AI ROI hits Wall: productivity gains, profit gaps persist
The gist
AI is turbocharging productivity but stubbornly failing to deliver profits, leaving most companies with ballooning bills and little to show on the bottom line.
What to know
- Since March 2023, AI spend flipped from fixed project budgets to metered, token-based billing—making costs less predictable and harder to tie to ROI.
- By 2026, PwC found only 12% of CEOs saw both cost and revenue benefits from AI, while a staggering 56% reported no significant financial upside at all.
- Token counts and usage stats can mislead: GPT-5.6 cut output tokens by 54% and task time by 57%, proving more activity doesn’t always mean real value.
AI Budgets Go Metered
Enterprises lost cost predictability when AI spend shifted from fixed project budgets to token-based billing, forcing finance teams to rethink how value is tracked and justified.
Before 2023, enterprises typically justified AI-adjacent technology investments using project-level ROI rather than usage-based cost models. As described in the background material, “Before 2023, enterprises approved technology investments using project-level ROI calculations that estimated total cost, expected savings or revenue, and payback period for bounded projects,” a framework built for contained implementations where costs were known in advance and benefits could be forecast against a fixed budget rather than metered continuously as consumption rose or fell.
That logic changed on March 14, 2023, when OpenAI introduced GPT-4 API pricing and recast AI spending as an operational bill tied directly to use. “[2023-03-14] On 14 March 2023, OpenAI announced the GPT-4 API, establishing a pricing model based on input and output tokens rather than fixed licenses or project charges,” explicitly displacing older fixed-cost software licensing metrics and pushing enterprises toward tracking AI spend through tokenized input/output consumption instead of the bounded, project-charge economics that had governed earlier technology approvals.
ROI Math Breaks at Scale
Most organizations lack the tools and shared frameworks to measure AI’s financial impact, leading to widespread uncertainty and project failures even as spending soars.
By mid-2026, the AI ROI problem had become quantifiable: a PwC study found only 12% of CEOs said AI had delivered both cost and revenue benefits, while 56% reported no significant financial benefit at all. That shortfall is colliding with scale, as Gartner forecast worldwide AI spending would reach $2.59 trillion in 2026, up 47% year over year, intensifying pressure to replace legacy ROI logic and token scrutiny with measures tied to work accomplished, reliability, and the cost of each successful task, as OpenAI’s Sarah Friar argued in CIO Dive.
The harder truth is that enterprises often lack the machinery to measure outcomes even after AI reaches production: IBM found only 29% of executives are confident they can measure returns, and Plug and Play reported 74% of the largest enterprises already have at least one AI system live while 93% are piloting or beyond. According to Forbes, only 21% of AI ownership sits with a CAIO or center of excellence, versus 37% with line-of-business heads, leaving no shared framework for finance; worse, most enterprises review on four quarters, a cadence mismatch that kills projects before the median 14-month path to positive ROI.
Productivity Gains Stall Profits
Despite widespread reports of AI-driven efficiency, only a minority of companies see these improvements reflected in actual earnings or margins.
The paradox is no longer whether AI helps people work better; it is whether those gains reach the income statement. Tech Disruptors pointed to Alaska Airlines, where an agent “save[s] 75% of their time” on destination planning, posts a 90% satisfaction rating, and customers using Alaska Inspires convert at 7% versus 5% through the standard booking widget, yet those are still framed as productivity, satisfaction, and conversion wins that must be connected to revenue creation before they become visible in EBIT or margin.
The broader data shows how often that translation fails: according to McKinsey reporting cited across September and October, “80% Of respondents… reported individual productivity gains,” and “Eighty percent of respondents said AI had improved their productivity, while half said it helped them make better decisions,” but only 30% reported any impact on earnings before interest and tax, while “Just 37 per cent of respondents said AI had contributed positively to their organisations’ EBIT, broadly unchanged from 2025,” and “McKinsey calls AI High Performers remains just 6% of respondents.” In sectors with thin economics, the gap is even starker: Nirav Shah of Endeavor Health said hospital margins are “averaging around 0.75% this year,” intensifying demands for hard ROI that AI often cannot immediately produce.
Token Counts Mislead Outcomes
Rising token usage and activity metrics often mask the true value of AI, as organizations optimize for usage over meaningful business results.
The bottleneck is not whether enterprises can count AI activity, but whether those counts mean anything. Mixture of Experts argued token metrics are vulnerable to Goodhart’s law: teams can inflate usage with elaborate prompts, bloated outputs, retries, or context engineering while the underlying work barely improves, which is why a viral warning that Uber had blown through its annual token budget in four months prompted leaders to rethink incentives; a bigger bill can reflect either valuable work or, as The AI Corner put it, compute “running in circles.”
That is why token efficiency and adoption telemetry can move in the wrong direction relative to value. OpenAI said token prices “plummeted 97% from GPT-4 to GPT-5.4,” and GPT-5.6 “slashed output tokens by 54% while cutting task completion time by 57%,” proving lower token volume can coincide with better performance; likewise, SAP said 30% to 40% of its software development lifecycle is now AI-assisted, yet warned activity is not the outcome. Sid Sharma, Head of Go to Market for ASEAN at OpenAI, made the same point at Techsauce Global Summit 2026, arguing that “The ROI of AI Should Be Measured by Work Completed, Not Just Time Saved,” explicitly moving beyond “hours saved and costs.” That framing also matches MIT’s NANDA study, which found that across hundreds of deployments, 95 percent produced no measurable profit impact because organizations optimized activity rather than value.
Governance, Not Tech, Drives ROI
Weak governance and unclear accountability—not technology—are the main barriers preventing AI from delivering measurable returns across organizations.
The scaling constraint is increasingly organizational, not technical. Business Wire reported NTT DATA found “only 22% of insurers have scaled AI to the production phase,” even though “66% of the insurance workforce has adopted AI tools,” because “the main constraints are not technology-related; they are related to trust, governance and operating models that were not designed for AI.” Earnix similarly concluded MGAA 2026 reflected a shift “from experimentation towards governed, insurance-specific decision intelligence designed to support better business outcomes,” meaning value depends on embedding AI into accountable operating workflows.
That is why ROI ownership is spreading beyond AI teams to finance, business units, and frontline operators. IT Brief New Zealand found only 8% of organisations in Australia and New Zealand can track AI impact with clear KPIs even as 53% use AI in workflows or strategy; 85% of individuals said AI made them more efficient, yet only a small minority measured business outcomes formally, while just 16% were confident staff knew what data was safe to enter and nearly one in four had never refreshed training or policy. The Agile Brand also found 16% of buyers are not tracking AI ROI at all, turning governance gaps into renewal and budget risk.
