AI writes over 75% of production code at Google DeepMind, yet about 88% of pull requests still need human steering

One of the most AI heavy engineering teams in the world still has a human steering almost nine in ten changes.

What happened

Google DeepMind's safety report for Gemini 4 includes a section on how much of its own work is automated, which The Decoder covered after Google's Haydn Belfield shared the findings.

The report says most production code at Google DeepMind is now more than 75% AI written. The share of fully delegated pull requests rose from 7.5% in April to 11.9% in September 2026. But about 88% of pull requests still require iterative, multi turn human steering or human code authorship, and code changes remain subject to human review before submission.

The team also looked at a sample of 1,389 Argon agent runs from internal use. The pattern was clear: the bigger the task, the less likely the model completed it on its own and the more likely a human stepped in. A survey of researchers found automation highest on well specified, verifiable tasks and lowest on open ended research. On Google's internal ML engineering benchmark, Gemini 4 Argon solves around half of the tasks. According to The Decoder, Belfield says there is no measurable acceleration of AI research driven by AI itself.

My take

This is the most honest automation data I have seen from a frontier lab, and it lines up with what I see in client work.

AI can do most of the typing. It does not yet do most of the deciding. The work that gets fully handed off is small, clearly defined and easy to check.

So when a business asks me to "automate the sales team" or "replace the VA", I break it down using the same lens:

1. Is the task well specified? Clear inputs, clear rules, a known shape for the output.
2. Is it verifiable? Can a check confirm it was done right, like a record that exists, a total that matches, a field that is filled?
3. How big is it? Long chains of decisions are where agents drift and humans end up stepping in.

Tasks that pass all three are candidates for full delegation. Everything else gets an AI draft plus a human approval step, and that is still a large time saving.

Track it like Google does. Count how many runs finish with no human touch, and watch that number grow before you remove the review step.