Intelligence is becoming abundant. Applied intelligence remains scarce.
A working thesis on where durable value moves as capable AI becomes cheaper, more accessible and easier to embed in ordinary workflows.
WORKING RECORD
- Research question
- Where does durable value move when increasingly capable intelligence becomes cheap and widely available?
- Thesis
- As access to capable AI becomes less scarce, durable advantage moves toward problem selection, proprietary context, workflow integration, accountable judgement, distribution and sustained execution.
- Evidence
- sourceThe cost of querying a model at roughly GPT-3.5-level performance fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a reduction of more than 280 times at that capability threshold.
- sourceIn a field study of 5,179 customer-support agents, access to a generative AI assistant increased issues resolved per hour by 14% on average, with substantially larger gains for novice and lower-skilled workers.
- sourceA study of 758 consultants found that AI improved speed and quality on tasks inside its capability frontier, but made participants 19% less likely to reach the correct answer on a task outside that frontier.
- operator observationIn operating work, useful outcomes still depend on choosing an appropriate task, supplying domain context, integrating the result into a workflow and assigning accountability for the decision.
- Counterarguments
- The owners of the strongest intelligence layers may continue to capture most of the value.
- Falling capability costs may compress the value of application businesses as quickly as they create them.
- Application advantages may be temporary when competitors can reproduce the same workflow, context or distribution.
- The capability frontier may advance quickly enough that judgement and integration advantages narrow rather than compound.
A change in the scarce input
The relevant shift is not that intelligence has become free or uniform. It is that useful levels of machine capability are becoming cheaper and easier to access across a growing range of tasks. That weakens access to a model as a standalone source of advantage.
The underlying intelligence layer can remain technically concentrated while application becomes broadly available. Both conditions can be true: model providers may capture substantial value, while organisations still need to decide where and how capability should enter real work.
Application is a system, not a prompt
Field evidence shows that the same technology can improve performance in one task and reduce it in another. The difference sits in the shape of the task, the context available to the system, the surrounding workflow and the person's ability to recognise when the output is outside the capability frontier.
Applied intelligence therefore includes problem selection, data and context, interfaces, workflow design, evaluation, human authority and feedback. A model output becomes economically useful only when those elements carry it into an accountable decision or action.
Where advantage may move
If baseline capability continues to diffuse, defensibility is more likely to come from access to a valuable problem, accumulated operating context, trusted distribution and the ability to improve a system through use. These assets are slower to copy than a prompt and more local than a general model.
This does not imply that every application becomes durable. Thin products remain vulnerable to model improvement and imitation. The thesis favours systems that become better through workflow integration, evidence and repeated execution.
What would change the thesis
The thesis weakens if model-layer economics continue to dominate while application margins collapse, if general systems absorb domain context without costly integration, or if workflow learning proves easy for competitors to reproduce.
The next research task is empirical: compare applications that produce durable operating gains with those that merely expose a temporary capability. The useful unit of analysis is not model performance alone, but a complete intervention in a real workflow.
UPDATES
Public working record opened with an initial evidence base and explicit counterarguments.