The Specific Problem on the Table

A recent analysis circulating in AI industry circles has identified a pattern called "tokenmaxxing," where employees, encouraged by corporate mandates to demonstrate AI usage, deliberately generate verbose, inefficient prompts that consume large volumes of tokens without producing proportionally better outputs (Companies With Goals Of AI Tokenmaxxing, 2025). The report frames this primarily as a cost problem. That framing is accurate but incomplete. What is actually being described is a structural breakdown in how organizations are attempting to build AI competence, and the mechanism behind that breakdown is worth examining carefully.

Procedural Mandates in Algorithmically-Mediated Environments

The tokenmaxxing pattern follows a predictable logic. When firms set quantitative targets for AI use, such as prompt counts, token volumes, or logged interactions, they are substituting procedural compliance for genuine capability development. This mirrors what Hatano and Inagaki (1986) called routine expertise: workers learn to execute a procedure to specification without developing the underlying principles that would allow them to adapt that procedure to novel situations. The employees engaging in tokenmaxxing are not failing to comply with their employer's directive. They are complying precisely, and the results expose how poorly constructed that directive was in the first place.

This connects directly to a theoretical problem I have been working through in my dissertation research. The Algorithmic Literacy Coordination framework proposes that competencies in algorithmically-mediated environments cannot be transferred through procedural instruction alone. Kellogg, Valentine, and Christin (2020) documented that workers subject to algorithmic oversight develop awareness of how systems function without converting that awareness into improved performance. Tokenmaxxing is the organizational equivalent: employees become highly aware that AI usage is being measured, but the measurement itself produces no schema about how to use AI effectively.

What the Usage Metric Actually Measures

There is a distinction in cognitive science between knowing the topology of a system and knowing its topography. Topology describes the structural shape of constraints; topography describes how to navigate within them. Corporate AI mandates built around token or usage metrics give employees no topological understanding of what distinguishes productive AI interaction from unproductive interaction. The metric is topographically neutral. It rewards volume regardless of structure, which is why workers who want to satisfy the metric without generating useful work can do so trivially.

Sundar (2020) observed that when users interact with systems they do not structurally understand, they develop folk theories, that is, individual impressions about how the system works that may have no relationship to its actual architecture. Tokenmaxxing is essentially the organizational institutionalization of a folk theory: the implicit belief that AI value is proportional to AI use, measured by volume. That belief has no basis in how large language models actually generate value, but it is the belief that usage-based mandates implicitly transmit.

The Costs Are Not Just Financial

The industry analysis focuses on the monetary cost of wasted token consumption, which is a real concern given that enterprise AI infrastructure carries significant per-token pricing. But the more durable cost is epistemic. Organizations that measure AI adoption through volume metrics are actively producing a workforce that has developed procedural habits around AI use without any structural understanding of what makes AI interaction effective. When those platforms change, when token limits shift, when model behavior updates, or when the task domain expands beyond current workflows, that procedural knowledge will not transfer.

Gentner's (1983) structure-mapping theory predicts exactly this failure. Transfer depends on mapping structural relations between domains, not surface features. A workforce trained to generate high token volume has learned a surface feature of AI interaction. That learning will not transfer to a context where the relevant structural question is how to construct a prompt that elicits precise, constrained output, which is the actual skill that produces organizational value from AI systems.

The Organizational Design Implication

The tokenmaxxing problem is not primarily a problem of employee motivation or honesty. It is a problem of schema deficit at the organizational level. Firms deploying AI adoption mandates without investing in what I would call schema induction, that is, training that targets the structural features of how AI systems process and respond to input, are producing compliance without capability. The variance in outcomes between employees with identical AI access, which is the core puzzle motivating my dissertation research, will almost certainly widen under these conditions. Workers who happen to develop accurate structural schemas through their own experimentation will outperform colleagues who followed the official metric, not because of natural ability, but because organizational incentives actively suppressed schema development in the latter group. That is a design failure, not a talent failure, and the distinction matters for how firms should respond.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). W. H. Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.