AI WORLD HQ
Picture someone on a software team staring at a pile of old testing code that nobody wants to touch. It is the kind of project that lives forever on "someday." Then one afternoon, an AI assistant finishes it. Roughly that is what just happened at Asana.
What actually happened at Asana
Asana is the work-management tool many teams already use to track tasks and projects. Internally, the company had a testing system — software that automatically checks whether a new feature accidentally breaks something older — that had grown old, slow, and painful to maintain.
The team's estimate: about five years of engineering work to rebuild it the normal way. With help from OpenAI's Codex (an AI model tuned to read, write, and modify code), they finished the rebuild in roughly two weeks, at a cost around $12,000.
That gap — five years versus two weeks — is the headline. But the real story is what kind of work the AI was asked to do.
So what is Codex, in plain language?
A model in AI is the trained "brain" that produces answers. Codex is OpenAI's model specifically tuned for programming. It belongs to the same family of technology that powers GitHub Copilot, the helpful autocomplete you may have already seen inside a code editor.
What changed recently is the size of job these assistants can take on. A few years ago, Codex-style tools mostly wrote small functions when you asked. Now they can take on an entire messy legacy project: reading old code, suggesting rewrites, running checks, and even validating their own work before handing it back.
A few terms you will see around this:
- Code is just instructions written in a language humans can also read. It tells software what to do.
- Testing is code that checks other code automatically — a stand-in for a careful human reviewer.
- Refactoring means cleaning up old code without changing what it does for the user. Think of it as reorganising a kitchen so you can actually find the spices.
Why this reaches beyond software teams
The lesson is not really about coding. It is about the pattern: AI is getting good at the tedious backlog work that nobody enjoys.
If a model can rewrite five-year-old testing code in two weeks, the same shape of job — long, dull, and structured — exists in writing, customer support, accounting, legal research, and data cleanup. The tools are catching up across the board, at different speeds.
That does not mean whole jobs vanish overnight. It means the boring part of many jobs is moving first.
Wrap-up
A five-year project cleared in two weeks is striking, but the more useful takeaway is the pattern behind it. AI assistants are moving up the stack, from writing a snippet to running an entire project. Pick one repetitive task in your own week and hand it to an AI today — that is the smallest realistic starting point, and it will tell you more about where this is heading than any headline.