Back to Blog
AI FinOps2026-03-22

Why AI Cost Attribution Is Broken (And How to Fix It)

AI adoption is accelerating across SaaS companies, ISVs, and enterprise software teams.

But while usage is scaling rapidly, one critical capability has not kept up:

Cost attribution.

Most organisations cannot answer simple questions about their AI spend:

  • Which feature is driving the most cost?
  • Which customers are the most expensive to serve?
  • How much AI usage is internal vs product?
  • How much is development vs production?
  • Which region or product line is driving spend?

This is not a reporting problem.

It is a structural problem.

The Illusion of Visibility

AI providers give you dashboards.

These dashboards show:

  • Token usage
  • Model usage
  • Total cost

At first glance, this looks like visibility.

In reality, it is not.

Provider dashboards answer:

"How much did you spend?"

They do not answer:

"Why did you spend it?"

The Root Problem: AI Lacks Context

An AI request is just an API call.

It contains:

  • A prompt
  • A model
  • A response

What it does not contain is business context.

It does not tell you:

  • Which feature triggered the request
  • Which product it belongs to
  • Whether it was dev or production
  • Whether it served a customer or an internal user
  • Which region it originated from

Without that context, cost attribution is impossible.

Why Traditional Approaches Fail

Most organisations try to solve this problem after the fact.

1. Spreadsheet Allocation

Finance teams take the provider invoice and attempt to allocate cost manually.

This is:

  • Time-consuming
  • Inaccurate
  • Not scalable

2. Engineering Logs

Engineering teams attempt to reconstruct usage from logs.

This creates:

  • Inconsistent tagging
  • Missing data
  • High operational overhead

3. Cost Per Token Thinking

Many teams rely on "cost per token" as a primary metric.

But tokens are not the problem.

Tokens are just the unit of measurement.

The real problem is:

What generated those tokens.

The Real Requirement: Dimensions

To understand AI cost, you need to understand it across dimensions.

Common dimensions include:

  • Feature
  • Product
  • Customer or tenant
  • Region
  • Environment (dev vs production)
  • Internal vs external usage

Without these dimensions, AI spend is a black box.

With them, it becomes measurable.

Example: When Attribution Fails

Imagine a SaaS company launches a new AI feature.

A month later, AI spend doubles.

Without attribution:

  • Finance sees increased cost
  • Engineering sees increased usage
  • No one knows which feature caused it

With proper attribution:

  • The increase is tied to a specific feature
  • Token usage per request is visible
  • Inefficient prompts are identified
  • Cost can be optimised

This is the difference between reactive and proactive cost management.

The Dev vs Production Problem

One of the most common attribution failures is mixing development and production usage.

Engineering teams experiment heavily with AI.

Without clear separation:

  • Development traffic inflates production costs
  • Product margins become unclear
  • Forecasting becomes unreliable

This is especially common in fast-moving ISVs and SaaS companies.

Internal AI vs Product AI

Another major blind spot is internal AI usage.

AI is used across:

  • Support tools
  • Sales automation
  • Internal workflows

If this usage is not separated:

  • Internal costs are mixed into product cost
  • Product profitability is distorted
  • Operational efficiency is hidden

Again, this is not a cost problem.

It is an attribution problem.

The Fix: Capture Context at the Source

AI cost attribution cannot be solved after the fact.

It must be solved at the point of request.

This means:

Every AI call should include metadata such as:

  • Feature
  • Product
  • Environment
  • Customer (where applicable)
  • Internal vs external usage

This transforms AI usage from:

Raw API calls

into:

Structured, analysable data

The Role of an AI Gateway or LLM Gateway

An AI Gateway or LLM Gateway provides the ideal control point.

By routing all AI requests through a central layer, organisations can:

  • Enforce required metadata
  • Standardise provider access
  • Capture consistent telemetry
  • Analyse usage across dimensions

Instead of relying on fragmented logs, the gateway becomes a system of record.

From Cost Tracking to AI FinOps

When cost attribution is solved, new capabilities emerge:

  • Feature-level cost analysis
  • Customer-level profitability
  • Region-based cost insights
  • Token efficiency optimisation
  • Budget control and forecasting

This is where AI FinOps begins.

Not with dashboards.

But with structured attribution.

The Bottom Line

AI cost attribution is broken because AI usage lacks context.

Provider dashboards show totals, not drivers.

Spreadsheets and logs cannot reconstruct intent reliably.

To fix this, organisations must move from:

  • Token-level metrics

to:

  • Dimension-based attribution

By capturing context at the point of request and analysing AI usage across features, products, environments, and customers, AI spend becomes understandable — and controllable.

In the era of AI-powered software, cost attribution is not optional. It is foundational.