I agree that automating a task, or even a collection of tasks, is not the same as automating a job. Jobs often sit on top of tasks and require human affordances such as context, responsibility, judgment, liability, and coordination.
I also agree that bottlenecks are a pragmatic route for quick wins. But I would be careful not to limit the AI opportunity space only to bottlenecks that are already visible and measurable. As the Solow productivity paradox put it, “You can see the computer age everywhere but in the productivity statistics.” Not everything that counts can be counted, and not everything that can be counted counts.
The risk is that firms focus only on measurable workflow bottlenecks while leaving a deeper “organization design bottleneck” outside the frame: could the organization itself be redesigned differently around AI?
In complex adaptive systems, the better approach may be to preserve ambiguity rather than prematurely collapse it into measurable bottlenecks. Firms need to ride both routines at once: exploitation, by using AI to improve visible constraints in existing workflows; and exploration, by asking whether AI makes entirely different organizational designs possible.
If incumbents restrict exploration to existing processes and measurable constraints, they may fall prey to the myopia of learning described by Levinthal: improving locally within the current architecture while missing the possibility of a different one. In that case, new organizations may have a structural advantage, because they can explore AI-enabled designs without the same legacy constraints.
Fascinating read...the key point is that AI should be used to reimagine workflows, rather than be deployed just for incremental improvements. Such an approach will help identify the bottlenecks(am reminded of the Theory of Constraints that Goldratt explained so well in his book, 'The Goal' and how improving system flow requires identifying and managing these constraints rather than optimizing every individual machine). If we take this as a basis, then successful implementation of AI should result in a 'improved system flow=workflow', rather than apply AI only at a task level...
Although I wonder if a good number of leaders do not know that change and outcomes will be a multiplicative version of all those factors and hence must go from one bottleneck to another? One at a time!!
My guess is, probably some of them if not many, do know that. But. It's the vested interests and incentives alignment that does not allow, for the honest admission and then taking a longer, slower and more grueling path to transformation. And then the long term view isn't safeguarded either with the speed of the change.
Yeah. Organizations tend to be very inert and slow to change. At INSEAD it took me so much effort to remove protections so I could take advantage of Claude Code
I agree that automating a task, or even a collection of tasks, is not the same as automating a job. Jobs often sit on top of tasks and require human affordances such as context, responsibility, judgment, liability, and coordination.
I also agree that bottlenecks are a pragmatic route for quick wins. But I would be careful not to limit the AI opportunity space only to bottlenecks that are already visible and measurable. As the Solow productivity paradox put it, “You can see the computer age everywhere but in the productivity statistics.” Not everything that counts can be counted, and not everything that can be counted counts.
The risk is that firms focus only on measurable workflow bottlenecks while leaving a deeper “organization design bottleneck” outside the frame: could the organization itself be redesigned differently around AI?
In complex adaptive systems, the better approach may be to preserve ambiguity rather than prematurely collapse it into measurable bottlenecks. Firms need to ride both routines at once: exploitation, by using AI to improve visible constraints in existing workflows; and exploration, by asking whether AI makes entirely different organizational designs possible.
If incumbents restrict exploration to existing processes and measurable constraints, they may fall prey to the myopia of learning described by Levinthal: improving locally within the current architecture while missing the possibility of a different one. In that case, new organizations may have a structural advantage, because they can explore AI-enabled designs without the same legacy constraints.
Fascinating read...the key point is that AI should be used to reimagine workflows, rather than be deployed just for incremental improvements. Such an approach will help identify the bottlenecks(am reminded of the Theory of Constraints that Goldratt explained so well in his book, 'The Goal' and how improving system flow requires identifying and managing these constraints rather than optimizing every individual machine). If we take this as a basis, then successful implementation of AI should result in a 'improved system flow=workflow', rather than apply AI only at a task level...
Good stuff as always.
Although I wonder if a good number of leaders do not know that change and outcomes will be a multiplicative version of all those factors and hence must go from one bottleneck to another? One at a time!!
My guess is, probably some of them if not many, do know that. But. It's the vested interests and incentives alignment that does not allow, for the honest admission and then taking a longer, slower and more grueling path to transformation. And then the long term view isn't safeguarded either with the speed of the change.
So there is at times, just no where to hide ;)
Yeah. Organizations tend to be very inert and slow to change. At INSEAD it took me so much effort to remove protections so I could take advantage of Claude Code
Whaouh. Thanks Pushan, that reading was amazing.
Long time I haven't read something that clear on AI usage.
Really appreciated the flow, the examples, the precision.
A must read for all people involved or at least interested in AI deployment.
Great thansk
Thanks Florent!
Much enjoyed reading this article.
Fascinating piece, Pushan.
Thanks. Will apply some of these ideas to the AI race between USA and China and between Google and OpenAI