The AI Economics Stack

Definition

The AI Economics Stack is a three-level framework for tracing how infrastructure and model consumption become useful work and enterprise value.

It answers three questions:

1. What did intelligence cost to produce and consume?
2. What did the AI-enabled application accomplish?
3. What changed for the enterprise?

The three levels

The stack organises AI economics into three distinct but connected layers. Each layer has its own economic logic, measurement challenges and ownership.

Level 1: Intelligence Production Economics

Object

The capacity and processing used to create AI outputs.

Includes

• Energy and facilities
• Accelerators, CPU, memory
• Storage and network
• Model access and tokens
• Throughput and caching
• Routing and reserved capacity
• Engineering and operations

Core questions

• What are we consuming?
• What determines unit cost?
• Where is demand appearing?
• Which workloads are efficient?
• Should capacity be rented, reserved, managed or owned?
• What is the provider and sovereignty exposure?

Measures

• Cost per input, output and cached token
• Cost per inference
• Tokens per second
• GPU utilisation
• Batch rate and cache hit rate
• Model mix
• Cost by workload
• Forecast variance
• Capacity commitment utilisation
• Energy and carbon where relevant

Primary owners

• FinOps
• Infrastructure and platform engineering
• AI engineering
• Procurement
• IT finance

Failure mode

The organisation optimises cost per token while ignoring whether output is usable.

Level 2: AI Application Economics

Object

The task, workflow, product feature or agent that uses AI.

Includes

• Prompt and context
• Retrieval and tool calls
• Orchestration and model output
• Evaluation and human review
• Correction and escalation
• Workflow integration
• Adoption and behavioural effects

Core questions

• What work is being attempted?
• What counts as success?
• Is output accepted and acted upon?
• What is the full cost per successful outcome?
• Is AI removing work or moving it?
• Are quality, latency and risk appropriate?

Measures

• Task completion and useful-output rate
• Acceptance and correction rates
• Rework and review time
• Escalation and latency
• Cost per accepted output
• Cost per successful outcome
• Workflow cycle time
• Agent calls per outcome
• Adoption and appropriate-use rate

Primary owners

• Product and engineering
• Operations and process owners
• Risk and compliance
• Business functions

Failure mode

The application produces impressive outputs but no net workflow improvement.

Level 3: Enterprise Value Economics

Object

The economic and strategic outcome.

Includes

• Revenue and contribution margin
• Operating cost and capacity
• Quality and customer outcome
• Risk and resilience
• Compliance and strategic option
• Capability and opportunity cost

Core questions

• What changed?
• Who owns the benefit?
• Was capacity or value captured?
• What evidence supports attribution?
• Is the outcome durable?
• Is this the best use of capital?
• Should the initiative scale, change or stop?

Measures

• Incremental contribution margin
• Realised cost reduction
• Redeployed capacity
• Throughput and loss avoided
• Risk exposure and customer retention
• Strategic milestone
• Attribution confidence
• Payback and economic yield
• Portfolio value at risk

Primary owners

• CEO and CFO
• COO and business unit executives
• Portfolio leadership
• Board

Failure mode

The organisation claims strategic value without a credible evidence chain.

Conversion between levels

Production efficiency

AI work capacity / production cost

Useful for engineering and sourcing decisions.

Application efficiency

Successful application outcomes / fully loaded application cost

Useful for product and operations decisions.

Enterprise yield

Measured enterprise value / full cost, risk and operational burden

Useful for capital allocation decisions.

Core principle

Examples of this principle in action:

• Cheaper inference can increase rework
• Higher token throughput can generate unused output
• Faster task completion can move a bottleneck downstream
• Higher adoption can increase risk
• Lower human effort can reduce learning and resilience
• A valuable use case can still be a poor portfolio investment

Relationship to AI Value Management Layers

The AI Economics Stack and the AI Value Management Layers solve different problems and should coexist.

Do not replace the five AIVM layers. Use both frameworks together: the Stack for economic analysis, the Layers for governance progression.

Diagnostic questions

For a material use case, ask:

Level 1: Production

• Can we allocate full production cost?
• Do we know model, provider and workload?
• Can we forecast demand?
• Do we know the sourcing alternatives?

Level 2: Application

• Is success defined?
• Do we measure accepted output and review?
• Can we calculate cost per successful outcome?
• Do we understand behavioural effects?

Level 3: Enterprise

• Is there a baseline?
• Is an owner accountable?
• Is value captured?
• Is attribution credible?
• Has the portfolio decision changed?

Worked example: Customer support

Production (Level 1)

• Model calls and retrieval
• Token cost and caching
• Routing efficiency

Application (Level 2)

• Resolved query rate
• Escalation and review burden
• Response quality
• Repeat contact rate

Enterprise (Level 3)

• Cost to serve
• Customer retention
• Customer satisfaction
• Agent capacity redeployment
• Risk exposure