The Delivery Gap: why 96% of your AI code is waste

AI coding tools increased PR volume by 98% — but net delivery throughput? Zero. In this episode, we unpack why individual AI productivity gains fail to become organizational outcomes, and what engineering leaders must do differently. In this episode, Romain sits down with Brenn Hill, Senior Engineering Manager at Delivery Hero, to discuss his book "The Delivery Gap" and the growing chasm between AI-generated code velocity and actual business delivery. 🔑 Key topics covered: Why 96% of AI-generated PRs can become organizational waste The "verification triangle" — your real speed limiter Cost per accepted change: the metric that exposes hidden waste Why specs are alignment documents, not source code The 80% illusion: AI prototypes that hide fundamental problems Agentic security risks most teams aren't prepared for Why "smaller, tighter agents" always outperform large ones The build vs. buy pendulum in the AI era Human-in-the-loop UX as an unsolved problem ⏱️ Timestamps: 00:00 — Intro & guest background 01:10 — What is Delivery Hero & HeroGen 03:20 — The journey: from AI Augmented Development to The Delivery Gap 04:09 — "It's too easy to write software with Gen AI" 07:10 — AI as an amplifier of engineering culture 09:18 — 100 PRs but only 5 merge: the waste problem 10:12 — FAROS report: 800% more defects 11:32 — "You're going 5x faster into a wall" 12:13 — The 80% illusion and rapid prototyping 14:03 — "Save me, I vibe coded this thing" 15:58 — Spec-driven development: alignment, not source code 19:23 — "We already have a language for telling computers what to do — it's called code" 23:13 — The CTO's transformation journey 25:01 — The verification triangle framework 28:57 — Cost per accepted change metric 31:00 — Theory of constraints applied to software delivery 35:16 — Token budgets and where verification costs go 36:13 — Cost per accepted action for agents 37:25 — "Smallest, tightest, most precisely aimed agent" 39:27 — MCP servers as cost multipliers and attack surfaces 41:16 — Shifting verification left: inner loop strategies 43:02 — AI-specific code review: testability, correctness, security 47:22 — Human-in-the-loop UX is broken 49:00 — Tiering your changes for review efficiency 52:33 — "Fix everything. You've got eight hours. Go nuts." 53:32 — What's still broken in AI development 58:37 — Build vs. buy in the AI era & European sovereignty 63:01 — Closing thoughts & recommendations 🔗 Resources mentioned: The Delivery Gap (book): [https://www.amazon.com/dp/B0GWRY2XH1](https://www.amazon.com/dp/B0GWRY2XH1) | [https://thedeliverygap.com/](https://thedeliverygap.com/) FAROS AI Acceleration Whiplash Report 2026: [https://www.faros.ai/research/ai-acce...](https://www.faros.ai/research/ai-acce...) HeroGen by Delivery Hero: [https://businesswire.com/news/home/20...](https://businesswire.com/news/home/20...) Warp Terminal (open-source): [https://www.warp.dev/](https://www.warp.dev/) The Phoenix Project (book): [https://itrevolution.com/product/the-...](https://itrevolution.com/product/the-...) The DevOps Handbook: [https://itrevolution.com/product/the-...](https://itrevolution.com/product/the-...) AWS DevOps Agent (previous episode): \[LINK TO EP 216\] Kiro (spec-driven development): [https://kiro.dev/](https://kiro.dev/) 👤 About the guest: Brenn Hill is a Senior Engineering Manager at Delivery Hero in Berlin, leading data science, ML engineering, and software engineering teams. With 17 years in software and experience from Vietnam's first unicorn to a DAX-listed company, he authored "The Delivery Gap" — a framework for engineering leaders navigating AI adoption. LinkedIn: [  / brennhill  ](  / brennhill  ) Book: [https://thedeliverygap.com/](https://thedeliverygap.com/) 🎙️ About the AWS Developers Podcast: Weekly conversations about Agentic AI and the future of software development with customers, AWS community members, and industry luminaries. Hosted by Romain Jourdan, AWS Chief Technologist Europe-Central.