The burgeoning integration of Artificial Intelligence (AI) across enterprise landscapes is presenting a significant challenge for financial leadership: the inability to definitively link AI expenditure to tangible business outcomes and profitability. As organizations enthusiastically scale their AI initiatives, with budgets often tripling and adoption rates soaring, a critical question from the CFO and the board looms large: which of these AI-driven endeavors are actually generating a positive return on investment? The stark reality, according to industry observers, is that most organizations are ill-equipped to answer this fundamental question, not due to a lack of cost visibility, but because the cost data provided by AI vendors is inherently insufficient for such granular analysis.
The Cloud Cost Management Analogy and Its Limitations
Lessons learned from the protracted journey of managing cloud spend offer a partial, yet ultimately inadequate, framework for addressing the AI and ROI quandary. For years, finance departments grappled with the complexity of cloud billing, recognizing that raw cost data without business context amounted to little more than noise. The path to achieving cloud ROI typically involved stitching together two primary data sources: cloud cost data and corresponding business data. Cloud providers like Amazon Web Services (AWS) offer a degree of granularity, detailing usage by account, region, resource, and metadata tags. By merging this with internal customer and product mappings, organizations could begin to discern the ROI of their cloud investments.
However, AI presents a far more intricate problem. To truly understand AI’s financial impact, a trifecta of data sources is required: cost, business context, and crucially, telemetry. Telemetry, in this context, refers to the automatic collection of data from disparate sources that illuminates the "what" and "why" behind AI usage. While an executive or engineering lead might possess AI invoices detailing token consumption and simultaneously track customer revenue, the critical link between the two remains elusive. The token count on an AI provider’s invoice, for instance, does not specify which customer initiated a particular query, which specific feature it served, or whether the output of the prompt translated into a measurable business outcome. This granular linkage data is simply not present in the provider’s standard billing statements.
AI Providers: Focused on Tokens, Not Enterprise ROI
The inherent limitations of current AI billing structures are unlikely to be rectified by the AI providers themselves in the near future. These companies are primarily in the business of selling computational resources, often measured in "tokens," rather than providing detailed enterprise-level cost attribution. Their billing systems are designed to reflect the granularity required for their own operational and sales processes, not the sophisticated business context demanded by a CFO or a board of directors seeking to justify significant investments.
A comparative analysis starkly illustrates this disparity. AWS, a mature cloud provider, offers billing statements that expose resource IDs, account hierarchies, geographical regions, Stock Keeping Units (SKUs), extensive tag metadata, and even usage broken down by the minute. This level of detail allows well-established FinOps (Financial Operations) teams to construct robust unit economics models, attributing every dollar spent to specific workloads, teams, or customer segments, provided the tagging strategy was implemented effectively.
In contrast, a typical AI provider invoice might reveal the total tokens consumed by a specific model, with an optional grouping by API key. This is often the extent of the provided resolution. There is typically no request-level attribution, no direct customer ID linkage, no feature mapping, no indication of prompt outcome, and no explicit identification of retries. Complex, multi-step agentic workflows, which can involve numerous individual calls to AI models, are often collapsed into a single aggregate token count. Consider a large financial institution receiving a multi-million dollar monthly AI invoice. Without granular visibility into which business units or specific applications were responsible for which portions of that cost, effective allocation and accountability become impossible.
The Imperative for Internal Data Capture
Consequently, if an enterprise aspires to understand the precise AI costs driving specific customer engagements or underpinning particular features, it must undertake the responsibility of capturing this data internally. This requires instrumenting applications to log relevant contextual information before an AI model call is even made to the provider. This proactive data collection forms the bedrock of any meaningful AI ROI analysis.
The Trifecta of Data Sources for AI ROI Measurement
Achieving robust AI ROI measurement necessitates the integration of three distinct data sources into a unified analytical model:
-
Cost Data: This encompasses the direct expenditure incurred for AI model usage, typically provided by the AI vendors through invoices and billing statements. While this data shows the raw expense, it lacks the crucial context for allocation.
-
Business Data: This involves internal organizational data that maps AI usage to specific business units, products, features, customer segments, or revenue streams. This data is generated and managed within the enterprise’s own systems, such as CRM, ERP, and product analytics platforms.
-
Telemetry Data: This is the most nuanced category, involving the automatic collection of operational data that provides a detailed account of AI interactions. This includes information about the specific prompts used, the parameters of the AI calls, the model versions invoked, the duration of processing, any intermediate steps within an agentic workflow, and crucially, the outcome or result of the AI’s processing. This data must be captured at the application or service level.
When these three sources are meticulously stitched together, they unlock the unit economics that are now indispensable for informed AI investment decisions. This allows for the calculation of metrics such as cost per customer interaction, margin per feature, profitability per agent workflow, and ROI per model choice. None of these critical metrics can be accurately derived from billing data alone, nor can they be calculated solely from telemetry or business data in isolation. They demand a holistic view, meticulously mapping the expenditure to tangible business outcomes.
The Escalating Urgency of Agentic AI
The complexity and urgency of this data integration challenge are amplified by the rise of agentic AI. While single-call inference models, where one request yields one cost and one outcome, represent a relatively straightforward scenario, agentic workflows introduce a new dimension of cost complexity. An AI agent is designed to decompose a larger task into a series of smaller, sequential steps. Each of these steps may involve a call to an AI model. Furthermore, these workflows can incorporate sophisticated logic: falling back to different models when an initial attempt fails, retrying operations when results are suboptimal, or integrating with external tools that themselves incur costs. Consequently, a single user request can trigger dozens, if not hundreds, of inference calls across multiple AI providers and services, with costs compounding in ways that are impossible to disaggregate from standard provider invoices.
Without robust telemetry that captures the granularity of each agent step, understanding which specific components of an agentic workflow are profitable becomes an insurmountable task. Aggregate costs will manifest on invoices weeks later, by which point the workflow may have been operating at scale, customers onboarded, and potentially unprofitable retry loops engaged thousands of times. As agents become more prevalent, the volume of cost-generating events without attached business context grows by an order of magnitude. The window of opportunity to instrument these systems effectively before they become unmanageable is rapidly closing.
Transformative Insights with Integrated Data
The integration of cost, business, and telemetry data fundamentally transforms the conversation around AI investment. Suddenly, seemingly equivalent AI development approaches cease to be interchangeable. Five different methods for achieving the same AI capability might appear comparable based on initial adoption metrics, but the integrated data can reveal cost differentials of up to tenfold. This enables teams to select the approach that delivers a similar business outcome at a fraction of the cost, simply because the financial implications are now transparent.
Product teams can design features with an inherent awareness of margin from the architectural phase, rather than discovering cost inefficiencies during a post-launch budget review. Engineering teams can make informed decisions about model architectures, weighing cost-per-outcome data alongside traditional metrics like latency and quality. Leadership can evaluate AI initiatives with the same rigor applied to any other capital allocation decision, focusing on unit economics rather than solely on engagement charts. Aggregated invoices, when properly contextualized, can track the cost per customer interaction, while engagement metrics can reveal the margin per feature. The reliance on gut instinct for model selection is replaced by data-driven comparisons of real cost-per-outcome results.
Within seconds, stakeholders can identify which AI features are driving profitability, which warrant scaling, and which are underperforming and should be deprecated. This level of insight is precisely what organizations are seeking to optimize the immense benefits that AI promises.
Navigating the "Build Trap"
The escalating nature of AI costs presents a significant challenge, with boards of directors unlikely to wait 18 months for an internal project to yield clear ROI. This creates a potent temptation to "build it anyway," a notion fueled by the advancements in AI coding tools that empower small engineering teams to ship substantial functionality within a quarter. The perceived tractability of the instrumentation layer, the normalization of cost data, and the potential for a senior engineer to draft a semantic model over a sprint can all contribute to this allure.
However, this path often leads to what can be termed a "build trap," primarily due to three critical factors:
-
Unprecedented Volume: A production AI footprint generates millions of telemetry events per hour, a volume that escalates dramatically with agentic adoption. Real-time ingestion, correlation, and attribution at this scale represent a fundamentally different and more complex problem than prototyping a feature in an afternoon. It requires a permanent, highly reliable operational system that must function accurately every minute of every day.
-
Dynamic Vendor Landscape: Cost data is delivered through delayed billing cycles from AI providers, each often possessing non-interoperable data schemas. These schemas can change without notice, and new AI providers enter the market monthly, each introducing its own unique taxonomy and metering approaches. The system required to manage this data is not a one-time build; it necessitates continuous maintenance against a rapidly evolving target that frequently outpaces internal release cycles.
-
Business-Critical Infrastructure: The confluence of volume and the dynamic vendor landscape means that the system for managing AI cost attribution becomes business-critical infrastructure. The CFO and the board will base significant capital allocation decisions on the data it produces. When schema drift goes unnoticed for weeks, or when an agent telemetry stream fails to correlate with a vendor that has quietly altered its billing API, the cost of error is not a minor cleanup effort. It translates into potentially a quarter of misallocated capital, impacting strategic planning and financial performance.
The traditional build-vs.-buy decision for engineering leaders is thus recontextualized. The question is no longer simply "Can we build this?" The honest answer for many might be yes. The more pertinent question is whether the marginal hour of the organization’s strongest engineers is best spent stitching together cost data, telemetry, and business outcomes, or building the AI products that are intended to generate the revenue being measured. The capability to achieve comprehensive AI cost attribution and ROI analysis is reproducible within weeks. The strategic choice lies between dedicating the next 18 months to building such a system or spending that same period actively leveraging the insights it provides to drive business growth and optimize AI investments.
