Drag the slider. See the decay.
Same engineer. Same tools. Different codebase. The X-factor is not a marketing number — it is AI efficiency × Human factor, and both terms shrink as complexity grows. Tune the inputs to match your team's reality.
97% × 97% × 97% × … × 97% = 73.7%95% × 95% × 95% × … × 95% = 59.9%01 · Method
Measured, not modeled
Pulled from 53 production codebases we built or co-built between 2024 and 2026 — greenfield MVPs, scale-ups, legacy rescues. Same toolchain, different terrain.
02 · Definition
Boost = shipped value
We track reviewed, merged, deployed work — not lines generated. Code thrown away, rewritten, or rolled back does not count.
03 · Mechanics
Why the curve bends
Below ~30k LOC AI compounds the engineer. Above ~150k LOC structural forces pull against it: duplicated logic, context rot, review load.
Why your client ships in an evening, and we ship in a quarter.
The honest answer to "why can't your team match what I did at home?" is not "they're worse than you". It is that you and the model worked on a fundamentally different problem. Six concrete differences — all measurable.
Empty repo vs. living codebase.
No review burden. No reviewers.
You + Claude vs. 12 humans + Claude.
2 integrations vs. 47 integrations.
No compliance. No audit. No data migration.
vercel deploy vs. blue-green across 3 envs.
What's actually shrinking the boost?
Six measurable forces — each backed by 2025–2026 research. Together they explain the curve you just dragged through.
Silent duplication inflates the diff
On small projects there is little to duplicate, so AI looks clean. On large codebases it often cannot see what already exists — so it rebuilds the same helper, validator, or fetch wrapper three folders away. Output volume goes up. Real progress does not.
Duplication compounds context rot
Duplication was always a tax. With AI it becomes a multiplier: two copies of the same logic mean two stale fragments in context, two reviewers tracing two call graphs, and two places where the next bug fix has to land.
AI does not know what it does not know
On real platforms the model loads fragments — a file here, a snippet there — and answers with full confidence. The output reads authoritative, but the relevant constraint often lived in a file the model never opened.
More generated code means more review burden
Volume is cheap to produce and expensive to vet. Every generated block has to be read, traced, and understood before it ships — because one confidently wrong block in a critical path is enough. The bottleneck moves from typing to review.
Why seniors on familiar code see ~0 speedup
Our working hypothesis: AI excels at isolated, low-context tasks — a new screen, a clean integration, a one-off script. It struggles where senior developers spend the hard hours: cross-module bugs and product logic that require holding the system in your head.
AI is powerful — when the system is designed for it
The answer is not less AI. It is better boundaries. Close context inside isolated modules, even if only at the logic layer. Plan the architecture, brainstorm with AI, challenge its proposals, cut the bad ones, develop the good ones. Soft skills matter: if a developer cannot articulate the need clearly, they will not vibe-code a great product.
The numbers — straight from the source.
We didn't make this curve up. Every percentage in our model maps to public, peer-reviewed or industry-scale research. Click through and read.