AI Productivity Bottlenecks: Why Faster Work Does Not Always Mean Faster Delivery

AI productivity bottlenecks can slow delivery even when drafting gets faster. Explore research on review queues and practical changes to ClickUp workflows.

Dtech SystemsClickUp Diamond Partner
4 min readClickUp

AI productivity bottlenecks can leave a business waiting for the final result even when individual tasks become faster. A team may produce more code, draft more proposals, or prepare more content, yet deliver little additional work because review and acceptance cannot keep pace.

For business leaders, the practical question is where work goes after AI produces its first output. That is the point at which a productivity claim needs to become an operating discussion.

What the research examined

In their working paper, Artificial Intelligence in the Firm: Bottlenecks in Software Production, Fiona Chen and James Stratton study software engineering using data from an engineering analytics platform. The version reviewed here, dated 9 November 2026, covers 445 million work events involving 720,542 workers at 655 firms.

The authors use differences in the timing of firms' AI adoption to estimate effects. This is an observational research design whose interpretation depends on assumptions about the comparability of earlier- and later-adopting firms. It is not a randomised trial of ClickUp or a study of Dtech clients.

The paper distinguishes intermediate coding activity from final software output. That distinction matters: producing more material is different from getting reviewed, accepted work into use.

More activity did not produce proportional delivery gains

For coding assistants, the authors estimate a 30.4% increase in lines of code 12 months after adoption. Estimated effects on resolved issues and epics were much smaller. For coding agents, the paper estimates larger increases in coding activity and smaller increases in final output. The agents' effect on resolved issues is described as borderline statistically significant.

The review stage helps explain the gap. After agent adoption, the authors estimate that submission-to-approval time increased by 56.1%, alongside longer waiting and active review. These are estimates from the studied setting, not universal performance forecasts.

The paper also reports no detectable employment effect in its empirical analysis, with uncertainty around the estimates. Its longer-term employment and output scenarios are model simulations rather than observed outcomes.

A practical implication for ClickUp workflows

The study did not test ClickUp. Dtech's practical application of the bottleneck idea is to inspect the whole delivery path before expanding AI-assisted production.

Consider a hypothetical content team that can prepare ten drafts each week but can review only four. Increasing drafting capacity to twenty does not, by itself, increase publication capacity. It can instead create a larger queue of unfinished work.

Map the process through drafting, readiness for review, active review, revisions, and acceptance. Record the accountable reviewer and the information they need to decide. Keep the distinction between waiting and active work visible, even if the first version uses a simple manual measurement.

ClickUp can organise the associated tasks and owners. A useful configuration follows the team's actual approval process; it does not assume that moving a task to a particular status enforces an approval control automatically.

Measure the complete outcome

During an AI pilot, track four things: accepted output, elapsed delivery time, review effort, and rework. Compare similar work over a defined period and record changes in staffing or complexity that could affect the result.

If drafting gets faster while acceptance slows, investigate the queue. The reviewer may lack context, the output may need more correction, or too much work may be entering the stage at once. Each cause calls for a different intervention.

Try a bounded change, such as a clearer brief, smaller review batches, or protected reviewer time. Check the result before increasing upstream output again. Review quality remains part of the outcome; a shorter queue is not useful if acceptance becomes unreliable.

Choose the next improvement around the constraint

The research offers a reason to be precise about productivity. Businesses need useful work completed and accepted, alongside the ability to produce it efficiently. A faster task can support that goal when the surrounding process is ready.

Before adding another AI tool, identify where the work waits today, who can move it forward, and which change would help that stage. Then evaluate the result through the complete workflow.

Source: Chen and Stratton, working-paper version dated 9 November 2026, abstract and sections 1, 4–6, especially introduction pp. 3–6. The authors' public version is linked below and may be updated independently.

Sources and further reading

Related Dtech guides

Build a workflow your team can operate

Dtech helps teams map processes, configure ClickUp and support adoption. Explore our ClickUp implementation services or book a call to discuss your workflow.

Chat on WhatsApp