56% of CEOs Can't Prove AI Works. Here's the Fix.

By Rajesh Beri | July 18, 2026

---

On July 17, 2026, OpenAI CFO Sarah Friar published something unusual: a detailed framework for measuring whether AI spending is actually creating value. Not a product announcement. Not a pricing page. A scorecard — with the implicit admission that most enterprises buying OpenAI's products cannot currently answer the most basic question about them: is this working?

The timing is not accidental. Axios reported the framework is OpenAI's direct response to a growing enterprise AI cost reckoning, as companies increasingly route work to cheaper models and demand clearer returns on their investments. Fortune noted that at McKinsey's 24th annual Global CFO Forum — roughly 100 finance chiefs from over 30 countries — about two-thirds reported that strategy now reports to the CFO, up from less than a third five years ago. The conversation at this year's forum shifted decisively from AI experimentation to enterprise-wide transformation and cost accountability.

Friar's scorecard arrives at a moment when the gap between AI spending and AI value has become a boardroom crisis. PwC's 2026 Global CEO Survey found that 56% of global CEOs cannot point to measurable business impact from their AI spend. Only 29% of executives say they can measure AI returns with confidence. Gartner forecasts total AI spending will reach $2.59 trillion in 2026 — a 47% year-over-year increase. And Larridin's AI ROI Measurement Framework estimates that 72% of enterprise AI investments are destroying value through waste — not because the AI does not work, but because measurement was never designed into the deployment.

The enterprise AI industry has a $2.59 trillion measurement problem. Friar just proposed four questions to solve it.

The Core Problem: You're Measuring the Wrong Thing

For two decades, enterprise software success was measured through adoption: seats purchased, users active, licenses renewed. Every SaaS company in the Fortune 500 runs on these metrics. When AI arrived, most organizations applied the same measurement playbook — and it broke immediately.

"The basic economic question facing CFOs and other business leaders is whether the value of the work AI completes grows faster than the cost of producing it," Friar wrote. "Answering that question requires looking more deeply than a metric such as cost per token."

This is a remarkable statement coming from the CFO of the company that invented cost-per-token pricing. Friar is explicitly telling her own customers: the number on your invoice is not the number that matters.

Here is why. A cheaper model may have lower token prices but fail more often, retry more frequently, or produce output that requires human correction. A more capable model may cost more per token but complete the same task in one pass. What matters is not what you paid per million tokens — it is the total cost of producing a successful outcome, measured against the value that outcome creates.

Consider a real-world example from McKinsey's research: roughly two-thirds of organizations have not begun scaling AI enterprise-wide, and only about 39% report any enterprise-level EBIT impact from AI at all. The MIT NANDA Initiative's study of 300 public deployments found that 95% of generative AI pilots fail to deliver measurable financial returns. The 5% that succeed share one common trait: they defined what "done" means before they started building.

That is precisely what Friar's scorecard demands.

The Four Questions: Unpacking "Useful Intelligence per Dollar"

Friar's framework asks four sequential questions that build on each other. Together, they form what she calls "Useful Intelligence per Dollar" — the ultimate metric for the AI age. Here is what each question means operationally, and why most enterprises are failing at every single one.

Question 1: Is AI completing work that matters?

Start with the output, not the input. Not how many tokens were consumed. Not how many API calls were made. Not how many users logged in. How many customer issues did AI resolve? How many code changes shipped? How many contracts were reviewed? How much time was returned to people?

"Tokens create value when they transform into work people can use," Friar writes. "The best place to begin is with one workflow. Define what 'done' means and measure that outcome in the system where the work happens."

For a support team, "done" might mean a customer issue resolved without escalation. For engineering, a code change that passes automated tests and review. For legal, a contract reviewed accurately and on time. For finance, a forecast model that produces numbers the team can act on without rebuilding the spreadsheet manually.

This sounds obvious. It is not. Genpact's research with HFS, surveying over 2,000 enterprise leaders, found that 85% of executives say fragmented data, ungoverned processes, aging systems, and undertrained talent are actively impeding their AI initiatives. Most organizations have AI tools running in production without a clear definition of what a "completed task" looks like. They are measuring activity instead of accomplishment.

Question 2: What does each successful task actually cost?

This is where the measurement gap becomes a financial gap. The calculation Friar proposes is deceptively simple:

Add the full cost of completing the work. Count the tasks that met the required quality bar. Divide.

But "full cost" is the operative phrase. For an AI-assisted workflow, the full cost includes: model API charges, tool and retrieval costs, retries and failed attempts, human review time, rework when the AI gets it wrong, and the infrastructure overhead of running the whole pipeline. Most enterprises track only the first item on that list.

This is why the token price war is a distraction. Gartner predicts that by 2028, the costs associated with AI coding will exceed the average developer's salary — not because tokens are expensive, but because the rate of consumption is outstripping price declines. Agentic AI workflows consume 5 to 30 times more tokens per task than a standard chatbot interaction. Your cost-per-token went down 97% since GPT-4. Your total bill went up.

Friar connects this directly to OpenAI's new model family: "GPT-5.6 Sol reached a new state of the art on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model." The pitch is not cheaper tokens — it is fewer tokens needed to get the same work done. That is the difference between cost-per-token and cost-per-successful-task.

Question 3: Can people depend on the result?

Dependability is the hidden multiplier in the AI ROI equation. When results are accurate, well-sourced, and consistent, people spend less time reviewing, correcting, and repeating the work. When they are not, every "AI-completed" task comes with a human tax that no one is tracking.

Friar proposes tracking three outcomes for every AI workflow:

• Ready to use: The result met the quality bar as delivered.
• Needs correction: The result required another attempt or human edits.
• Needs escalation: A person needed to step in and finish the work.

These three categories are far more useful than aggregate accuracy metrics. A model that scores 92% on a benchmark but produces output that requires human correction 40% of the time in your specific workflow is not 92% accurate for your purposes — it is 60% dependable. And that 40% correction rate has a cost: the botsitting problem that is consuming enterprise productivity gains before they reach the bottom line.

"Capability earns first use," Friar writes. "Dependability makes AI part of how work gets done."

Question 4: Does each AI dollar produce more value as usage grows?

The final question tests whether your AI investment compounds or plateaus. Track the same workflow over time: How many tasks met the quality bar? What was the total cost? What was the cost per successful task? If completed work grows faster than total cost while quality holds or improves, each AI dollar is producing more value.

This is where compute strategy meets business strategy. OpenAI's argument — and it is an argument for buying more OpenAI products — is that better models, more efficient inference, purpose-built hardware, and smarter routing all compound over time. Each generation of infrastructure trains more capable models. Better models improve products. Better products drive adoption. Adoption funds the next generation.

The enterprise version of this argument is more nuanced: it depends on whether you are building reusable AI infrastructure or launching disconnected pilots. A company that invests in shared identity, trusted connectors, curated knowledge bases, evaluation frameworks, observability, model routing, and reusable agent patterns will see compounding returns. A company that funds 47 independent AI experiments with no shared infrastructure will see 47 independent cost centers.

Framework #1: The Enterprise AI ROI Calculator

Friar's scorecard provides the theory. Here is how to turn it into a practical calculation your CFO can run this quarter. This calculator translates "Useful Intelligence per Dollar" into a number you can put on a slide.

Step 1: Define Your AI Workflow Unit

Pick one workflow. Define what "completed task" means. Examples:

Function	Workflow	"Completed Task" Definition
Customer Support	Ticket resolution	Issue closed without human escalation, CSAT ≥ 4.0
Engineering	Code review assist	PR review completed, all comments addressed, merged
Legal	Contract review	Contract analyzed with no missed clauses, delivered on time
Finance	Report generation	Forecast model delivered, accepted without manual rebuild
Sales	Lead qualification	Lead scored and routed, converted within SLA

Step 2: Calculate Cost Per Successful Task (CPST)

CPST = Total Fully Loaded Cost / Number of Successful Tasks

Where Total Fully Loaded Cost includes:
  + Model API spend (tokens consumed × price per token)
  + Tool/retrieval costs (RAG, web search, database queries)
  + Retry costs (failed attempts × per-attempt cost)
  + Human review cost (review hours × hourly rate)
  + Rework cost (correction hours × hourly rate)
  + Infrastructure overhead (hosting, orchestration, monitoring)

Step 3: Calculate the AI Value Ratio

AI Value Ratio = Value Created per Task / CPST

Where Value Created per Task =
  (Human time saved × hourly rate)
  + Revenue protected or generated
  + Risk/cost avoided
  + Capacity created for higher-value work

If AI Value Ratio > 1.0: AI is creating net value
If AI Value Ratio < 1.0: AI is destroying value on this workflow

Step 4: Track Dependability Rate

Dependability Rate = Ready-to-Use Tasks / Total Tasks Attempted

Target by maturity stage:
  Pilot: ≥ 60% (explore, learn)
  Validation: ≥ 75% (prove the workflow)
  Production: ≥ 85% (scale with confidence)
  Optimized: ≥ 92% (compound returns)

Step 5: Calculate Return on AI Compute (ROAC)

Month-over-Month ROAC:
  = (Successful Tasks This Month / Total AI Cost This Month)
  ÷ (Successful Tasks Last Month / Total AI Cost Last Month)

If ROAC > 1.0: Each AI dollar is producing more value over time
If ROAC < 1.0: You're hitting diminishing returns — investigate

Run this calculation quarterly. Compare across workflows. Fund what compounds. Kill what does not.

Framework #2: The AI Investment Maturity Assessment

Friar's blog post also contains a buried investment framework that most readers will skip: she argues enterprises should manage AI investments as a portfolio across three tiers, with funding decisions tied to maturity stage. Here is that framework made explicit, with governance gates for each stage.

Tier 1: Broad Access (Explore)

What it is: Organization-wide access to AI assistants for everyday productivity — email drafting, meeting summaries, research, code completion.

Budget allocation: 20-30% of total AI spend

Measurement: Adoption rate, user satisfaction, time saved per user per week

Governance gate: Usage analytics and spend controls in admin console. Track which teams adopt, which models they use, and whether demand is growing or plateauing.

Kill signal: Adoption stalls below 40% after 90 days. Users revert to non-AI workflows despite access.

Tier 2: Function-Specific Workflows (Validate)

What it is: Dedicated AI workflows built for repeatable, measurable business processes — support ticket resolution, code review automation, contract analysis, financial reporting.

Budget allocation: 40-50% of total AI spend

Measurement: Cost per successful task, dependability rate, AI Value Ratio, cycle time reduction

Governance gate: Defined quality bar for each workflow. Evals that reflect real tasks including edge cases. Clear ownership and accountability for workflow performance.

Kill signal: Dependability rate below 60% after two quarters. CPST exceeds human-only cost for the same task. AI Value Ratio below 0.8 with no improvement trajectory.

Scale signal: Dependability rate above 85%. AI Value Ratio above 2.0. Month-over-month ROAC consistently above 1.0. Clear demand from adjacent teams for the same capability.

Tier 3: Strategic Bets (Transform)

What it is: AI-first business processes built around proprietary company context — workflows that create competitive advantage because they leverage data, domain knowledge, or process expertise that competitors cannot replicate.

Budget allocation: 20-30% of total AI spend

Measurement: Revenue impact, market position, competitive moat created

Governance gate: Executive sponsorship. Dedicated engineering team. Privacy and compliance review before scaling. Defined exit criteria if the bet does not pay off within 12-18 months.

Kill signal: No measurable business impact after 12 months. Technical debt accumulating faster than value creation.

The Critical Layer: Shared Infrastructure

Budget allocation: 10-15% of total AI spend (funded centrally, not by individual teams)

What it includes:

• Identity and access management for AI agents
• Trusted connectors to enterprise systems
• Curated knowledge bases and RAG infrastructure
• Evaluation frameworks and quality benchmarks
• Observability and monitoring (token usage, cost attribution, error rates)
• Model routing logic (match task complexity to model tier)
• Reusable agent patterns and templates

This layer is what separates organizations that see compounding AI returns from those stuck in pilot purgatory. Every dollar spent on shared infrastructure reduces the marginal cost of launching the next AI workflow.

Why OpenAI Is Publishing This Now

Friar's scorecard is not philanthropy. It is strategy.

OpenAI faces three converging pressures. First, the model price war has compressed margins across the industry. From GPT-4 to GPT-5.4, token prices fell 97%. Chinese models like DeepSeek are offering comparable performance at a fraction of the cost, and enterprises are noticing. The Georgetown CSET researcher Sam Bresnick told the Financial Times: "Enterprises have an incentive to shift some of their workload to cheaper models. Why would you pay a premium for Anthropic, OpenAI models when for a lot of the workloads you need, the Chinese models are generally workable?"

Second, OpenAI is approaching an IPO. The company is valued at $852 billion and approaching the $1 trillion range. An IPO could come as soon as this summer. To justify that valuation, OpenAI needs its enterprise customers to prove — with numbers — that AI spending is a strategic investment, not an experiment. If CFOs cannot quantify AI's value, they will cut AI budgets. If they can quantify it, they will increase them. Friar's scorecard is designed to produce the second outcome.

Third, GPT-5.6 just launched with a three-tier model family — Sol, Terra, and Luna — that only makes economic sense if enterprises are measuring cost per successful task rather than cost per token. If you are optimizing for token price, you buy Luna for everything. If you are optimizing for cost per successful outcome, you route complex tasks to Sol (fewer retries, higher first-pass success) and simple tasks to Luna (fast and cheap). The scorecard framework justifies the tiered pricing model.

This is not a criticism. It is exactly what a good CFO should do: create the measurement framework that makes your product's value proposition visible. But enterprises should adopt the scorecard while understanding who wrote it and why.

The Scorecard's Blind Spots

Friar's framework has real value, but it also has deliberate omissions that enterprises should fill.

It does not address vendor diversification. The scorecard assumes you are measuring AI value within a single vendor's products. In reality, most enterprises run multiple models from multiple providers. The cost-per-successful-task calculation becomes significantly more complex when you are comparing OpenAI, Anthropic, Google, and open-source models across the same workflows. The scorecard should be applied per-workflow and per-model, then compared across vendors.

It underweights the cost of model switching. Friar argues for matching the model to the task — use Luna for fast work, Sol for complex reasoning. But model switching has hidden costs: prompt engineering per model, eval suite maintenance per model, integration testing per model, and the cognitive overhead on engineering teams managing a model zoo. The model-agnostic architecture debate is real, and this scorecard sidesteps it.

It does not account for opportunity cost. The AI Value Ratio measures whether AI creates net positive value. But it does not measure whether that value is the best available use of the same dollars. An AI workflow with a 1.5x Value Ratio sounds good until you discover that investing the same budget in process redesign would yield 3x. The scorecard measures AI efficiency, not capital allocation efficiency.

It ignores governance and compliance costs. As China's new agent recall framework and Illinois' SB 315 demonstrate, AI governance costs are rising rapidly. Mandatory audits, compliance reporting, agent identity management, and recall infrastructure are all costs that should be in the "fully loaded" calculation but are absent from Friar's framework.

What to Do Monday Morning

If you are running AI in production — or spending money on AI that is not yet in production — here is your action list:

Week 1: Pick your top three AI workflows by spend. For each one, define what "completed task" means. If you cannot define it, that is your first finding: you are spending money on a workflow with no success criteria.

Week 2: Calculate Cost Per Successful Task for each workflow. Include everything: API costs, retrieval costs, retries, human review, rework, infrastructure. Compare this to the human-only cost of the same task. If the AI version costs more, you have a problem to solve — or a pilot to kill.

Week 3: Implement the three-outcome tracking system: Ready to Use, Needs Correction, Needs Escalation. This is a logging change, not a product change. Add it to your existing AI observability stack. Start calculating Dependability Rate.

Week 4: Run the AI Investment Maturity Assessment against your current AI portfolio. Classify every active initiative into Tier 1 (Explore), Tier 2 (Validate), or Tier 3 (Transform). Apply the governance gates. Identify what should scale, what should be killed, and what shared infrastructure gaps are holding back the entire portfolio.

Ongoing: Calculate Return on AI Compute monthly. Track the trend. If ROAC is declining, investigate whether the problem is model performance, workflow design, data quality, or simply using a frontier model for a task that a smaller model handles fine.

The Real Scorecard

Friar's framework is the first time an AI company's CFO has publicly told the market: stop measuring adoption, start measuring work accomplished. That alone makes it worth reading.

But the real scorecard is broader than any single vendor's framework. It includes vendor diversification risk, governance compliance costs, opportunity cost against non-AI alternatives, and the shared infrastructure investments that determine whether AI compounds or stalls.

The enterprises that win in the next 12 months will not be the ones spending the most on AI. They will be the ones that can answer, with data, one question: for every dollar we spend on AI, how much useful work comes back?

If you cannot answer that question today, you now have a framework to start. The 56% of CEOs who cannot prove AI is working have run out of excuses.

---

Continue Reading

• 100% of CIOs Budget for AI. Half Already Blew Their Budgets — The enterprise AI FinOps crisis that created the demand for Friar's scorecard.
• 89% of S&P 500 Firms Are Stuck in AI Purgatory. MIT Proved It — Why most enterprises cannot answer the four questions yet.
• The $9 Billion AI Deployment War Just Went Nuclear — Microsoft Frontier Company and the deployment model that sits behind the scorecard.