Enterprise AI spend has crossed the threshold at which token consumption has become a primary financial signal, yet token volume measures model utilization, not business outcome. The practitioner term "tokenmaxxing" has surfaced to describe a pathology that has no formal definition: the organizational tendency to maximize inference throughput as a proxy for AI productivity, without establishing a causal link between consumption and financial return. This paper introduces the AI Value Accounting Layer (AVAL) and three original constructs that provide the formal architecture to address this gap:
The term "tokenmaxxing" entered enterprise AI vocabulary in mid-2026 to describe organizations that treat inference token throughput as the primary signal of AI program health [1]. The term is pejorative: it captures the organizational pathology of maximizing a proxy metric rather than the underlying outcome. Yet the criticism remains largely rhetorical. The industry's response has been to recommend "measuring value, not tokens," advice that is correct in direction and entirely absent in mechanism. No framework specifies what value signals to instrument, how to bind them to token expenditure at authorization time, or how to surface the divergence between consumption and return as a named, auditable quantity.
The financial scale of this gap is material. Global enterprise AI spend reached an estimated $2.59 trillion in 2026 [2], with per-organization AI expenditure rising substantially year-on-year [3]. Yet fewer than one in three executives can name a specific financial outcome attributable to their organization's AI program, according to directional survey data [4]. Reported incidents of organizations exhausting annual AI budgets before mid-year, with no corresponding business outcome inventory to justify continuation, have accelerated [5]. The problem is not that organizations are unwilling to measure value; it is that no instrumentation architecture has been formally specified to make value measurement structurally possible at the point of spend.
Existing frameworks address adjacent concerns. NIST AI RMF [6] focuses on risk identification and management, not financial return attribution. ISO/IEC 42001 [7] establishes AI management system requirements without specifying financial instrumentation. FinOps practice [8] addresses cloud cost optimization but treats AI inference as a cost-to-minimize rather than a value-to-attribute. AI observability platforms instrument latency, error rates, and throughput: not business outcome linkage. The gap is structural, not a matter of incomplete vendor coverage.
This paper's contributions are:
The remainder of this paper is organized as follows. Section II surveys related work and the gap. Section III defines TVD and Consumption Decoupling formally. Section IV specifies the AVAL architecture. Section V presents the Maturity Model. Section VI discusses governance alignment and limitations. Section VII concludes.
Token consumption as a management signal emerged from a structural artifact: AI API pricing is denominated in tokens, making token volume the most readily available financial metric in AI observability dashboards. As organizations scaled inference workloads, token dashboards became the de facto reporting surface for AI program health, despite the absence of any causal model linking tokens to business outcomes [1].
The practitioner literature identified this pathology in 2026 under the term "tokenmaxxing" [9]. The diagnostic observation is that high-consumption AI programs are not systematically higher-value than low-consumption ones: a legal summarization agent that processes ten thousand documents monthly and reduces review time by a verifiable margin may consume far fewer tokens than an internal chatbot generating long-form responses to questions already answered by search. Token volume discriminates consumption level, not value level.
FinOps, as defined by the FinOps Foundation [8], provides a cloud financial management practice that addresses cost visibility, cost allocation, and continuous optimization. Its unit economics framework, cost per request and cost per user, represents the closest existing practice to token-level financial accountability. However, FinOps practice treats AI inference as a cost center to be minimized, not a value-generating investment to be attributed. The FinOps model has no native construct for binding an authorized expenditure event to a prospective value signal at authorization time.
AI observability platforms, including purpose-built LLMOps tooling, instrument token consumption, latency distributions, error rates, and prompt/completion pairs [10]. These platforms provide the data substrate from which value attribution could be computed, but they do not define or surface value attribution as a native construct. The gap is definitional, not infrastructural.
No existing framework defines a formal architecture that:
This is the gap the AI Value Accounting Layer (AVAL) is designed to fill.
Let C(t) denote an organization's cumulative verified AI token expenditure (in normalized cost units) over interval [0, t], for t > 0. Let V(t) denote the organization's cumulative verified value delivery (in the same normalized units, derived from named value signals mapped to financial outcomes) over the same interval [0, t]. The Token-Value Displacement TVD(t) is defined as:
TVD(t) = C(t) − V(t)
For non-empty deployment set D and time window [t₀, t₁] where t₁ > t₀, an organization exhibits TVD Growth when ΔTVD = TVD(t₁) − TVD(t₀) > 0. An organization is in TVD Deficit when TVD(t) < 0 (value delivery exceeds expenditure in the defined window, indicating underinvestment). TVD is undefined when no value signals have been registered for a deployment (see Definition: Value Opacity).
TVD requires that both C(t) and V(t) be defined on commensurate units. Practical implementation maps token expenditure to monetary cost via provider pricing, and maps value signals to monetary equivalents via pre-agreed business case assumptions. The precision of TVD as a management signal depends on the quality of the value signal registry, not on consumption measurement, which is already available in existing observability infrastructure.
Of the four quantities required to compute TVD(t): token volume, token cost, value signal events, and value signal monetary equivalents. Organizations currently instrument the first two natively through AI API billing and observability platforms. The third and fourth are absent from all existing observability, FinOps, and AI cost management platforms reviewed in this survey. The result is that TVD is structurally unmeasurable without the AVAL architecture, regardless of the sophistication of an organization's existing AI monitoring stack. This is an instrumentation gap, not a data gap.
An AI deployment D exhibits Consumption Decoupling if and only if: for every authorized token expenditure event e in D, there exists a named value signal s in the Value Signal Registry such that s is bound to e at the time of authorization, s has an agreed monetary equivalent m(s) > 0, and s is verifiable within a defined measurement window w after e. A deployment that does not exhibit Consumption Decoupling exhibits Value Opacity: its token consumption events are not individually attributable to named value signals at authorization time.
An AI deployment D is Value Opaque if and only if TVD(t) is undefined for D at time t, which occurs when no value signal s is registered in the Value Signal Registry for any expenditure event in D over the measurement window [0, t]. Value Opacity is not equivalent to zero value delivery; it is equivalent to the inability to distinguish zero value delivery from positive value delivery using any available instrumentation. In Value Opaque deployments, budget continuation decisions are made on consumption trajectory alone, a condition that structurally enables TVD Growth to persist undetected.
Value Opacity is the named condition that makes TVD unmeasurable. An organization with multiple AI deployments may have some that are Value Opaque and some that are Consumption Decoupled. AVAL's Displacement Monitor surfaces this heterogeneity, enabling governance decisions to target Value Opaque deployments specifically rather than applying budget constraints to the entire AI program.
The AI Value Accounting Layer (AVAL) is a formal instrumentation architecture comprising four components. AVAL is positioned between an organization's AI inference infrastructure and its financial reporting systems. It is provider-agnostic: the same architecture applies regardless of whether inference is consumed from a cloud provider API, a self-hosted model, or a managed AI platform. Fig. 1 illustrates the component architecture.
The Consumption Ledger is a structured, append-only record of all authorized AI token expenditure events, organized at the deployment level. Each ledger entry records: deployment identifier, timestamp, token volume consumed (input and output separately), provider cost per token at time of consumption, total monetary cost, and the value signal identifier bound to this expenditure event at authorization time (or a null marker indicating Value Opacity).
The critical implementation requirement is that the Consumption Ledger operates at event granularity, not aggregate granularity. Monthly cost reports are not sufficient: TVD requires C(t) to be computed over arbitrary time windows and attributed to specific deployments. Event-level logging enables TVD computation at any granularity and makes post-hoc attribution impossible to disguise as prospective binding.
The Value Signal Registry is a catalog of named value signals, each associated with a monetary equivalent m(s) and a measurement window w. A value signal is any verifiable business outcome that can be causally linked to an AI deployment's operation. Examples include: legal review cycle time reduction (measurable in hours recovered per document, convertible to cost savings via loaded attorney rate), customer resolution deflection (measurable in support tickets avoided, convertible to per-ticket handling cost), and code review acceleration (measurable in PR cycle time reduction).
The Registry enforces two constraints that distinguish it from a simple metrics catalog. First, each signal must have a pre-agreed monetary equivalent established at registration time, not retrospectively. Second, each signal must have a defined measurement window: the period within which the signal event is expected to occur following an authorized expenditure event. Signals with indefinite or retrospective measurement windows are not registerable and trigger a Value Opacity flag for the associated deployment.
The Displacement Monitor continuously computes TVD(t) for each deployment and for the organization's AI program in aggregate. It surfaces three alert conditions: (1) TVD Threshold Breach, when TVD(t) for any deployment exceeds a pre-configured organizational threshold; (2) TVD Growth, when the rolling rate of change of TVD(t) is positive over a defined lookback window; and (3) Value Opacity Detection, when a deployment's ledger contains null value signal bindings for more than a configured proportion of expenditure events.
The Displacement Monitor does not make budget decisions. It surfaces information to the Governance Attestation Engine and to financial reporting systems. Budget decisions remain with human governance actors; AVAL provides the information substrate for those decisions, not a replacement for them.
The Governance Attestation Engine enforces Consumption Decoupling as a governance gate. It intercepts spend authorization requests for AI deployments and evaluates whether each request can be bound to a registered value signal at authorization time. Requests that cannot be bound are flagged as Value Opaque; depending on organizational policy, they may be permitted (with mandatory disclosure to financial reporting) or blocked pending signal registration.
The Attestation Engine also generates Consumption Decoupling Attestations: signed records confirming that a specific expenditure event was bound to a specific value signal at authorization time. These attestations are the auditable artifact that enables post-hoc verification of Consumption Decoupling claims and supports regulatory or board-level reporting on AI financial accountability.
Fig. 2 illustrates the TVD trajectories of three archetypal deployment patterns across a representative six-month observation window. Values are directional illustrations of structural patterns, not derived from empirical measurement of specific organizations. The patterns are derived from reported practitioner observations in the sources cited [3][5][11].
Pattern A ("Coupled Deployment") represents a deployment in which value signals were registered at authorization and Consumption Decoupling was enforced from the start. TVD remains near zero throughout the observation window. Pattern B ("Late Attribution") represents a deployment in which value signals were registered retrospectively at month 3: TVD grows unchecked through months 1-3, then stabilizes after attribution is established. Pattern C ("Value Opaque") represents a deployment in which no value signals are registered: TVD grows monotonically and TVD is undefined (Value Opacity) throughout the window. The structural difference between patterns B and C is the eventual attribution in B; both produce identical TVD trajectories through month 3, illustrating why Value Opacity is not detectable from consumption data alone.
| Framework / Practice | TVD as Named Quantity | Value Signal Registry | Consumption Decoupling Gate | Value Opacity Detection | Attestation Record |
|---|---|---|---|---|---|
| NIST AI RMF [6] | No | No | No | No | No |
| ISO/IEC 42001 [7] | No | No | No | No | No |
| FinOps Framework [8] | No (cost-focus only) | No | No | No | No |
| AI Observability Platforms [10] | No | No | No | Partial (cost anomalies) | No |
| EU AI Act (Annex IV) [12] | No | No | No | No | Partial (technical docs) |
| AVAL (this work) | Yes | Yes | Yes | Yes | Yes |
Organizations adopt AVAL capabilities incrementally. The following four-tier maturity model specifies concrete criteria for each tier, enabling self-assessment and targeted investment. Fig. 3 illustrates the four tiers and their relationship to TVD visibility and Consumption Decoupling coverage.
Criterion: The organization has deployed the Consumption Ledger for all production AI deployments and can compute C(t) for any deployment over any time window. Value Signal Registry is empty or minimal; TVD is undefined for most deployments. Governance state: Value Opaque across the portfolio. Characteristic: The organization knows what it spends; it does not know what it receives.
Criterion: The organization has registered value signals for at least 50% of active AI deployments by expenditure volume, with agreed monetary equivalents and measurement windows. TVD is computable for the covered deployments. The Displacement Monitor is operational. Governance state: Partial Value Opacity; TVD visible for covered deployments. Characteristic: The organization can distinguish high-spend high-value from high-spend low-value deployments in covered segments.
Criterion: The Governance Attestation Engine is operational. All new spend authorizations above a defined threshold require value signal binding at authorization time. Retrospective attribution is flagged and disclosed. TVD is computed in near-real-time across at least 80% of expenditure. Governance state: Consumption Decoupling enforced for new deployments; Value Opacity resolved for existing covered deployments. Characteristic: TVD Growth is detectable and reportable within a defined measurement window.
Criterion: Consumption Decoupling Attestations are generated for all authorized expenditure events. TVD is surfaced in financial reporting alongside traditional cost metrics. Board-level or executive reporting includes a TVD summary and Value Opacity coverage ratio. The organization can produce a signed Attestation Record for any expenditure event within 24 hours of request. Governance state: Full AVAL deployment; financial accountability auditable at event granularity. Characteristic: AI program financial accountability is equivalent to capital project accountability: every authorized spend is linked to a named expected return, tracked, and reported.
Of the four organizational conditions relevant to AI program continuity, namely budget availability, stakeholder support, technical capability, and demonstrable ROI, only demonstrable ROI is not surfaced by any existing monitoring system without an AVAL-equivalent architecture. In Value Opaque organizations, budget continuation decisions are made on consumption trajectory alone. This creates a structural asymmetry: consumption growth is visible and generates financial pressure; value delivery is invisible and cannot generate corresponding justification for continuation. Organizations with growing TVD and no AVAL are structurally predisposed to budget events that suspend programs regardless of actual value delivery. This is an existence result: the mechanism exists and produces the outcome; the frequency distribution across organizations is not quantified here.
NIST AI RMF's GOVERN function addresses organizational accountability structures for AI risk, but does not specify financial accountability instrumentation [6]. ISO/IEC 42001's performance evaluation clauses (Section 9) require monitoring and measurement of the AI management system but leave financial outcome attribution out of scope [7]. AVAL is complementary to both: it provides the financial instrumentation substrate that neither standard specifies, and its Attestation Records can serve as evidence artifacts in GOVERN and performance evaluation audits.
The EU AI Act's Annex IV technical documentation requirements for high-risk systems include general performance monitoring obligations [12]. AVAL's Attestation Records are structurally suited to satisfy these obligations for AI systems whose "performance" includes financial return attribution, though regulatory interpretation of this alignment remains to be established.
AVAL's TVD computation requires V(t) to be denominated in units commensurate with C(t). For AI deployments whose value is primarily qualitative, such as improved decision quality, reduced employee cognitive load, monetary equivalents require estimation assumptions that introduce measurement uncertainty. The framework does not resolve this uncertainty; it makes the estimation assumptions explicit and auditable, which is a governance improvement over the current state of no estimation at all.
Consumption Decoupling as defined here applies to spend authorization events. It does not address value signal attribution at the level of individual inference calls within an authorized deployment, which would require instrumentation at a granularity that is technically feasible but organizationally intensive. Tier 3 and Tier 4 organizations may pursue call-level attribution as an extension; the framework's definitions are compatible with that extension.
This paper has introduced three original constructs for enterprise AI financial accountability. Token-Value Displacement (TVD) is the signed divergence between an organization's cumulative AI token expenditure and its cumulative verified value delivery, a quantity that is currently unmeasurable in all existing financial reporting and observability systems. Consumption Decoupling is the governance property of AI deployments in which token expenditure events are bound to named value signals at authorization time; its absence is Value Opacity, a condition in which TVD is undefined and budget continuation decisions are made on consumption data alone. The AI Value Accounting Layer (AVAL) is the four-component instrumentation architecture that makes TVD measurable and Consumption Decoupling enforceable.
The central finding of this paper is an existence result: the structural conditions for TVD Growth and program discontinuation risk are present in any organization that does not implement the AVAL architecture or its functional equivalent. The frequency of this condition across the enterprise AI market is not quantified here; the mechanism is. Organizations at any AVAL maturity tier are better positioned than those with no AVAL instrumentation; the Tier 1 minimum viable deployment, comprising the Consumption Ledger and Value Signal Registry alone, closes the most critical visibility gap with a modest engineering investment.
Future work includes: empirical calibration of TVD distributions across industry verticals, development of a standardized value signal taxonomy by AI deployment type, and integration of AVAL attestation records with AI governance audit frameworks.
© 2026 Arjun Jaggi and Aditya Karnam Gururaj Rao. All rights reserved. Academic citation permitted with attribution; commercial use and derivative frameworks require written permission. The terms Token-Value Displacement, AI Value Accounting Layer (AVAL), Consumption Decoupling, and Value Opacity originate with this work and are subject to the foregoing license.