Datagras
Back to blog

Research11 June 2026

The Hidden Cost of AI: Why Token Counting Isn't Enough

The FinOps Maturity Gap

The 2025 State of Cloud FinOps report (n=1,400 enterprises) contains a striking statistic: 98% of respondents now have formal processes for tracking AI and LLM spend. Yet only 2% can answer the question "what business outcome did that spend produce?" This is the FinOps maturity gap - organisations have mastered cost visibility but have not yet connected spend to value.

The gap matters because without outcome attribution, AI budgets are managed by the wrong metric. Teams optimise for cost-per-token without understanding which tokens are generating revenue, which are serving compliance obligations, and which are pure waste. The result is blunt cost-cutting that reduces capability indiscriminately, or unchecked spend growth because nobody can prove the cost is unjustified.

Why Token Counting Falls Short

Token counting is necessary but not sufficient for AI FinOps. The fundamental problem is that tokens are an infrastructure metric, not a business metric. A 10,000-token call that summarises a contract clause that unblocks a $2M deal has a very different value profile than a 10,000-token call that regenerates a tooltip that nobody reads. Identical infrastructure cost, radically different business impact.

To close the attribution gap, you need three additional data layers on top of token counts: (1) workflow tagging - which business process or feature triggered this LLM call; (2) outcome linkage - what downstream event (conversion, resolution, decision) correlates with the call completing successfully; and (3) quality scoring - did the output actually serve its purpose, or was it discarded, retried, or overridden by a human.

Without these layers, cost-per-token is like measuring a factory's efficiency by counting electricity usage without knowing what the factory makes or how much of it sells.

Building an Outcome-Attributed AI Cost Model

The practical path to outcome attribution follows three phases. Phase 1 is instrumentation: integrate SAVI's SDK at the LLM gateway layer and begin tagging every call with a workflow_id and feature_id. This takes one to two weeks for most teams and requires no changes to the underlying LLM integration. Phase 2 is baseline: run for 30 days to establish call volume, cost, and quality distributions per workflow. Phase 3 is attribution: connect your CRM, product analytics, or ERP data to SAVI via the Outcome API, linking downstream business events to the upstream LLM calls that contributed to them.

After Phase 3, SAVI's attribution dashboard shows cost per outcome - not cost per token. You can see that your customer-service summarisation workflow costs $0.003 per resolved ticket, that your contract review feature costs $0.18 per reviewed clause, and that your internal knowledge Q&A system costs $0.0009 per answered question. These are numbers a CFO can evaluate against a business case.

The 2% Playbook: What Leading FinOps Teams Do Differently

The enterprises in the 2% cohort share three practices that distinguish them from the 98%. First, they treat AI spend attribution as a product requirement, not a finance afterthought - every new AI feature ships with outcome tracking instrumentation as part of the definition of done. Second, they review AI unit economics monthly at the product level, not quarterly at the infrastructure level. Third, they use cost attribution data to inform model selection: workflows where outcome quality is high even with cheaper models get migrated to cheaper models proactively, without waiting for a cost spike to force the conversation.

Getting Started

SAVI's 14-day free trial includes full workflow tagging, cost attribution, and the Outcome API with sample connector templates for Salesforce, HubSpot, and Mixpanel. Most teams reach Phase 2 (baseline established) before their trial expires. The gap between the 2% and the 98% is not technical sophistication - it is instrumentation. Start there.