Slide contents

1. 1. AI-Native Development in Big Tech in 2026

   Microsoft, GitHub, Google, Amazon, Meta, Uber, Stripe, Netflix

2. 2. 00. AI Joins Engineering Process

3. 3. A New Operating Model

   AI joins delivery system

   AI becomes operating-model layer

   Metrics move beyond adoption

   DORA: 90% AI use; 80%+ gains

   Processes matter more than permission

4. 4. Metrics Are Evolving

   'How many lines did AI write' no longer works

   Early metrics are gameable

   GitHub: adoption → PR activity

   AWS: watch delivery system

   Focus: velocity and economics

5. 5. 01. Microsoft & GitHub

6. 6. AI Code Review at Company-Wide Scale

   Source: Microsoft Engineering

   90%+ of PRs covered

   600,000+ PRs per month

   10–20% faster PR completion

   AI sits in normal PR workflow

7. 7. Coding Agent as Part of the Platform

   Agent asynchronously opens draft PRs.

   Metrics GA: adoption, PR, CLI, tokens.

   AI-native is managed platform capability.

   Observability layer for AI adoption

8. 8. 02. Google

9. 9. Velocity as the Key KPI

   Source: Google AI workplace examples

   30% — Of new code is generated with AI assistance

   10% — Estimated engineering velocity improvement

   12% — Duplicate bugs automatically handled by an AI agent

10. 10. Key Principle

    Humans still approve

    AI accelerates the whole lifecycle

    Review, tests, migrations

    The focus is velocity, not AI-code share

11. 11. 03. Amazon & AWS

12. 12. The Language of Economics

    Source: AWS DevOps Blog

    450,000+ hours — Reduced waiting for technical answers via Amazon Q

    4,500 dev-years — Saved on AI-driven software transformations

    15.9% — Year-over-year cost reduction in 2024 through end-to-end improvements

13. 13. CTS-SW: Cost to Serve Software

    Source: Amazon Science + AWS Enterprise Strategy

    DevEx and AI connect to ROIC

    Goal: reduce cost and friction

    AI goals use economic language

14. 14. 04. Meta

15. 15. Ranking Engineer Agent (REA)

    Source: Meta Engineering

    Generates hypotheses and jobs

    Humans set scope, budget, and quality criteria

    2× accuracy over baseline

    5× engineering output

16. 16. 05. Uber

17. 17. uReview: Production Scale

    Source: Uber Engineering

    90%+ — Of weekly ~65K diffs analyzed by AI

    75% — Of comments rated useful by engineers

    ~1,500 hrs/wk — Estimated time savings

18. 18. AI Across All SDLC Stages

    Source: Principal Engineer — AI Tools

    AI covers the whole SDLC

    Metrics: velocity, review, tests

    Outcome: more feature launches

    Planning language for tech leads

19. 19. 06. Stripe & Netflix

20. 20. Goal: 2× Engineering Velocity via LLMs

    Source: Stripe benchmark + hiring

    Benchmark: real integrations

    Deterministic evals and e2e

    Measured by closed scenarios

    Main asset: eval system

21. 21. Platform → Evals → Scaling

    Source: Anthropic + Netflix

    Centralized infrastructure

    Strict evaluation frameworks

    Context quality comes first

    Then evals and scaling

22. 22. 07. AI-native planning has four layers

23. 23. Adoption → Throughput → Quality → Economics

    Adoption — Users, agents, teams

    Throughput — PR cycle time, time to merge, engineering velocity

    Quality / Risk — Comments, defect prevention, tests, incident frequency

24. 24. 08. AI-Native Becomes Operating Model

25. 25. Key Takeaways

    Humans become orchestrators

    Key question: delegated scenarios, flow impact, economics

    AI becomes a standard participant

    AI-native becomes operating model

26. 26. References and Materials

    Key sources used in this presentation

    DORA; Microsoft; GitHub.

    Google; Amazon CTS-SW.

    Meta REA; Uber uReview.

    Stripe benchmark; Netflix scaling.

27. 27. Thank You!

    AI-Native Development in Big Tech

    For more materials on this topic, visit the "Book Cube" channel — all links are collected there

    Alexander Polomodov, Technical Director & Fellow, T-Technologies

    @Book_Cube