OpenAI Proposes 'Useful Intelligence per Dollar' as the AI Value Metric

OpenAI Reframes AI ROI Beyond Token Economics

According to the OpenAI Blog, enterprise finance leaders increasingly ask how to extract more value from AI spending. OpenAI proposes abandoning traditional software metrics—seat adoption, active users, license renewals—in favor of a new framework: “Useful Intelligence per Dollar,” a four-part scorecard that measures whether the economic value of AI-completed work grows faster than production costs.

The shift reflects a fundamental tension in enterprise AI deployment. Cheaper tokens do not guarantee better outcomes if a model requires multiple attempts, extended human review, or iterative refinement. Conversely, expensive tokens may justify their cost if the model completes complex tasks in a single pass, reducing downstream labor. OpenAI’s argument centers on total cost of success, not marginal cost of computation.

The Four Pillars of the Scorecard

According to OpenAI, the metric answers four distinct questions:

Work volume and relevance. Does the AI system complete tasks that matter to the business? Examples include customer issues resolved, code changes shipped, contracts reviewed, time returned to teams, or decisions improved through better information availability. The measure captures whether tokens translate into actionable outputs.

Cost per successful task. What is the full economic cost—compute, human intervention, retry overhead—to deliver one completed task? This differs fundamentally from cost-per-token by including the friction of reliability failures.

Reliability and dependability. Can users trust the results without rework? A 90% accurate output that requires human validation carries hidden costs that token pricing never surfaces.

Scaling economics. Does each AI dollar produce more value as usage grows? This tests whether adoption compounds returns or encounters diminishing returns due to task saturation or quality degradation.

Practical Implementation: A Finance Workflow Example

OpenAI illustrates the framework through a finance team preparing a forecast review. The pre-decision work—finding forecasts, moving data into spreadsheets, identifying deltas, reconciling tabs, rebuilding slides, verifying sums—is automatable. ChatGPT Work, OpenAI’s assistant platform, can handle much of this workload, freeing the team to focus on judgment-intensive questions: what changed, why, and what action follows.

This reframing inverts traditional productivity metrics. Rather than measuring hours saved in aggregate, the scorecard captures how AI reallocates time from routine data-wrangling to expert reasoning—a qualitative shift with downstream business impact.

Why This Matters

OpenAI’s scorecard directly challenges how CIOs and CFOs justify generative AI spending to boards and investors. If enterprises adopt this framework, it will shift vendor competition from model-scaling races (bigger context windows, faster tokens) toward outcome engineering: architecting workflows where AI eliminates the most labor-intensive, error-prone steps. This may accelerate adoption of agentic AI—systems that operate across tools and maintain context—since multi-step workflows generate measurable, auditable value. However, the framework’s success depends on enterprises actually implementing outcome tracking, a data discipline many organizations lack. Teams that cannot define “done” or measure task completion rates will struggle to adopt the scorecard, potentially widening the adoption gap between sophisticated and basic AI users.