How AI is reshaping software development: lessons from GitHub's CPO

Executive overview

Most developers spend less than 25% of their time writing code — the rest goes to meetings, reviews, legacy code, and collaboration overhead. AI tools like Copilot return time to developers without replacing them; a human remains essential for innovation and system thinking.

The shift isn't from human to AI — it's from writing code to understanding systems.

What's overhyped and underhyped

• Generative AI replacing human developers is overhyped — creativity and innovation require humans in the loop
• AI needs humans to generate the data it depends on; the cycle is symbiotic
• AI-driven testing is underhyped — unit, integration, performance, security, and load testing all need coverage as code output scales
• More code being generated raises the stakes for test coverage, not lowers them

The developer role in an AI-assisted world

• Junior developers can focus on system architecture from day one instead of spending years learning syntax
• Senior developers will set the template: systems thinking, connected architecture, big-picture context
• Hardware development becomes more central — GPU/CPU optimisation is a niche today but will broaden
• AI integration knowledge becomes a baseline expectation for all developers
• Copilot is a co-pilot, not a pilot — the human must remain in the loop

Copilot by the numbers

• Over 1.5 million developers and 37,000 organisations using Copilot
• Developers write code 55% faster on average
• 85% of users felt more confident in their code quality
• Code reviews completed 15% faster
• 88% of users felt less frustrated and more focused
• At Accenture, 88% of suggested code was retained

Design philosophy

• Built developer-first: GitHub engineers and designers used it before any external release
• Core principle: zero friction — if adoption requires effort, developers won't use it
• Features must be seamless and intuitive; adding complexity kills uptake
• GitHub runs on GitHub — every internal team uses the platform, giving real-world signal before launch
• Nothing ships until it has been proven inside GitHub first

Common mistakes when adopting AI

• Treating AI as magic: deploying a tool and expecting instant results without change management
• Starting from "what should we do with AI?" instead of "what problem are we solving?"
• The right sequence: identify the workflow with the most friction, then ask how AI can reduce it
• Working backwards from the customer problem leads to better AI use than plastering AI onto existing surfaces

Measuring success

• No single metric works — productivity, code quality, security, and collaboration must be combined
• Time alone is a flawed metric: fast bad code is worse than slow good code
• Preferred framing: time to value — from task assignment to realised business outcome
• Developer happiness is the ultimate leading indicator; it predicts retention, innovation, and output
• GitHub tracks: secrets prevented from leaking, issues detected before deployment, review cycle time

The future of the field

• Hybrid model landscape: general-purpose LLMs alongside specialised models for high-stakes domains (aerospace, automotive, self-driving)
• Multi-model architectures will emerge — each LLM contributing where it excels
• Sketch-to-product tools are better framed as collaboration aids than production tools; they reduce back-and-forth, not headcount
• Developers will continue choosing their own AI flavour — no single correct way to use the tools

How GitHub runs innovation (GitHub Next)

• GitHub Next is a team of applied and research scientists focused on a 3–5 year horizon
• Their mandate is to invent the future of software development, not solve current-quarter problems
• Copilot originated here
• Key to success: tight feedback loops between Next, product, and engineering — ideas must flow to production
• Failure modes of similar teams elsewhere: becoming too academic (ideas never ship) or too tactical (becoming a second engineering team)
• Innovation at GitHub is organic, not time-boxed — structured creativity kills the instinct

Leadership and career development

• CPO role spans product vision, go-to-market, sales alignment, and engineering collaboration — not just roadmaps
• Influence without authority is the core skill: PMs must bring engineering, revenue, and design along
• Learning from leaders above, peers across, and reports below all contribute to a rounded toolkit
• Biggest early-career lesson: driving change too fast without explaining the why loses the team
• Entering a new organisation with high energy and a fix-everything mindset is a common failure pattern
• The correction: slow down, communicate the why, bridge what you see to what the team has normalised