Is Attention the New Scarce Resource?
Cheap generation, finite judgement. Or is the bottleneck somewhere else entirely?
The Question
The AI Abundance Paradox argues that when generating work becomes cheap, review, decision and change capacity set the ceiling on AI value. Is human attention really the scarce resource that matters, or is that a knowledge-work habit of mind?
The optimist's case
Output is cheap. Judgement is not. Attention is where the queue forms.
Herbert Simon made the point in 1971: a wealth of information creates a poverty of attention. Generative AI is an information-producing technology. Every draft, summary and option it creates must be read, checked or chosen between by someone.
Interpretation The evidence that review is a real constraint is already visible. In a controlled study, experienced developers expected AI to speed them up and took 19% longer when allowed to use it. In the 2024 DORA survey, higher AI adoption was associated with small estimated falls in delivery throughput and stability. Neither proves the thesis, and both are consistent with it.
If attention is scarce, the economics change. Token cost becomes a minor line next to the hours of human judgement the output consumes, and the sensible metric is value per unit of reviewer time.
The sceptic's case
Attention is a symptom. The binding constraint is usually data, integration or permission.
Calling attention the constraint is easy and hard to test. In many organisations the output of AI is blocked long before a human reads it: data is not accessible, integrations do not exist, security will not approve the use, or nobody has the authority to change the process the output would feed.
Interpretation The Tokenomics Foundation survey points the same way. The largest challenge respondents named was proving value, at 43%, not review capacity. Ownership and attribution of spend, not attention, separated the organisations that could show value from those that could not.
There is also a technical answer. Automated checks, evaluations and policy gates can absorb a large share of verification. If those work, attention is not scarce, it is simply being used badly.
Synthesis · house view
It is true where work ends in human judgement. Find the real queue before budgeting for it.
Both are right about different workflows. Where AI output ends in a person who must read, approve or act on it, human attention sets the ceiling. Where it is blocked earlier, or where automated assurance is good enough, something else does.
Interpretation The test is empirical and local. Measure queue age, review minutes per item and the load on reviewers for one workflow. If the queue is human, attention is the constraint. If it is not, look upstream.
If We Get It Right / If We Get It Wrong
Right + act: Review and decision capacity is budgeted like output. Low-value work is stopped before it consumes it.
Right + ignore: Generation grows, queues lengthen and quality slips silently as people review faster.
Wrong + act: Organisations protect reviewer hours that were never the problem and delay fixing data and integration.
Wrong + ignore: The workflow works, and effort on attention budgets was unnecessary.
The Author's Honest Position
Interpretation In knowledge work, attention is usually the first constraint that bites, and it is the one least often priced. It is not the only one, and a good deal of the argument rests on untested heuristics such as the AI Pressure Index.
What would change our mind: workflows with heavy AI output and short, stable review queues over a long period.