Jakob Nielsen: AI value comes from redesigning workflows end-to-end, not from automating tasks

What the article is about

Jakob Nielsen, Ph.D., synthesizes findings from a large-scale field experiment run by researchers at INSEAD and Harvard involving 515 startups. The study compared two groups: one that received standard AI tool training and entrepreneurship support, and one that was also taught principles for redesigning workflows around AI from the ground up. The revenue difference between the groups was 90% in favor of the workflow redesign group, measured over a sustained period.

Context

The article opens by distinguishing between two types of productivity gain. Task productivity is a local efficiency gain — a single step in a process gets faster. Workflow throughput is a systems-level gain — the entire sequence from input to output produces more because the structure of the work changed, not just the speed of one step. Nielsen argues that most teams are capturing task productivity and leaving workflow throughput on the table.

The research found that optimizing individual steps shifts bottlenecks downstream rather than eliminating them. A faster first step in a five-step process does not increase overall output if step four is the constraint. Only restructuring the whole sequence captures the full potential.

Key method and takeaways

Nielsen identifies four moves that appear across effective AI workflow redesigns: removing handoffs by combining steps that were previously separated across roles or tools; parallelizing work by running variants simultaneously rather than sequentially; moving humans to exceptions by letting AI handle the standard case and reserving human attention for cases that fall outside defined parameters; and adding evaluation loops to continuously measure output quality rather than assuming it.

Treated firms in the study also required 40% less external capital while maintaining flat headcount — meaning the gains came from structural changes in how work flows through a team, not from hiring more people or spending more on tools.

The article specifically flags “human glue” as a target for elimination: manual steps that exist solely to bridge disconnected systems, such as copying data from one tool to another, reformatting documents for different audiences, or coordinating hand-offs between departments. These steps add no value and are often invisible in process maps.

Who it is useful for

Product managers who have experimented with AI tools but have not seen the productivity gains they expected. Nielsen’s framework gives a diagnostic lens: if adopting an AI tool did not change the structure of the workflow — who does what, in what order, with what reviews — then the efficiency gain is local and the throughput of the team remains the same. The article is also useful for anyone leading an AI adoption initiative at a team or organizational level, where individual tool rollout is easier to track than end-to-end process impact.