You accept an AI suggestion, commit it and push. Two weeks later production breaks; someone runs git blame, and the name on that line is yours.

AI now writes a big share of our code, faster than we can, and if we keep measuring developers by output, we lose to the machine every time.

What AI can’t do is answer for the work. Air Canada learned this when it told a tribunal its chatbot was responsible for its own words, and lost.

I build AI systems for Indian MSMEs. In one of them, factory staff send updates on WhatsApp and an AI turns them into records. We let the system drop messages it wasn’t confident about, and a staff member’s update vanished with no error. We found out only when someone at the client’s office asked why it was missing. I’ll share how we fixed it, and why the AI’s decisions need tests as much as the code it writes.

You’ll leave with SIGN, a four-step way to own AI-assisted work (Scope, Inspect, Guard, Name), and templates for the CODEOWNERS and AGENTS.md files.