The Quality Dividend: Why the Real AI Opportunity Isn’t Speed | by Daniel Vaughan | May, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

The Quality Dividend: Why the Real AI Opportunity Isn’t Speed

Daniel Vaughan

13 min read·<br>1 day ago

Listen

Share

‘When the cost of production collapses, the winners are not those who produce more. They are those who produce better.’

Press enter or click to view image in full size

The question nobody is asking<br>Making scrambled eggs takes five minutes. Suppose AI cuts that to one. You have a choice. Make five batches of passable eggs. Or spend the full five minutes on one plate, this time with a world-class chef beside you, guiding your technique, telling you when to lower the heat, when to fold, when to pull the pan. Same five minutes. Completely different result. That chef is what AI can be: not a replacement for your skill, but a guide that elevates it.<br>Most people would choose the excellent eggs. Yet in software engineering, the overwhelming instinct has been the opposite. Use AI to build five projects in a week. Ship more features. Merge more pull requests. Produce, produce, produce.<br>The problem is simple: if you make five batches of eggs, you are not going to eat them all. You throw most of them away. Overproduction is not free. It creates waste in code, in maintenance burden, in cognitive load, in things that exist but serve nobody. Speed without purpose does not fail to add value. It actively destroys it.<br>The compromise between speed and quality has fundamentally changed. Most people have not noticed.<br>What I mean by quality<br>Not gold-plating. Not perfectionism. Not three months of polishing a landing page.<br>Quality is work that does what it claims to do and handles the cases it will actually encounter. It does not silently fail when the world is messier than your assumptions. Specifications thorough enough that the implementer does not have to guess your intent. Tests comprehensive enough to catch regression before your users do. Architecture is considered enough that the next engineer can extend it without first decoding it.<br>Quality is the absence of future pain. Code that does not generate support tickets, specifications that do not spawn ambiguity meetings, and systems that do not wake people at three in the morning. None of this is glamorous. All of it is valuable. Until recently, all of it was expensive enough that most teams rationally chose to skip it.<br>That is what changed.<br>Why speed wins first<br>The instinct to choose speed is not irrational. A 160-year-old economic principle predicts it.<br>In 1865, William Stanley Jevons observed that when James Watt’s steam engine made coal more efficient, Britain did not burn less coal. It burned more. Efficiency created demand. Roads get wider; traffic increases. Bandwidth gets cheaper; we stream higher resolution. This is Jevons Paradox, and it is the dominant narrative for AI in 2025 and 2026.<br>Torsten Slok, Apollo’s chief economist, applied Jevons directly to AI in April 2026: as AI makes professional work cheaper and faster, demand for those services expands rather than contracts. According to a December 2025 Vanguard report, the roughly 100 occupations most exposed to AI automation are outperforming the rest of the labour market in both job growth and real wages.<br>The Faros AI Productivity Paradox Report measured this at scale across 10,000 developers. Developers with high AI adoption complete 21 per cent more tasks and merge 98 per cent more pull requests. More. Faster. Jevons in action.<br>At the organisational level, the same report found no significant correlation between AI adoption and improvements in company-wide throughput, DORA metrics, or quality KPIs. Team-level velocity gains evaporate when aggregated.<br>The speed is real. The value is not.<br>The evidence against speed<br>Some will argue that earlier data measured 2025 tools, and 2026 improvements have already fixed this. They may be right for the narrow question of raw coding speed. The evidence here is not about whether AI can generate code faster. It is about whether faster code generation translates into better outcomes. That question has only grown more pointed as the tools improve.<br>Speed alone does not deliver.<br>JetBrains’ April 2026 telemetry study of 800 developers found a pattern: AI ‘redistributes and reshapes developers’ workflows in ways that often elude their own perceptions.’ Developers reported productivity gains that their actual behaviour did not confirm, with no statistically significant change in debugging patterns despite perceived quality improvements.<br>The Faros report quantifies the downstream cost: 91 per cent longer PR reviews, nine per cent more bugs per developer, and pull requests 154 per cent larger.<br>2025 was the year of AI speed. The bill is arriving in 2026.<br>Through 2025 and into early 2026, AI coding tools were heavily subsidised. GitHub Copilot costs Microsoft money on almost every...