At one mid-sized software company tracked by researchers over two years, AI-authored code changes grew to nearly 90% of volume. Merge rates held steady and cycle times improved. The share of changes that received a written comment from a human reviewer fell from roughly 39% to 21%.
That second number matters more than it looks, because review comments are mostly not about catching bugs. Decades of research on review practice find that the most common feedback concerns structure, documentation, placement, alternative approaches: interpretive judgments about what the code should become. When a senior engineer writes "have you considered doing this differently," she is transferring architectural understanding to someone who didn't have it. Nobody budgeted for that transfer. It happened because she was there, reading the code, while doing something else.
In July 2025, Commonwealth Bank of Australia announced that an AI voice bot justified eliminating 45 customer-service roles. The bot's metric said call demand had dropped. The workers said the opposite: volumes rising, managers offering overtime, team leaders pulled back onto the phones. The bank reversed the decision the following month.
Two signals pointed in opposite directions, and only one of them was instrumented. The other survived because the employees were still on the floor to carry it and the Finance Sector Union gave them somewhere to carry it to. A few months later, there would have been no one holding it.
What links the two cases is something organizations have never had to account for. People doing routine work produce two outputs. One is the output the job was designed around: processed claims, reviewed code, answered calls. The other arrives free, as a byproduct of attention — pattern recognition, interpretive feedback, a sense of how today compares with last Tuesday. A branch employee who notices that a loan applicant's circumstances don't fit the online form, and moves the case into a manual process, is exercising a discretion that exists only because she is there, doing the routine thing, with enough context to see when the routine doesn't apply. Workarounds signalled where a process had quietly broken. Complaints revealed what a policy actually did to actual people. None of it appeared in a job description. Together it amounted to a kind of organizational peripheral vision: not something anyone looked at directly, but something that shaped what the institution believed was happening.
Peripheral vision is hard to value because its absence is not an event. By the time anyone goes looking, the people who carried it have moved on, and the incentive structure that removed their positions has no line for what it removed.
Where that leads over the next few years deserves more attention than it gets. Those ambient signals weren't only useful information. They were the main way organizations encountered problems they hadn't anticipated in advance. The CBA workers did more than contradict a metric; they were the mechanism by which the bank could discover that its metric was wrong. An organization that loses that capacity doesn't become less informed in any way it can measure. It becomes faster inside what it already understands, and quieter at the edges where understanding hasn't reached yet.
