When I say that half of our company’s code is now AI-generated, I don’t mean that we just copy-pasted something from ChatGPT or Claude or that we’re relying on a few auto-suggestions in order to ship to production.

I mean that in the last 6 months, a significant portion of the software we’ve shipped has had some kind of input from tools like GitHub Copilot, ChatGPT, and local LLMs tuned to our codebase.

However, I'll preface by saying that this only possible because the bulk of our application was already written, tested and optimized for the past several years by human developers. I know the codebase better than any non-sentient being ever could, because I know the reason behind every single edit.

But moving forward, the percentage of AI-generated code is only going to move up.

The Turning Point

A few years ago, AI-assisted coding felt like a clever autocomplete. A tool for getting through boilerplate faster or whipping up a quick regex pattern that normal humans can't decipher. Today, it’s something else entirely.

We reached the 50% mark not because we aimed for it, but because AI tools became reliable enough that not using them felt like wasting time. Every developer knows that creating a new simple component encompasses much more than just coding a function.

You start with a function, then add a test suite, then scaffold a whole microservice, verify data, validate input and then continuously polish. And before you know it, your simple component has taken many hours to complete.

AI is now able to work on each step with little interference and with less oversight.

What "AI-Generated" Actually Means

Let's clarify what that 50% represents in a single day of development:

• Boilerplate and scaffolding: Entire front-end layouts, CRUD APIs, and React components are generated in minutes.
• Unit tests: Our AI tools generate 70-80% of our test coverage automatically.
• Refactoring and optimization: We’ve used AI to refactor legacy codebases that hadn’t been touched in years.
• Docs and typing: Adding JSDoc comments, TypeScript annotations, or even inline explanations is now mostly automated.
• Bug Hunting: As a human, I tend to err more often than I'd like to admit. So there are bugs in our codebases, they just so happen to be non-breaking, so nobody is the wiser. But GitHub Copilot can scan my components faster than it takes for me to brew the next cup of coffee.

It’s not perfect, it still hallucinates, gets things subtly wrong, or misses edge cases. But the success rate is high enough to rely on, especially when paired with senior-level review.

The Human Shift

The role of the engineer is changing. We’re not replacing ourselves just yet but we are rethinking what we actually do:

• We design more than we code.
• We debug more than we build.
• We ask better questions, because AI answers are only as good as the prompt.

When you pair an experienced developer with a powerful model, you get productivity that was previously reserved for teams 2–3x the size.

While my time spent writing boilerplate code has gone down dramatically, my time reviewing code has slightly gone up. And this is definitely a trade-off that I'm comfortable making.

Risks and Checks

There are risks, of course. AI can introduce subtle bugs, legal gray areas around code origin, or false confidence in unverified logic. We’ve had to:

• Increase code review rigor Not just for bugs, but for model mistakes.
• Build internal prompt libraries Good prompts become company assets.
• Version-control AI outputs Because reproducibility matters when audits come around.

In short, AI raised the floor and the ceiling, but it also added new responsibilities to my daily to-do list.

I'm not comfortable currently with where AI is at, to fully hand over responsibility to a model that has its good days and bad days. And truth be told, I don't think we're going to get there for quite some time.

I've had many instances where an AI code edit goes rogue and makes adjustments to logic that was crucial for the operation of our business. And that's because it doesn't know any better. It's not sitting in on our 2-hour morning standup meetings and taking notes.

And even if it were, there are still subtle nuances that AI isn't good at picking up on just yet.

What This Means for the Industry

We’re not a FAANG company by any means, just a lean dev team building real products and software solutions. If we’re at 50% AI-generated code today, bigger orgs with more resources are likely on a similar curve, or will be soon.

This isn't hype. It's not the future. It's definitely happening now.

The question isn't whether AI will write your code though.

The question is: How are you going to lead, manage, or innovate in a world where it already does?