The companies announcing “AI-first” org changes in 2025 divided into two groups. One moved fast, cut deep, and is now quietly rehiring. The other is still running the same pyramid org chart, calling Copilot access a transformation.

Neither is right. The actual structural shift is less dramatic and more durable than either camp admits.

The pyramid is the right starting point

Traditional engineering orgs look like pyramids. Lots of junior engineers at the base, a mid-tier of seniors and leads, a handful of architects and managers at the top. Entry-level roles feed the machine: they do the scoped tickets, the boilerplate, the translation between spec and code.

AI can do most of that work now. Not all of it, but enough that the economics of the base are changing. Harvard Business Review reported in January 2026 that two-thirds of companies are already slowing entry-level engineering hiring. This is not a prediction. It is happening.

But the response to this shift has been almost universally wrong.

What went wrong at Klarna

Klarna replaced 700 customer service workers with an AI assistant in 2024. The CEO announced savings of roughly $40 million. Investors cheered. Then CSAT dropped on complex cases: multi-step billing disputes, fraud escalations, anything requiring judgment that wasn’t baked into training data. By early 2026, Klarna was rebuilding its human support capacity, now structured as a hybrid where AI handles volume and humans handle escalations.

The mistake was assuming all interactions are equivalent. Sixty to seventy percent of queries are routine. Fifteen percent require judgment. Klarna cut the workforce before it built the infrastructure to tell them apart.

Forrester’s Predictions 2026 report found 55% of employers who laid off for AI now regret it. Forrester predicts half of those cuts will be rehired, offshore or at lower pay, but rehired.

What Shopify got right

Shopify went in a different direction. CEO Tobi Lütke declared in April 2025 that AI use is a baseline expectation across all roles and that teams must show AI cannot do a job before requesting headcount. That is not a layoff policy. It is a leverage policy.

The operational details matter. Shopify built a centralized LLM proxy, one internal gateway routing all AI requests, so engineering leadership can see usage, compare costs, and swap models without disrupting engineers. They moved from measuring lines of code and pull requests to weekly demos. They made comprehension a hard requirement: engineers need to understand systems two to three levels below where they are working, because AI will generate code in those layers whether the engineer understands them or not.

The result: Shopify’s VP of Engineering estimates a 20% productivity gain from the shift. Not headcount reduction. Output increase with existing headcount.

The shape that’s actually emerging

PwC called the new structure a diamond: small leadership tier, strong middle layer of seniors and leads, narrower base of entry-level talent. The “No More Pyramids” analysis published in 2026 argues that AI absorbs the base, repurposes the middle for agent oversight, and makes the leadership tier smaller but higher-leverage.

That is about right in shape, but wrong about where the pressure lands. The middle layer is where the real friction accumulates. Seniors and leads are now expected to:

  • Write specs precise enough that agents can execute without constant correction
  • Review agent-generated code they did not write and have no intuition about
  • Maintain quality across outputs they can no longer inspect line by line

PR review time increased 91% in high-AI-adoption teams. Bugs per developer went up 9%. The bottleneck moved from generating code to verifying it. Most engineering orgs have not moved their senior layer to match.

The Pragmatic Engineer’s 2026 survey found something telling: engineering managers are becoming more hands-on with technical work, not less. As junior output floods in through AI-assisted generation, someone has to catch what is wrong. That someone is increasingly the senior engineer or the lead. Sometimes the manager.

What the org chart actually needs to change

The right structural moves are smaller and more boring than the AI-first announcements suggest.

Stop defaulting to headcount when a team is slow. Before you hire, diagnose whether the bottleneck is generation speed (almost certainly not) or spec quality, review throughput, or architectural clarity. Hiring a junior engineer to solve a spec problem makes things worse.

Raise the floor on spec quality. If agents are doing more of the writing, humans have to do more of the thinking. A vague two-paragraph brief produces a plausible-looking wrong implementation. Precision up front is where senior time now earns its keep.

Build the review layer deliberately. Faster generation without faster review creates a merge jam. Teams that jumped ahead on AI adoption without redesigning review are the ones seeing rising bug rates and incident counts. Review is not reading every line; it is knowing what to look for and having the system context to catch it.

Hold on entry-level hiring, but hire better. The roles have not disappeared. They shifted from writing boilerplate to verifying output, understanding why the agent’s implementation is wrong, and developing the judgment that eventually becomes senior capability. That takes time and structure. Hire for curiosity and systems thinking, not typing speed.

The org that emerges from this isn’t smaller. It’s differently weighted — fewer people generating code from scratch, more people owning the quality and direction of what the agents produce. Companies that treat this as a headcount reduction story will find out, like Klarna did, that they cut the judgment they still needed.

If you are trying to figure out how your engineering team should be structured for this, talk to us.

Sources