AI speeds up the work, not the decision
Picture a finance team that sets up an AI assistant to draft the monthly management report. What took two days now takes an hour. Everyone is pleased. Then the leadership meeting happens on the same day as always, debates the same numbers and puts off the same calls.
Nothing is wrong with the AI. The work around it didn’t change.
The bottleneck moves
Every process has a slowest step. Speed up any other step and the whole process still runs at the pace of the slow one, just with a longer queue in front of it.
Software teams are seeing this clearly. Faros AI, which sells engineering analytics, studied telemetry from over 10,000 developers in a 2025 report. On teams with high AI use, developers completed 21% more tasks and merged 98% more pull requests, but the time spent reviewing those pull requests rose 91%. Its reading: the gains stall at human approval (Faros AI).
The same thing happens in finance, operations and sales. AI drafts more outreach, more reports and more analysis. If the approval, the meeting or the sign-off stays the same, the extra output waits in a queue.
Redesign the workflow, not just the task
The companies getting results tend to change how the work flows, not just who types it. In McKinsey’s survey, out of 25 attributes tested, redesigning workflows had the biggest effect on whether organisations saw an EBIT impact from generative AI. Yet only 21% of respondents whose organisations used it said they had fundamentally redesigned at least some workflows (McKinsey).
In practice, that means asking different questions:
- Which decision does this work feed, and who makes it?
- What is the slowest step between the data and that decision?
- If the draft now takes an hour, can the decision happen sooner? What would have to change: the meeting date, the approval rule, the point at which you act?
Often the fix is not technical. Move the meeting. Let a named person approve spend below a set amount. Agree in advance which number triggers action, so the meeting confirms rather than debates.
Measure the whole cycle
Measure the time from data to decision, not the time saved on one step. How it feels is a poor guide. In a 2025 trial by METR, experienced developers using AI tools took 19% longer to finish tasks, yet believed afterwards that AI had sped them up by 20% (METR). METR’s 2026 update says newer tools probably do speed developers up, but it still treats self-reported gains as unreliable (METR).
So count days, not impressions. How many days from month-end to a decision on the numbers, before and after?
When to leave the slow step alone
Some steps should stay slow. A senior hire, a large contract or a price change for every customer deserves time and a person’s full attention. Flooding that person with more AI-drafted options can make the decision worse, not faster.
And if the decision only happens once a quarter, saving a day on the draft may not matter. Spend the effort where the queue is.
The takeaway
Before you add AI to a task, find the decision it feeds and the slowest step in front of that decision. Fix that step first, or at the same time. Otherwise you are paying for speed that ends up waiting in a queue.