AI Value Is Reaching Earnings Calls, but the Attribution Test Has Barely Begun
Company disclosures and a practitioner’s account of agent use point to the same management problem: AI spending is easier to see than the work, revenue, and risk it changes.
By Mira Solis · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human research credentials or firsthand experience.
Key points
- Santander reported business value from AI initiatives, while other companies described operating benefits without isolating AI’s share of financial performance.
Sources: S1
- A Fortune commentary argues that firms should judge AI by completed work rather than token consumption, but its central examples are anecdotal and its cited developer study is observational.
Sources: S2
- The practical gap is attribution: reported company outcomes may be real without proving that AI, rather than other changes, produced them.
The new question is not whether AI has value
Corporate discussion of AI is moving beyond broad productivity promises and toward financial results. CNBC reported that companies are beginning to identify benefits in earnings discussions as investors look for measurable returns from large technology outlays. Santander said its AI initiatives generated business value during the first nine months of the year, framing the benefit as both lower-value work removed from employee workloads and potential effects on revenue and costs. That is a more concrete claim than a generic assertion that AI will transform the business, but it is not the same as a clean measure of incremental profit.
Sources: S1
The distinction matters because AI is entering businesses through processes already affected by management decisions, staffing, pricing, customer demand, software modernization, and economic conditions. Santander’s finance chief also said AI was being embedded across the bank’s processes, making it difficult to state precisely how much came from AI. That caveat is unusually important: broad deployment can make a tool operationally useful while making causal attribution harder, not easier.
Sources: S1
Other reported earnings comments illustrate the same tension. Novartis’s chief executive said AI had started to improve its profit-and-loss performance on a tangible basis, pointing to marketing and sales productivity while describing research-and-development productivity as an area under examination. Standard Chartered’s finance chief linked AI to customer relationships, income growth, core-process improvement, and operating leverage. Carlsberg, by contrast, said AI had affected internal, commercial, pricing, marketing, and supply-chain processes but that benefits were not yet visible at scale. These are different stages of adoption, not interchangeable proof points.
Sources: S1
Sources: S1
Measured business value is stronger than a broad earnings narrative
Santander’s reported business-value figure is the clearest disclosed outcome in the supplied material, but the evidence packet does not explain the methodology used to assign that value to AI. The reported result therefore establishes that management has put a number on the initiative; it does not, on its own, establish how an outside evaluator could reproduce the estimate or separate it from related operational changes. Novartis and Standard Chartered offer an even looser form of evidence: company executives connected AI to broader performance measures, while neither disclosure in the supplied material assigns a specific portion of those measures to AI.
Sources: S1
Big technology companies face a related attribution issue from the other direction. Meta said AI investment was delivering incremental advertising revenue as total quarterly revenue rose, while Google Cloud revenue more than doubled. Those results may be commercially significant, but a company-level revenue change contains many drivers. A statement that AI contributes to an outcome is not equivalent to a controlled comparison showing what would have happened without the AI investment. This is not a claim that the reported benefits are false; it is a limit on what the cited disclosures can demonstrate.
Sources: S1
Inference: the most credible executive AI claims will increasingly be those that link a defined workflow to a before-and-after outcome, identify the people and systems involved, and account for the full cost of delivery. Earnings commentary can signal that value is emerging, but it is a weak substitute for evidence that survives a repeatable external review. The supplied company disclosures support the need for that distinction because they range from a stated business-value estimate to qualitative claims tied to broad financial performance.
Sources: S1
Sources: S1
Usage cost is a dependency, not a value measure
A Fortune commentary supplies a useful counterweight to earnings-call attribution: it argues that firms should assess the cost of finished work rather than impose blanket limits on AI use. The authors describe an internal funding arrangement at Accenture that covered a set monthly amount for Claude and required higher usage to be assigned to a project, while Accenture disputed that it had imposed a firm limit on access. The account highlights a real management dependency: where AI work is performed under fixed-price contracts, usage costs can affect the provider’s margin even when the employee sees only an access restriction.
Sources: S2
The commentary’s strongest operational point is that token charges are not necessarily the largest cost in an AI-enabled workflow. It says a model of deployments in several industries found that repeated attempts, idle capacity, monitoring, and staff time spent checking outputs made total costs materially higher than the token bill. If that framing applies to a company’s environment, reducing visible model usage may leave the harder costs intact. But the article does not provide the underlying model, its assumptions, or independent results in the supplied material, so its claim should be treated as an argument rather than a generally established benchmark.
Sources: S2
The same caution applies to its anecdotes. The authors say a portfolio company used agents to create a website rather than use an agency, and that a non-technical marketing hire took on work beyond a conventional marketing role. Such examples describe expanded capability, not merely faster completion. Yet the material does not provide a controlled comparison of quality, maintenance burden, security, intellectual-property exposure, customer outcomes, or the counterfactual cost of the work. They are plausible management cases, but they cannot prove that the pattern will hold across organizations.
Sources: S2
Sources: S2
What independent evaluation would need to test
The Fortune commentary cites an observational study of developers that found large differences in token spending and completed work across levels of usage, while explicitly acknowledging that association does not establish causation. That admission is central. Heavy users may finish more work because they are assigned different tasks, have better tools, are more experienced, or work in teams that already operate differently. A completed-work metric is more relevant than raw consumption, but it still needs comparable workloads and quality checks before it can support a hiring, budgeting, or restructuring decision.
Sources: S2
For executives, the practical test is to examine a specific workflow rather than ask whether AI is raising the whole company’s productivity. Define the task, baseline labor and vendor cost, output standard, review effort, failure and rework rate, usage cost, and the business outcome that matters. Then compare results across similar work with and without the intervention where feasible. This approach follows the tension visible across the supplied sources: earnings disclosures seek financial relevance, while the commentary argues that local work outcomes reveal value that seat-level token budgets can miss.
What could change this assessment is more granular evidence. Santander could disclose how it estimated business value and how it prevented overlap with other initiatives. Novartis, Standard Chartered, Carlsberg, Meta, and Alphabet could connect particular AI systems to defined operating metrics and disclose relevant costs and controls. The commentary’s claims would become more persuasive with underlying data on its examples, quality audits, and independent replication of the developer findings. Until then, the evidence supports cautious confidence that AI can create value in specific settings, not a conclusion that broad enterprise gains have been cleanly attributed.
Why it matters
The next phase of enterprise AI will be governed less by headline adoption and more by whether companies can connect a tool’s full operating cost, output quality, and business effect without overstating causality. That discipline affects capital allocation, contract pricing, workforce decisions, and investor confidence.