The content team drafts a campaign in a morning. Product approval still takes a week. The analyst produces the diagnosis faster, but the budget decision waits for the next meeting. The organisation has acquired a productivity gain and kept its delivery schedule.
This is one reason AI adoption needs an operating discussion alongside a tool discussion. A task can become faster while the flow of useful work remains unchanged.
The revised Shifting Work Patterns with Generative AI field experiment reports individual time savings but no detected change in the quantity or composition of tasks from individual-level provision in its setting. It is not a marketing-agency study. It does provide evidence against assuming that personal efficiency automatically reorganises collective work.
Follow one campaign all the way through
Map the work from a recognised customer problem to something live in the market. Include the time spent waiting for product evidence, legal or commercial review where applicable, localisation, production and publishing. A tool's contribution is easier to judge once the whole route is visible.
In a hypothetical campaign, AI might reduce initial drafting from six hours to two. If an unresolved product claim creates five days of waiting, the draft saving is real but does not remove the dominant delay. Worse, greater drafting capacity can create a larger review queue and make priorities harder to see.
The response should depend on what the queue represents. Some review is necessary judgement. Some delay reflects missing facts or unclear authority. Some arises because work was commissioned before the business knew what decision it wanted to make.
A useful intervention might be a verified product evidence record, a narrower brief or a clear commercial owner. None looks as dramatic as generating another hundred variants. Each can matter more to delivery.
Measure the gain where the organisation receives it
Track task time and completed work separately. A writer's saved hours, an approved campaign and a commercially useful result are related but different outcomes. If saved time allows a person to do more thoughtful analysis, that may be worthwhile even when publication volume does not rise. State the intended benefit before choosing the metric.
The Generative AI at Work study also reports heterogeneous effects across workers in customer support. That cautions against applying one average productivity assumption to every role. An experienced reviewer and a new operator may need different assistance and produce different gains.
For Frontier's integrated marketing work, the practical unit is often a completed growth intervention: the issue understood, the material produced, the change implemented and the result reviewed. Improving that unit requires cooperation across functions.
AI can expose a bottleneck by making the preceding work abundant. Treat that exposure as useful evidence. If review capacity is now scarce, change the way briefs arrive and evidence is supplied. If implementation is constrained, reduce unfinished work and resolve ownership before accelerating another upstream task.
The promise of AI in growth operations is easier to evaluate when the team asks where the saved effort goes. Sometimes it reaches the customer. Sometimes it becomes a queue. The next investment should depend on which is happening.