I’ve recently been looking at how AI is influencing productivity and performance in Global Business Services (GBS) organizations. Productivity is about efficiency: output per unit of input, such as invoices processed per person. Performance is broader: whether we’re delivering the outcomes that matter, including quality, accuracy, customer satisfaction, meeting SLAs. For years these have sat in quiet tension. Push volume too hard and quality slips. Focus only on quality and throughput slows. What’s caught my attention is whether AI actually dissolves that trade-off or just moves it somewhere we’re not looking.

Picture any process where a task that took three hours now takes one. The two hours saved are real. But the work still moves to the next step on the same schedule, and the whole thing still finishes when it always did. The speed happened, but the saved time just went nowhere.

This is the puzzle of AI at work right now. Individuals are getting faster, and the organizations they work for are not. In Microsoft’s 2026 Work Trend Index, 58% of AI users say they produce work they could not have produced a year ago. Yet the returns to the organization are hard to find. A 2025 study reported that roughly 95% of enterprise generative-AI pilots showed no measurable return.

The reason is that work does not happen in a step. It happens in a chain. A draft is prepared, then reviewed, then approved, then it waits for the closure. Each link hands off to the next. The output of the chain is set by the slowest link and by how the links are joined, not by how fast any one worker can go. AI makes a worker faster. It does not, by itself, touch the chain.

So the saved time pools as slack between the links. It doesn’t become more cases cleared. It becomes waiting. The processor’s work now lands on the reviewer’s desk hours earlier, but the reviewer still works to the same schedule. Speed at one step is invisible to a chain that was timed around the old speed.

The bottleneck moved. It used to be how fast people could do the work. Now it is how the work is arranged.

The fix is to redesign the work. Remove steps. Collapse approvals that exist only out of habit. Re-time the batch so it runs when the work is ready, not on a calendar set for a slower era. This is the hard part, and it is the part being skipped. Buying tools is fast, cheap, and uncontroversial. Redesigning a process is slow, and it is political, because every step belongs to someone, and collapsing a step means telling that someone their check is no longer needed. Organizations are doing the easy half and calling it a strategy.

So the instruction for managers is narrow and unglamorous. Stop measuring adoption. Measure the chain. Find the step where AI-saved time is pooling as slack, then go redesign the links around it, even though that means the meetings and the arguments that buying a tool let you avoid. The model is not the constraint anymore. Your process is. Until you change the process, the time your people save will keep going nowhere.