The AI bubble is a measurement problem. Boards want the return on every dollar.

By COGScontrol Team · October 31, 2026

Proving AI value to the board means showing, initiative by initiative, the fully loaded AI cost of each business outcome — not the size of the AI bill. For customer-facing AI, that is AI gross margin and a unit cost such as cost per loan or cost per daily active user. For internal AI, it is cost per feature shipped or per case resolved, against the pre-AI baseline.

That distinction has stopped being academic. In October 2026 the AI bubble debate moved from op-eds into share prices, and the question investors put to the hyperscalers is the one boards are now putting to every company that spends on AI: what is each dollar returning? Most finance teams can answer what was spent. Far fewer can answer what it bought.

Is the AI bubble popping?

The people asking are no longer fringe skeptics. Speaking at the Forbes Global CEO Conference, Bridgewater founder Ray Dalio placed AI near the end of a debt-fuelled cycle: "We’re in the part of the cycle that is before that but approaching that." The IMF’s October Global Financial Stability Report warned that markets are increasingly exposed to an AI-driven correction if AI investment fails to produce the profits investors expect. Bank of America’s October fund manager survey found 45% of respondents expect an AI bubble in 2027, even as tech overweights hit their highest level since 2020. And Citadel’s Ken Griffin said investors are almost certainly in one: "Any radical transformation of our economy is generally intertwined with a bubble."

Then the earnings arrived. Alphabet shares fell more than 10% after it reported negative free cash flow of $9.07 billion and its CFO said capex would “increase significantly in 2027.” Amazon posted trailing twelve-month free cash flow of negative $42.4 billion on $77.3 billion of quarterly capex, with Andy Jassy pointing to a “clear line of sight to attractive financial returns.” Microsoft’s CFO Amy Hood, guiding to roughly $170 billion of calendar-year capex, gave analysts an assurance rather than a payback period: "We feel good that the return on every dollar of CapEx we're spending, we're confident in that." Cboe’s JJ Kinahan summarized the mood of the week: "You want to make sure that the reward is worth the risk."

Read carefully, none of this is a verdict that AI doesn’t work. It is a demand to see the return. Whether or not the bubble bursts, that demand does not stay on Wall Street. It flows into every company’s budget review.

Why are boards suddenly asking for proof?

Because governance has caught up with spending. The Deloitte and Center for Audit Quality audit committee survey found 75% of audit committee members now rank AI governance a top-three priority, up from 35% a year earlier — and only 56% are confident they can oversee it effectively. Gartner’s 2027 CEO survey found 44% of CEOs struggle to demonstrate AI-driven financial returns to stakeholders.

The CIO view is starker. In Gartner’s 2027 CIO Agenda survey of 2,002 CIOs, only 13% said their enterprise has realized significant or transformational value from AI, and just 31% have rules for stopping failed pilots. Gartner analyst Alicia Mullery put it in four words: "Volume is not value. AI is creating a reckoning for organizations that lack fundamental disciplines."

Vendors see it in their customers. On IBM’s third-quarter call, Arvind Krishna said "larger enterprises are becoming a lot more cognizant and a lot more aware of what is the true ROI of all of their AI spend." The follow-up to this summer’s tokenmaxxing retreat came in October, when Microsoft reportedly cut its projected internal Claude spending by more than a third and Meta’s internal Claude Code users fell from roughly 60,000 to around 30,000.

Why can't most companies answer the question?

Most companies watch the total AI bill; very few know the unit cost underneath it, and fewer still tie that unit cost to an outcome. The 2026 survey data is consistent on this point:

• Only 12% consistently assess AI value against cost, according to KPMG’s Q3 2026 Global AI Pulse of 2,131 leaders — rising to 48% among organizations with established ROI. Most review cost at approval (61%) and monitor it in operation (59%); almost none close the loop to value.
• 71% track total AI inference and API spend, but only 44% track per-unit inference cost, in Deloitte’s State of AI in the Enterprise survey of 3,500 leaders. Deloitte’s Costi Perricos: “Many enterprises are getting better at watching their total AI bill, but far fewer understand the unit cost underneath it.”
• 53% of CFOs mostly review AI costs after the fact, and the most-cited barrier is “a lack of AI cost governance, processes, or policies,” per Deloitte’s Q3 2026 CFO Signals. Only 25% require a quantified ROI and formal business case before approving AI investment.
• More than 80% of AI spending now sits outside the IT budget, according to BCG’s Applied AI Index, which also found corporate AI spend has doubled to 3.3% of revenue.

Put those together and the problem is structural, not motivational. AI spend arrives as a handful of provider invoices with no project, product, or outcome attached; it is scattered across business-unit budgets; and the business metrics it is supposed to move live in other systems entirely. Total spend is visible. Value per unit is not — because nobody joined the two.

What does proof look like for customer-facing AI?

For AI you sell, proof is gross margin: the AI cost of serving each unit of the business — a loan, a daily active user, a transaction — set against the revenue that unit earns. Customer-facing AI is cost of goods sold, so it shows up first on the gross-margin line. That is exactly where it surfaced this earnings season. GoDaddy’s CFO told analysts its Airo AI product would weigh on margins: "Yes, the gross margin is going to see a little deterioration as we move forward." Figma’s CFO described bearing "the cost of inference without offsetting consumption revenue" on beta AI features, with gross margin improving in its first full quarter of AI credit monetization.

The counterexample shows what proof looks like when a company measures at the unit. Duolingo’s CEO Luis von Ahn reported that the cost of its AI video-call feature, originally estimated at $0.30 per call, "is now under $0.01 per video call. And the reason for that is mainly a move towards open source models." That is a board-ready sentence: a named feature, a unit cost, a before and after, and a reason. It is only possible because someone was tracking cost per call, not just the total model bill.

We go deeper on this end of the ledger — per-feature margin, cost per loan, and stress-testing against token price changes — in AI product gross margin. The underlying formulas are in AI unit economics, and what belongs in the cost base is defined in what is AI COGS.

What does proof look like for internal AI?

For AI you use to run the business, proof is a lower cost per unit of work the company already produces — a feature shipped, a case resolved, a ticket closed — measured against the pre-AI baseline. The October bank earnings calls made the contrast vivid. Citi CEO Jane Fraser gave analysts a measured outcome: "14,000 of our engineers use our main agentic coding tool, Devin, making it much more efficient to do work in-house. We've been in-sourcing work and have reduced our tech vendor spend by 30% year-over-year as a consequence." She added that “the financial benefits from AI and automation significantly outweigh the investments we've made year-to-date.”

The same week, JPMorgan CFO Jeremy Barnum told analysts he was "very reluctant to spend a lot of energy quantifying it," acknowledging that “the benefits are not observable in our current quarterly earnings” and that some would be “offset with token cost.” BNY CEO Robin Vince declined to break out AI spending at all, saying "it could be counterproductive to break it out and to start obsessing about it publicly."

Barnum’s point is fair — heavyweight, bespoke ROI studies do waste people’s time. That is an argument for measurement that runs continuously and automatically, not for no measurement. Outside the largest banks, the cost of not measuring is visible: Apollo chief economist Torsten Slok observed that mainstream companies are treating "token spending not as a source of growth but as a cost center that weighs on margins." An internal AI initiative without a value denominator will be read exactly that way at the next budget review. The mechanics of attributing internal AI cost by team and project are covered in measuring the ROI of internal AI initiatives.

What belongs on a board AI value scorecard?

One page, two sides: the AI in your product and the AI in your operations, each expressed as a cost per business outcome that reconciles to the P&L. A board does not need token counts or model names. It needs to see, for each material initiative, what a unit of value costs, whether that is improving, and who owns it.

Two disciplines make the scorecard credible. First, every row must reconcile to what the company actually paid — costs estimated from list prices won’t survive an audit committee. Second, the value side must come from systems of record, not surveys: loans funded from the loan system, DAU from product analytics, features shipped from Jira or GitHub, cases from the support platform.

What should finance do this quarter?

The companies that keep investing through a correction will be the ones that can show the return. Five steps get a finance team from “here is the bill” to “here is what it bought” before the next board meeting:

1. Inventory every AI initiative and label its side. Customer-facing (cost of goods sold) or internal (operating expense). The split decides which P&L line the value appears on.
2. Pick one value denominator per initiative. Cost per loan, per DAU, per feature, per case — a metric the business already reports. The AI Value Ledger framework maps common initiatives to their denominators.
3. Attribute the real invoices, fully loaded. Model tokens, cloud infrastructure, and the platforms around them, allocated to initiatives and reconciled to the bill — not estimated from a price list.
4. Connect the business metrics. Bring the denominators in from their systems of record so unit cost updates on its own instead of in a quarterly spreadsheet exercise.
5. Agree on targets and stop rules. Gartner found only 31% of enterprises have rules for stopping failed pilots. A scorecard row without a target is a report; with one, it is a decision.

This is the discipline of AI value management: cost attribution, unit economics, and P&L reconciliation applied to AI at both ends of the business. COGScontrol was built to run it continuously — aggregating spend across OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google Vertex AI, and the underlying cloud, attributing it to initiatives, and joining it to business metrics that arrive via CSV upload, API, direct entry, or Google Sheets. Try your own numbers in the AI unit economics calculator, or see the full system on the features page.