Fully loaded cost
The fully loaded cost of AI is the complete cost of operating an AI-enabled workflow during a defined period. It includes direct technology spend and the labor, implementation, review, governance, and support required to produce accepted work.
This measure prevents a license price or model bill from standing in for the economic cost of the workflow. It is useful when comparing AI options, calculating unit economics, or testing whether an initiative produced net value.
What costs belong in the calculation?
Choose a cost boundary that matches the decision. Relevant categories can include:
- software subscriptions, licenses, and committed spend;
- API tokens, inference, hosting, storage, retrieval, and networking;
- integration, evaluation, testing, monitoring, and maintenance;
- implementation, training, change management, and program work;
- human review, correction, escalation, and support;
- security, privacy, legal, procurement, and governance effort; and
- incidents, failed runs, rework, migration, and retirement.
Not every item must be allocated perfectly. Material exclusions should remain visible, and shared costs should use an approved allocation rule or stay explicitly unallocated.
How is fully loaded AI cost calculated?
A general formula is:
Fully loaded AI cost = direct AI spend + attributable labor + allocated shared costs + implementation and governance costs
State the period and classify each amount as observed, contracted, forecast, or estimated. Keep one-time implementation costs separate from recurring operating costs before combining them for a specific analysis.
For workflow comparison, calculate:
Fully loaded cost per accepted unit = fully loaded AI workflow cost / accepted completed units
Use accepted work in the denominator. Generated drafts, model calls, and attempted runs may explain consumption but do not represent completed business output.
How should labor cost be calculated?
Use the organization's approved labor-cost method. Salary or wages alone omit employer-paid benefits and other employment costs. The US Bureau of Labor Statistics Employer Costs for Employee Compensation program measures wages, salaries, and employer-provided benefits per hour worked (BLS calculation method).
An internal fully loaded rate may also allocate payroll costs, facilities, management, or other overhead. Document the method and apply it consistently. Do not copy a generic external percentage into an investment case when internal finance data is available.
Include time spent reviewing and fixing AI output. Correction rate identifies pre-completion edits, while rework captures repair after completion. Both can materially change cost per accepted unit.
Fully loaded cost versus AI TCO
AI total cost of ownership covers the lifecycle cost of acquiring, operating, governing, and retiring an AI system. Fully loaded cost usually applies that broad logic to a defined period, team, workflow, or output.
TCO can support a build-versus-buy or multi-year architecture decision. Fully loaded workflow cost can support a monthly operating review or comparison between use cases. The boundaries may overlap, so every report should state what is included.
How does fully loaded cost reveal value leakage?
AI value leakage occurs when expected gains fail to become captured outcomes. Hidden review, rework, duplicate tools, unused commitments, and support work can consume an apparent benefit.
Compare fully loaded cost with quality, speed, accepted throughput, and realized value. The cheapest model or tool may produce a higher total cost if it requires more review or fails more often.
The FinOps Foundation describes AI cost management as a cross-functional practice that connects consumption and allocation with business value (FinOps for AI).
Common fully loaded cost mistakes
Common failures include counting only vendor invoices, omitting internal labor, hiding shared infrastructure, mixing forecast and actual amounts, ignoring failed work, and dividing cost by raw model output.
Another mistake is adding every possible overhead allocation until the measure becomes unusable. Use a proportionate boundary, document it, and keep the calculation reproducible enough for the decision at hand.