The question I hear from many CFOs is simple: how do you get more value from AI spending?
For years we measured software success by adoption: licenses sold, seats used, renewals. With AI that’s not enough. What really matters is the work AI completes: useful outcomes that replace or speed up human tasks.
A new framework: Useful Intelligence per Dollar
OpenAI proposes a more powerful way to measure impact: Useful Intelligence per Dollar. What does this metric answer? Four concrete questions every company needs to ask:
- Does the AI complete work that actually matters?
- How much does each successful task cost?
- Can you trust the result?
- Does each dollar invested produce more value as usage scales?
These questions force you to look beyond price per token. A cheap model might require many attempts and human review. A pricier model may solve the task in one pass. What counts is the full cost to produce a useful result compared to the value it creates.
Start with the work, not the model
The best starting point is a real workflow. Define what it means for something to be “done” and measure it where the work happens.
- Customer support: done = ticket resolved.
- Engineering: done = code change that passes tests.
- Legal: done = contract reviewed to the required quality.
A practical example: a finance team preparing a forecast review. Before the final decision there are hours of reconciliations, moving data, rebuilding slides. Tools like ChatGPT Work can handle large parts of that process, leaving people time to ask what changed and what decision to make. That’s Useful Intelligence per Dollar in action.
How to calculate the cost per successful task
The formula is simple and honest:
- Add the full cost of completing the work (models, compute, human time, reviews, retries).
- Count the tasks that reached the required quality.
- Divide total cost by successful tasks.
This explains why price per token isn’t the whole story. A cutting-edge model can be more efficient at producing the correct result in one attempt, reducing latency, reviews and rework.
OpenAI talks about a family of models GPT-5.6 with three tiers: Sol (top capability), Terra (cost-performance balance) and Luna (fast and cheaper). The idea is to use the layer that optimizes the equation for the workflow’s complexity: Luna for high volume and speed, Terra for depth, Sol when higher capacity reduces retries and reviews.
Trust and dependability matter as much as accuracy
Adoption happens in stages: first AI helps draft, then it integrates context and reasons, later it takes actions and handles exceptions with human oversight.
Dependability has direct economic value. When results are accurate, well-documented and consistent, people spend less time reviewing and fixing, and the organization gains confidence to use AI in more critical processes.
Three practical measures for teams:
- Ready to use: the output meets quality on delivery.
- Needs correction: requires another pass or human editing.
- Needs escalation: someone must intervene to complete the task.
Those metrics tell a richer story than a single accuracy number. Also, before AI acts on real systems, be clear about what data it can access, what systems it can change and when human approval is required.
Scale, compute and the feedback loop
Do the numbers improve as you scale? To find out, track the same workflow over time: tasks that reach quality, total cost and cost per successful task. If completed work grows faster than cost and quality holds or improves, then each AI dollar yields more.
Compute is the heart of this equation. Training models creates future capacity; inference delivers useful work today. Better hardware, more efficient algorithms, smart product design and higher utilization lower the cost per outcome. It’s a cycle: better infrastructure speeds research, research builds better models, better models improve products, better products drive adoption and funding for the next generation.
Practical recommendations for teams and leaders
- Measure results, not tokens. Define “done” for a workflow and measure there.
- Calculate the full cost per successful task, including human time and retries.
- Evaluate dependability with the categories ready to use, needs correction and needs escalation.
- Use model tiers according to the workflow’s economics: speed versus depth.
- Set clear limits on data, access and approvals before allowing automated actions.
If you run finance, operations or product, this gives you a clear map: it’s not just about saving on tokens, it’s about getting more useful work from every dollar you spend on AI.
A final thought
The real promise of AI isn’t just doing the same things faster; it’s letting people spend more time on judgment, creativity and decision-making that truly add value. Measuring “Useful Intelligence per Dollar” turns that promise into a practical goal: more capable models, more reliable outcomes and lower cost per successful task.