Sundar Pichai put the number in a blog post during Google's Cloud Next 2026 conference on April 22. Seventy-five percent of all new code at Google is now AI-generated and approved by engineers, up from 50% the previous fall. It appeared as a productivity proof point, not a warning. The 75% figure arrived as good news.

Six weeks later at Cisco Live 2026, I watched Jeetu Patel open his Day 2 keynote with a customer service scenario. A refund agent, operating without oversight, had paid out $1,200 to a customer who had paid $700. The answer Cisco demonstrated was not to remove the agent but to add a review layer. Galileo, which Cisco acquired specifically for agent observability, flagged the transaction, held the payout, and routed the discrepancy to a human for approval before it completed. The $500 error never went out.

Then on June 4, Anthropic published "When AI Builds Itself." More than 80% of the code in Anthropic's own production codebase is now written by Claude. Not assisted. Written. Engineers merge roughly eight times as much code per day as in 2024. The report called for a coordinated pause among frontier artificial intelligence labs before development crosses into recursive self-improvement, the point at which an AI system can design its own successor without human intervention.

On June 4, Jack Clark, one of Anthropic's co-founders, gave his own BBC World Service interview about what the data meant. Nigel Doran had originally emailed about Anthropic's IPO filing, but Clark's interview that day shifted what we actually talked about. By the time Krupa Padhy introduced me on the weekend program on June 7, the IPO was the secondary story. My answer was what I would tell any enterprise technology leader: the message is not that AI should stop. Every output needs a human to review and approve it before it goes further.

The IPO Framing Is Worth Taking Seriously Before Dismissing It

The prevailing read on Anthropic's pause call is that it is credibility-building timed to a capital raise. Anthropic filed confidentially for its IPO one week before publishing this report.

What it cannot account for is the data Anthropic chose to disclose alongside the argument. More than 80% of merged production code written by Claude. Engineers outputting eight times the code of 2024. Leadership estimating the real figure may exceed 90% including scripts and experimental work. These figures complicate a public offering. They raise questions about auditability and what happens to valuations if AI-written code fails at scale. A company does not file for an IPO and then volunteer data that makes institutional investors nervous unless the concern is genuine.

And Pichai's April post came with no comparable motive. Google has no IPO pending, no safety positioning to establish. The 75% figure was productivity evidence. Two of the most sophisticated engineering organizations on earth disclosed the same structural shift from opposite ends of the motivation spectrum, six weeks apart. That convergence is the evidence.

The Network Demo Shows What the Solution Looks Like

The second scenario Cisco demonstrated was a network operations failure. An IT monitoring agent detected the fault, traced it to root cause, and worked out the fix. Then it stopped. It sent the engineer its full reasoning through WebEx and waited for approval before acting. The engineer reviewed it, agreed, and the network recovered in 45 seconds. After that, the system offered a standing choice: for this class of problem going forward, should the agent act without waiting? The engineer decides. Trust extends one verified decision at a time.

The observability layer that makes this practical at enterprise scale is Galileo, acquired by Cisco, whose Luna model routes all agent evaluations through Cisco's own systems at 98% lower cost than conventional monitoring, according to Cisco's figures. Splunk, now part of Cisco, provides the unified view across network, application, and AI system data that made the 45-second diagnosis possible.

Pichai's disclosure carries the same logic. "AI-generated and approved by engineers." He put the approval clause in the same sentence as the percentage. Whether enterprise leaders treat that approval step as essential infrastructure or as friction to eliminate is the operational question Cisco's demos were designed to answer.

The Jobs Question Has a Longer Answer Than Most Coverage Gives It

The leaders worth listening to are not framing this as replacement. They are asking what their teams can accomplish now that was not possible before. That is a different question, and it tends to produce different answers than the ones dominating the public debate about AI and jobs.

Some roles will evolve. Some will disappear, as there are no longer operators running elevator cars or flagmen walking ahead of automobiles. That has always been true of technology. The difference now is speed, and whether the transition is managed or damaging depends on whether a human is still in the room when the consequential decisions get made.