The AI ROI is Hidden in the Lack of Cost Data
Without internal telemetry, companies cannot attribute AI costs to customers or functions, repeating cloud computing mistakes but with greater complexity.
July 26, 2026 · 4 min read
TL;DR: Measuring AI ROI requires three data sources: costs, business, and telemetry. AI providers do not provide customer-level attribution, so companies must capture that data internally. Without it, scaling AI is a blind bet.
Business leadership wants to scale AI. Budgets are tripling. Adoption is growing. But when the CFO asks which initiatives are actually profitable, the answer is often an awkward silence. It is not that organizations lack visibility into costs, but rather that the cost data they have was never designed to answer that question, according to an InfoWorld analysis.
The three-data problem
Managing cloud spending taught CFOs that billing without business context is just noise. To calculate cloud ROI, two sources were merged: cost data and business data. AWS reveals account, region, resource, and tags. By combining them with customer and product maps, ROI becomes clear. But AI is more difficult. It requires three sources: costs, business data, and telemetry—the automatic collection of data from disparate sources that helps clarify what happened and why.
An executive may have AI invoices and customer revenue, but they have no way of connecting them to business value. The token count on an OpenAI invoice does not specify which customer triggered which call, which function it served, or if the prompt produced a commercial result. That data does not exist in the provider's billing.
AI providers won't fix it
The situation will not change soon, because AI providers are not in the business of attributing a company's costs to its customers. They are in the business of selling tokens. The granularity they expose is what their billing systems require, not what a CFO needs.
Compare what AWS gives you with what an AI provider gives. AWS exposes resource IDs, account hierarchies, region, SKU, tag metadata, and usage per minute. Every dollar can be attributed to a workload, a team, or a customer segment if tagged correctly. An AI provider's invoice only gives tokens consumed by model, with optional grouping by API key. No request-level attribution. No customer ID. No function mapping. No prompt result. No identification of retries. Multi-agent workflows collapse into a single token count.
Imagine a large bank receiving a multi-million dollar AI invoice every month. It has no visibility into which parts of the business generated which part of the cost, so it cannot allocate them. If a company wants to know which AI cost drove which customer or function, it must capture that data itself, within its application, before the call goes out.
Cloud lessons that don't apply
Cloud computing taught CFOs to demand tags and metadata. But AI is inherently more opaque. Language models do not generate detailed invoices per request; they aggregate consumption. Furthermore, AI workflows are complex: a single prompt can involve multiple calls to models, vector databases, and orchestration. Without internal telemetry, it is impossible to know which part of the cost corresponds to which business outcome.
Some AI FinOps startups are emerging, offering proxy solutions that capture metadata before sending requests to providers. But adoption is slow. Meanwhile, CFOs face a paradox: they want to scale AI, but they cannot justify the spending because they lack the data to calculate ROI.
Consequences for companies and users
This lack of visibility has direct consequences: companies may be investing in AI projects that do not generate value, or worse, cannibalizing margins without knowing it. Product teams cannot optimize costs because they do not know which functions are more expensive. AI providers, for their part, have no incentive to change the billing model, as opacity benefits them by making price comparisons difficult.
For end users, this can translate into more expensive or less optimized AI products, as companies pass on cost uncertainty to prices. Additionally, the lack of attribution hinders auditing and regulatory compliance, especially in regulated sectors like finance or healthcare.
What readers should know
If you work in a company implementing AI, demand that your engineering team implements telemetry from day one. Without request-level attribution data, you will never be able to calculate the real ROI. AI providers will not do it for you. The solution involves capturing metadata (customer ID, function, result) in your application before sending the request to the model. Observability and AI-specialized FinOps tools can help, but the ultimate responsibility is yours.
In short, AI ROI is not a problem of missing data, but of poorly designed data. Companies that solve this will have a competitive advantage: they will be able to scale AI with confidence, knowing which initiatives are profitable and which are not.