Most enterprises don't model AI cost - they model headcount reduction. The invoice that follows is not a billing surprise. It's a structural misclassification that's been accumulating for quarters.

There is a pattern repeating itself across enterprise AI programmes. The business case gets approved on the strength of infrastructure investment, model licensing, and a projected headcount saving. The CFO signs off. Six months later, the invoice bills are itemised, unreadable, and somehow always going up.

The cause is not that AI is expensive. The cause is that AI cost is routinely mismodelled from the start.

92% Increasing AI investment (of enterprises plan to scale AI spend in the next 12 months)

80% Failing to deliver ROI (of enterprise AI programmes fail to translate spend into measurable return)

1% At AI maturity (of CEOs believe their organisation has reached AI maturity, per McKinsey 2025)

The misclassification problem

Most enterprises that attempt a TCO analysis before onboarding AI account for the obvious capital expenditure i.e. infrastructure, licensing, tooling. What lands in the OpEx bucket is narrower than it should be. Token consumption, ongoing data pipeline refresh, model re-evaluation cycles, compliance monitoring, and the engineering time spent maintaining governance controls are routinely treated as capital investment or ignored entirely.

They are not capital investment. They are recurring operational expenditure. And because they are misclassified from the start, they stay invisible until the CFO sees the report.

By the time the exposure surfaces, the misclassification has been accumulating for quarters. The shock is not a billing surprise but it is a structural governance failure.

The CapEx/OpEx boundary in enterprise AI is far more porous than any pre-migration business case acknowledged. The hosting model decision alone, self-hosted versus SaaS, has significant balance sheet implications that are rarely modelled explicitly before architecture commitment.

Classification is not enough

Fixing the classification problem gets you visibility. It does not get you control. The more common failure point is accountability. When no function clearly owns a cost category, it goes unmanaged regardless of how visible it is.

Recognising token spend as OpEx is the classification fix. Attributing it to specific workflows, teams, and decision points is the governance fix. The two are related but distinct problems - one for FinOps and CFO observability, the other requiring architectural enforcement upstream of the model.

Architecture note

Every model and agent call must pass through a centralised gateway enforcing a mandatory metadata schema - WorkflowID, CostCentreCode, BusinessUnitID, UseCaseTag - on every request. Direct model access must be blocked at the IAM layer. Without this, attribution relies on self-reported data that degrades under operational pressure. The gateway is what makes chargeback mechanical and cost control executable.

Who owns what

The whitepaper structures cost ownership across four functions with distinct boundaries. GRC owns the governance policy and the governance cost line. If a function owns the policy, it should own the associated cost, charged back to GRC rather than to the application owner. FinOps monitors, attributes, and runs the chargeback engine. Business units own their token consumption and are accountable for the business outcome that consumption is meant to deliver. Engineering builds and operates the controls.

A further dimension that is often missed: every function consuming AI agents for its own operational purposes generates token spend that must be attributed to its own cost centre. FinOps using a cost monitoring agent owns that agent's OpEx. Engineering using coding assistance agents owns theirs. The attribution principle applies without exception.

What the whitepaper covers

The full paper - published at https://linkedin.com/in/iamrahulus, covers the complete enterprise AI lifecycle from initiation through decommission, with CapEx/OpEx classification at each stage, a worked Australian Big 4 bank chatbot example with cost benchmarks, the gateway-with-mandatory-metadata attribution architecture, and the four-function cost governance model.

It is written for CFOs, heads-of, and principal architects who need a structured framework rather than a vendor pitch.

Browse and download here: https://www.linkedin.com/in/iamrahulus/overlay/1780000523898/single-media-viewer?type=DOCUMENT&profileId=ACoAAAF8aV4Bvw6r2l4RbtjpdYXo8AtL7kQ67iE&lipi=urn%3Ali%3Apage%3Ad_flagship3_profile_view_base%3BN5YOVzMOSVChB%2BCoh0RspQ%3D%3D