Two years after ChatGPT launched, Danish workers who used AI chatbots reported productivity benefits. Their employers had reorganized jobs around the chatbots and added new tasks, such as checking AI output and fitting the tools into existing processes.
Companies have found the payoff harder to see. In McKinsey’s 2026 survey of 1,719 respondents, 80 percent said AI had improved their individual productivity. Only 37 percent said AI had contributed to their organization’s EBIT, about the same share as a year earlier. McKinsey calls the roughly 6 percent who attribute at least 5 percent of EBIT to AI high performers. Nearly three-quarters of them had fundamentally redesigned workflows because of AI, compared with a quarter of other respondents.
The gain people feel at their desks seems to be taken up by new tasks and existing processes before it reaches the bottom line. Economists have seen gains trail adoption like this with earlier general-purpose technologies. Brynjolfsson et al. called this pattern the productivity J-curve in 2021. They argued that a technology like AI tends to pay off only after an organization invests in redesigned processes and new skills. For anyone planning a rollout, most of the change lies in the work around the tool.
Where the extra work piles up
Eliyahu Goldratt’s The Goal explains one of the main reasons gains from new tools often don’t show up in the bottom line. A system produces only as much as its slowest step allows, and that step is known as the bottleneck. A faster step elsewhere mostly adds to the work waiting at the bottleneck, which at best caps the benefit of the new tool and more likely slows the whole flow down. Little’s Law describes the slowdown mathematically: with the bottleneck’s pace unchanged, more work waiting means each piece takes longer to get through.
Three things to watch after the pilot
A pilot seldom surfaces the side effects of an AI rollout, because a small trial team can absorb the extra output without anyone else noticing. Once the tool is deployed at scale, though, the side effects tend to appear in predictable places and can cause serious problems there. The tips below can help you find those places.
-
Audit queues and handoffs. To see where extra output will pile up, follow a single piece of work from request to result and note every place it sits waiting. Any of those places can become the next bottleneck once the current one is relieved.
-
Carefully watch which tools people use. In May 2026, Ontario’s Auditor General reported that the government’s secure, approved AI tool accounted for only about 6 percent of public servants’ use of popular generative AI websites. The other 94 percent went to public sites that lack its data protection. The government had neither blocked those sites nor tried to raise use of the approved tool. Training on responsible use was optional, and only 3 percent of staff had completed it. In my experience training is often neglected, but on its own it is unlikely to change which tools people reach for. Finding out why staff preferred the other sites, and enforcing the approved tool where needed, will probably do more.
-
Mind the metrics. Performance measures have a long history of changing how people behave in unintended ways. A team measured on tickets closed or documents produced may use AI to produce more of both, regardless of how that extra volume affects the steps downstream. Chasing the measured number can also squeeze out work that isn’t measured, and quality problems can follow quickly. Measures set before a rollout can keep steering behaviour long after the tool arrives, so revisiting them has to be part of the long-term strategy.
Before the tool arrives
Before choosing where to deploy AI, look at the whole system, because the most convenient place to start may simply feed a bottleneck and slow the whole organization down. It can help to pair whoever implements the technology with someone experienced in systems thinking and organizational flow.
Four questions asked early can head off unwanted side effects:
- Where will the extra output wait, and who will receive it?
- Who decides how to use the time the tool frees up, and with what information?
- Who answers for the data that staff put into AI tools, approved or not, and how soon would that person hear about a problem?
- What will your current measures reward once the work gets faster?
If your team suddenly produced twice as much as today, where would that extra work be sitting tomorrow?
Much of the potential of people and AI goes unrealized because of how the work around them is organized.
At Klose Coaching, we help you identify where your delivery system gets in the way and how to redesign it with AI built in. Book a free 30-minute consultation.
Sources
- Anders Humlum and Emilie Vestergaard, “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI,” NBER Working Paper 33777 (2025).
- McKinsey & Company, The State of AI in 2026: On the Road to ROI (2026).
- Erik Brynjolfsson, Daniel Rock and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics 13, no. 1 (2021): 333–372.
- Eliyahu M. Goldratt and Jeff Cox, The Goal: A Process of Ongoing Improvement (North River Press, 1984).
- John D. C. Little, “A Proof for the Queuing Formula: L = λW,” Operations Research 9, no. 3 (1961): 383–387.
- Office of the Auditor General of Ontario, Use of Artificial Intelligence in the Ontario Government, Special Report (May 2026).