AI does not accelerate every part of work equally
A new workplace study of generative AI reveals a pattern that matters more to executives than an adoption chart. After enablement, intensive AI users performed 21.2% more actions in Word, Excel and PowerPoint than a matched group that received access later. Activity in Outlook and Teams increased by 7.1%.
The study analyses digital traces from eleven large international companies over twenty weeks. It counts only human-initiated actions. Among frequent AI users, work shifted relatively towards individual, documentation-focused activity. They read and organised less email and sent fewer messages to small groups.
That may be good news. Less email management and more substantive work can mean that AI is removing noise. But the researchers identify an important limitation themselves: application actions are not a direct measure of time allocation, productivity or business outcomes. Opening, pasting into, writing and editing more documents does not prove that a customer receives a faster answer or that an order moves through the process sooner.
Local acceleration can create a new queue
When AI speeds up one step, the work does not automatically disappear. It can move to the next handoff. A consultant drafts a report faster, but the subject-matter reviewer receives more drafts. A planner processes more files, but exceptions still wait for information from another department. An account team produces more proposals while pricing decisions and approvals continue at the same pace.
Local output rises. End-to-end cycle time may not. Extra documents can even increase the queue at review, decision or execution.
That is the difference between personal productivity and operational capacity. An AI tool can improve the first without changing the second. Organisations that track only licences, prompts or office-software activity will miss that distinction.
Less communication is not automatically a loss
The finding also needs nuance. Not every email transfers valuable knowledge. Summaries, suggested replies and better documents can remove repetitive coordination. An earlier field experiment with 776 professionals also found that individuals using AI could approach the performance of teams without AI on a product-development task and partly bridge functional knowledge silos.
The question is therefore not how to preserve communication volume. It is which communication the workflow genuinely needs: a decision, missing context, an exception, informed disagreement or a transfer of ownership. Those interactions must remain visible. Status email and copying work can disappear.
Measure the workflow, not the tool
A serious AI rollout should therefore begin with a bounded workflow and an owner. Before implementation, record how work enters, where it waits, which handoffs it requires and when the outcome is accepted. Then measure more than AI usage:
- end-to-end cycle time from request to completed outcome;
- waiting time between teams, systems and approvers;
- number of exceptions and escalations;
- rework after expert review;
- output that is actually used or executed;
- quality, error rate and customer impact.
These measures reveal whether AI releases capacity or merely generates more intermediate output.
Design the handoffs as part of the system
Introducing AI as a standalone assistant leaves most of the existing process untouched. The employee produces faster, but the environment in which that work must land does not change. Operational value requires a wider design: relevant data and methods, integration with the systems people use, clear ownership, an exception path and feedback from execution.
The practical next step is small: select one workflow, make its handoffs visible and decide where AI prepares information, where a person decides and where a system executes. Establish a baseline first. Only then does higher activity become meaningful.
The new study offers strong evidence that AI can shift the balance of knowledge work. The deciding business question comes next: has work merely moved on the screen, or has the outcome of the whole operation improved?