on it. Each layer added its own interpretation and political considerations. By the time it arrived, it was three weeks old and barely recognizable.
In the new org, that engineer's judgment goes directly into a system the VP can query. "What's the biggest technical risk in the Q3 launch?" The answer comes from the people closest to the code, synthesized by an agent, without the game of telephone.
This is the same shift I described in The AI Product Engineer: one person doing what used to require a team, not because people are unnecessary, but because the translation layers between them are. The humans who remain should be in high-bandwidth collaboration, not isolated by process. That's why I've moved to mob prototyping: one day, one room, PM plus designer plus engineer building a working prototype together. The coordination overhead disappears because everyone is in the same room reacting to the same artifact. Agents handle the status reporting. Humans handle the creative problem-solving.
The new structure looks like this
Leadership owns the vision and the customer signal. This gets more important, not less. When coordination layers thin out, the quality of the input matters more because fewer people are adding interpretation. Bad vision propagates faster.
Small teams execute with agents. Three to five people with clear ownership of an outcome, supported by agents that handle status, scheduling, and mechanical overhead. No project managers tracking timelines. No program managers aligning dependencies. The PRD is dead. The agents do the coordination.
Agents absorb the overhead. Status reporting, dependency tracking, resource allocation, progress monitoring, risk flagging. All automatable with tools that exist today.
What to do Monday morning
Stop optimizing developer productivity. Start restructuring the company around it.
The AI subscription is a rounding error. The org chart is the real cost center. Every manager-as-router, every alignment meeting, every status report that exists because information doesn't flow naturally, that's where the budget is.
Here's your audit checklist:
- Count your coordination roles. How many people in your org exist primarily to move information between other people? That number should shrink by half within 18 months.
- Count your status meetings. How many hours per week does your team spend reporting status vs. making decisions? Automate the reporting. Keep the decisions.
- Count your review bottlenecks. What's your average time from PR creation to merge? If AI produces code 10x faster but review takes 3x longer, you've moved the bottleneck, not removed it.
- Ask the uncomfortable question. If you removed your coordination layer tomorrow, would the result be chaos, or would it be faster? If chaos, that tells you your information infrastructure is the real problem. Fix that first.
The companies restructuring now will outrun the ones still debating tool licenses. The subscription works. The org chart doesn't. Fix the org chart.
Sources:
- METR Early-2025 Developer Productivity Study, experienced OSS devs 19% slower with AI tools, perceived 20% faster
- 2025 Stack Overflow Developer Survey, AI-authored code at 26.9% of production, mixed developer sentiment
- CodeRabbit AI Pull Request Analysis (Dec 2025), AI PRs 18% larger, 1.7x more issues, 3.6x review time
- Logilica: The Shifting Bottleneck, 21% more tasks completed, 91% longer review time
- Gartner Top Predictions 2025-2026, 20% of orgs will cut 50%+ of middle management
- From Hierarchy to Intelligence (Dorsey/Botha, Sequoia Capital), structural argument against hierarchical management
- Shopify CEO AI-First Hiring Memo (CNBC), "prove AI can't do it before hiring"
- Coinbase 100% Cursor Mandate (Fortune), full adoption in under 6 months
- Klarna AI Reversal (CX Dive), 7,400 to 3,000, then partial reversal after customer complaints
- Fred Brooks, No Silver Bullet (1986), essential vs. accidental complexity in software