A large-scale experiment in Alibaba's Taobao customer-service operations found that agentic AI handling eligible chats under human supervision cut average chat duration by 16.8 percent. Customer satisfaction fell by 0.412 points on a five-point scale. Faster service, worse outcomes. But the most revealing number was in the escalation data.
When the system flagged emotional escalations late in a conversation, after customer sentiment had already deteriorated, satisfaction dropped 0.928 points. Workers who inherited those late escalations wrote fewer messages, asked fewer diagnostic questions, offered fewer solutions. The researchers' interpretation: by the time the system surfaced the problem, the damage was entrenched. The human recognized that recovery was unlikely and adjusted accordingly.
Not through negligence. Through the rational assessment of someone who could see the situation had been decided before they arrived.
I keep returning to this finding because it exposes a structural problem that extends well beyond customer service. The escalation system existed. The human was in the loop. But the loop activated too late for the human's intervention to matter, and the human, being a reasonable person in an unreasonable position, behaved accordingly. The system's speed didn't remove human judgment. It rendered human judgment ceremonial.
You see the same structural logic in software engineering, though the surface looks entirely different. A Microsoft study of coding agents found that engineers using command-line AI tools merged 24 percent more pull requests per day. The authors were careful to note that merged PRs are a throughput proxy, not a measure of delivered value, and that the field lacks agreed-upon measures for software-quality effects. What the paper did not measure, and could not easily measure, was what happened on the receiving end of those PRs: review time, reviewer workload, review quality. Twenty-four percent more artifacts entered the pipeline. Whether anyone was positioned to meaningfully evaluate them is a question the data doesn't answer. But the organizational physics are not subtle. Reviewers are finite. Artifacts are not.
And evaluation requires more than time. It requires context. Current observability standards for agent systems capture what happened: model calls, tool invocations, error codes. What they generally do not capture is why the action was authorized, whether it fell within the scope of what was delegated, or what the reviewer should actually be checking for. The same trace entry, "agent accessed the database," means something entirely different depending on whether the user requested that access or the agent wandered into unrelated data. A reviewer looking at the log sees identical lines. The distinction between appropriate action and scope violation lives outside the trace, in delegation context that most systems never record.
I've argued earlier this year that production agents mature by becoming institutional clerks: they gather, organize, and present artifacts for human review. That framing assumed the review layer would hold. What the Taobao and Microsoft findings suggest is that the review layer is precisely the part most vulnerable to agent-driven acceleration. A reviewer who must evaluate 24 percent more pull requests does not receive 24 percent more judgment. A service worker who inherits a conversation after sentiment has collapsed does not get a fresh start. Review is a capacity. It degrades under load like any other capacity. And it degrades fastest when the artifacts look complete, the logs look clean, and the context needed to challenge them was never captured in the first place.
Can your institution challenge decisions at the speed it now produces them?
Organizations have spent decades building the institutional rhythms of challenge, review, and correction: code review pipelines, escalation paths, audit trails, approval chains. These rhythms evolved at human tempo because humans were the ones producing the work. Now the production side accelerates and the contestability side stays fixed. The result is outcomes that no one had the time, the context, or the standing to question before they became permanent. In institutional design, there is a name for this condition. It is the appearance of efficiency with the substance quietly removed.