AI coding tools promise big productivity gains.
But how much do they really help?
Optimistic findings:
55.8% faster on simple HTTP server tasks
(MIT/Microsoft, 2023)
26% more tasks completed in a fixed time window
(Accenture/GitHub RCT, 2024)
~25% of all code is AI-assisted at Google, Microsoft, Amazon
Sobering findings:
Experienced OSS developers were 19% slower using AI on their own repos
(METR, 2025)
Gains shrink to <10% on high-complexity tasks
(McKinsey, 2024)
84% of developers use AI — but only 29% trust AI output to be accurate
(Stack Overflow, 2025)
In the METR study, developers believed AI made them 20% faster — when it actually made them 19% slower.
That’s a 39-point gap between perception and reality.
Always measure. Don’t guess.
High impact (30–55% time savings)
Low impact or risky (0–10%, sometimes negative)
For low-impact tasks, use AI as a brainstorming partner, not a code generator.
Speed comes at a cost if you’re not careful:
AI-generated PRs have 1.7x more issues than human-written PRs
(CodeRabbit, 2025)
4x growth in code duplication from 2021 to 2024
Refactoring dropped from 25% to <10% of changed lines — devs generate instead of refactor
(GitClear, 2025)
Code revised within 2 weeks grew from 3.1% to 5.7% — more premature commits
AI adoption reduces delivery stability by 7.2% at the org level
(Google DORA, 2024)
More code ≠ better code.
Speed without discipline creates debt.
The 2025 DORA Report found the most important thing about AI in teams:
AI doesn’t fix broken processes. It accelerates them.
Delivery (DORA metrics)
Quality
Team Health
Avoid vanity metrics:
Developer cost: €80,000/year (~€50/hour)
Work hours per month: 160h
Realistic AI uplift: 20–30% on routine tasks (conservative)
Routine task share: ~50% of work
Hours saved per month = 160h × 0.50 × 0.25 = 20h
Value of saved time = 20h × €50 = €1,000
Tool cost per month = ~€20–40
Net benefit per dev ≈ €960–980/month
Even conservative estimates pay for the tools 25–50× over.
But only if quality doesn’t regress — otherwise you spend the saved time fixing AI-generated bugs.
Ask quarterly:
If the answer to several of these is “no” — the problem is workflow and discipline, not the tools.