Ethan Mollick described the pattern precisely: “When workers gain productivity from AI, the rest of the bottlenecks in the system can eat up any gain.” This is not a prediction about what might happen in organizations that deploy AI badly. It is an observation about what is already happening in organizations that have deployed AI well — technically well. The automation works. The organization around it does not change at the same rate. The efficiency gain evaporates.
ishikawa
Why AI automation gains disappear
Org Design
Approval pace unchanged
Reporting unchanged
Stakeholders
Misaligned goals
Buy-in not secured
Process
Workflow not redesigned
Handoffs unchanged
Data Trust
Outputs re-verified manually
Unfamiliar format flagged
Scope
Adjacent steps ignored
Automation isolated
The Automation Works. The Approval Process Does Not.
Consider what actually happens in a well-executed AI automation deployment. The automated portion of the workflow runs at ten times the previous speed, with higher accuracy, at lower cost. A document processing workflow that took two days now takes two hours. A data extraction process that required three analysts now runs overnight without them. The technology delivered exactly what was promised.
Then nothing else changes.
The reports still go through the same sign-off chain they always did. The downstream team that receives the outputs of the automated process still applies the same manual review they always did — because no one updated their workflow, the output format changed slightly, or they simply do not trust a system they did not build. The approval committee that met weekly to review what the old process produced is still meeting weekly. The automation outpaced the organization and then waited for it to catch up.
I saw this pattern directly during an engagement with a class-action settlement administration company. We built a well-designed, customer-facing case management platform — the product was solid, the architecture was clean, the development executed well. The problem was that the business had not fully aligned on one critical dimension before the build began: how legal counsel representing both sides of each case would access the platform. The settlement administrator sat in the middle as a neutral party, and the access question had legal and contractual implications that had not been resolved before work started.
The platform stalled — not because the technology failed, but because stakeholder buy-in was not fully secured before the build began. The owner of the project should have paused until every key stakeholder was aligned and the legal implications were resolved. In retrospect, that was the prerequisite the project needed before technical work started. The organizational problem was not created by the project — it was inherited by it, unaddressed.
That pattern repeats in AI automation deployments constantly. The technology works. The organizational alignment that should have preceded it did not happen.
Three Bottlenecks That Most Often Absorb the Gain
Approvals designed for the old throughput rate. Most organizational approval structures were designed around human processing speed. When AI accelerates the upstream work, approvals — which exist to catch errors at human-review rates — become the rate-limiter. An invoice that used to take three days to process now takes three hours to process and two and a half days to approve. The approval process was not wrong in the original system. It was calibrated for a different volume. Recalibrating it requires someone with authority over both the automation and the approval structure — and that person is rarely the same person who built the automation.
Adjacent teams who did not agree to the workflow change. AI automation rarely affects only the team that deploys it. It changes what adjacent teams receive, how fast they receive it, and what format it arrives in. If those teams were not part of the design conversation, they are not prepared for what lands in their queue. Their response is often to re-verify AI outputs manually — which absorbs the time savings — or to escalate every edge case as an exception — which creates new bottlenecks. Alignment across stakeholder groups is a prerequisite for the efficiency gain to survive the organizational boundary.
Outputs that are not trusted. AI automation produces outputs that look different from what humans produce — in format, in confidence expression, in how exceptions are surfaced. If the people downstream of the automation do not trust those outputs, they will re-check them. That response is rational, not obstructive. The fix is not faster automation — it is transparency about how the automation works, what it gets right, and what it flags for human review. Trust is designed in or it does not appear.
What the Fix Requires
The problem is not the automation. The problem is deploying automation without redesigning the organizational processes surrounding it.
Before deploying AI automation, the team responsible for it needs to be able to answer: where does this automation hand off to a human process, and has that human process been updated for the new throughput? Who receives outputs of this automation, and do they trust them enough to act without re-verification? What approval structures will constrain the downstream impact, and have those structures been adjusted?
These are not engineering questions. They require someone with standing to get honest answers across organizational boundaries and authority to drive the workflow changes that follow. The automation team usually has neither. That is why the efficiency gain survives the deployment and disappears in the quarterly review.
The automation worked. Someone needed to manage what happened next.
