July 17, 2026, (Inside AI) — Chief financial officers are abandoning seat-based software metrics. The new question is not how many employees logged in, but how much work AI actually completed.
OpenAI this week proposed a unified scorecard for the age of AI: Useful Intelligence per Dollar. The framework arrives as enterprises struggle to connect model benchmarks to balance-sheet impact. It defines value through four lenses: useful work produced, cost per successful task, dependability, and value at scale.
The shift is overdue. For decades, software success was measured by adoption—licenses sold, users active, renewals secured. Generative AI breaks that model because output, not access, creates economic value. A model that answers a question is not the same as one that resolves a customer issue, ships a code change, or reviews a contract.
OpenAI’s core argument is simple: the full cost of a successful outcome matters more than the price per token. A cheaper model may require more attempts, more human review, or more time. A frontier model might complete the same task in one pass, reducing total compute and labor. The metric forces organizations to measure what AI finishes, not what it attempts.
The Four-Part Scorecard
The framework starts with work accomplished. OpenAI urges teams to define “done” for a single workflow—customer issue resolved, code change passing tests, contract reviewed accurately—and measure that outcome in the system where work happens. For a finance team preparing a forecast review, AI can handle data gathering, reconciliation, and slide rebuilding, freeing analysts to ask what changed and why.
Next is cost per successful task. The calculation is straightforward: total cost divided by number of successful outcomes. Total cost includes model fees, employee time, human review, retries, and rework. OpenAI acknowledges that complex coding or research workflows consume more compute but can create much more value. The lowest price per token does not always produce the lowest cost per outcome.
Dependability is the third pillar. As AI moves from drafting to taking action, accuracy, sourcing, and consistency become economic variables. OpenAI recommends tracking three outcomes: tasks completed without human correction, tasks escalated for human review, and tasks requiring rework. These reveal whether AI is genuinely reducing the work involved in completing a project. Clear boundaries around safety, security, and governance create the foundation for deeper use.
Finally, value at scale examines whether economics improve over time. Companies should track the same workflow, measuring how many tasks meet the quality bar, total cost, and cost per successful task. If completed work grows faster than total cost while quality holds, each AI dollar is producing more value. Compute sits at the center: training compute builds future capability, inference compute delivers useful work today.
OpenAI’s new GPT‑5.6 model family, released last week, illustrates the tiered approach. Sol is the flagship; Terra balances performance and cost; Luna is the fastest and most affordable. On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with max reasoning set a new state of the art while using 54% fewer output tokens than another leading model. The company says each generation should make existing tasks cheaper and enable new kinds of work.
Where the Framework Falls Short
The scorecard is elegant but incomplete. It assumes organizations can cleanly define “successful” outcomes, yet many knowledge-work tasks have ambiguous endpoints. A legal contract review might be “accurate” but still require negotiation. A code change might pass tests but introduce technical debt. The framework also sidesteps the cost of integration—connecting AI to legacy systems, retraining staff, and managing change.
Competing voices offer different lenses. Andreessen Horowitz has argued that AI ROI should be measured by revenue growth and margin expansion, not task-level efficiency. McKinsey emphasizes the need to track “decision velocity”—how quickly organizations act on insights. OpenAI’s focus on cost per task may undervalue strategic shifts that AI enables, such as entering new markets or launching AI-native products.
Historical parallels are instructive. The 1990s productivity paradox showed that IT investments took years to appear in national statistics because firms needed time to reorganize work. AI may follow a similar path: early adopters see task-level gains, but systemic value emerges only when processes and roles are redesigned. OpenAI’s framework is a starting point, not a destination.
OpenAI positions ChatGPT Work as the enterprise layer that builds on ChatGPT Enterprise’s security and compliance foundation. The company says its shared intelligence platform—spanning ChatGPT, Codex, and the API—allows improvements in one layer to benefit all products. The ultimate goal, it says, is AI that helps people do more meaningful work and spend more time on judgment and creativity.
The scorecard’s real test will be whether CFOs adopt it. Most still rely on vendor-reported metrics like token usage or model accuracy. Useful Intelligence per Dollar demands a deeper look at work systems, outcomes, and costs. It may be the first framework that treats AI not as a tool, but as a worker whose output can be measured, managed, and improved.
