AI Is Getting Cheaper. So Why Is ROI Still So Hard?
Artificial intelligence is becoming cheaper, more capable, and more widely available at extraordinary speed. Normally, economics tells us what should happen next: when the cost of supplying something falls, more of it becomes available at a lower price and usage expands.
That appears to be happening with AI: model competition is increasing, smaller specialized models are proliferating, inference costs are falling rapidly, and capabilities continue improving. Companies are responding by spending aggressively.
Yet there is a strange contradiction. According to McKinsey's Technology Trends Outlook 2025, 78% of surveyed organizations already use AI in at least one business function, and 92% of executives plan to increase investment over the following three years. But only 1% of leaders say their organizations have reached full maturity in AI deployment.
AI supply is expanding, investment is exploding, and adoption is widespread. So where is the mature economic payoff?
The Supply of AI Is Expanding Rapidly
One of the most important developments in McKinsey's report is not simply that AI models are getting better—they are getting cheaper.
As inference becomes less expensive and smaller models perform tasks that once required giant frontier models, the effective supply of AI shifts outward:
Lower cost → greater access → greater usage
McKinsey describes a "small-model explosion" where quantization and distillation allow smaller models to deliver strong performance using far less computation, with median inference costs declining roughly 50-fold annually across benchmarks.
Simultaneously, model capabilities are advancing in reasoning, multimodal tasks, coding, and agentic behavior. Supply is shifting outward while demand shifts outward too, driving accelerating investment.
Companies Are Spending as Though the Future Has Already Arrived
McKinsey estimated that Alphabet, Amazon, and Microsoft were each on track to spend roughly $70 billion to more than $100 billion on AI-related capital expenditures in 2025 across data centers, chips, models, software, infrastructure, and enterprise deployment.
This connects directly to broader economic trends: business investment has been strengthening, capital-goods orders are rising, corporate profits have rebounded, and productivity is improving. The latest BLS data show nonfarm business productivity rising 2.2% over the past year, with output up 2.5% and hours worked up only 0.2%.
However, that does not prove AI caused those productivity gains, nor does it mean the investment boom has already paid off. McKinsey suggests almost the opposite.
The Technology May Be Ahead of the Organization
McKinsey notes that the gap between AI's potential and realized progress exists because companies need time to adapt organizations, develop complementary innovations, and reskill workforces. The sequence is rarely instantaneous:
Buy AI → Experiment → Redesign workflows → Reorganize teams → Change skills → Build governance → Measure results → Improve → Realize productivity
This explains a major economic puzzle:
Why can companies spend enormous amounts of money on AI while still struggling to demonstrate enterprise-wide ROI?
The bottleneck is no longer the model—it is the organization wrapped around it.
AI Is Moving From a Technology Problem to an Operating-Model Problem
Phase one centered on capability (writing, coding, analyzing, reasoning). Phase two centers on hard operating-model questions: Which workflows change? Which tasks automate? Where do deterministic rules override models? How are outcomes measured and held accountable?
McKinsey emphasizes that organizations must change operating models, talent, and governance quickly enough to capture value. Michael Chui notes that the primary differentiator is rewiring these components around AI to create measurable impact.
The core question shifts from: “Can we deploy AI?”
To: “Can we redesign the business around it without losing control of cost, risk, quality, and accountability?”
Cheaper AI Does Not Automatically Mean Better ROI
As inference costs collapse, companies may simply consume more AI across teams via chatbots, coding assistants, internal apps, and customer agents. Infrastructure, monitoring, and governance costs rise alongside subscriptions.
The cost per inference falls, but the total AI bill still rises. ROI requires a tangible numerator:
Revenue gained + Cost avoided + Time saved + Quality improved + Risk reduced + Decisions improved
Without connecting usage to outcomes, cheap AI simply creates more inexpensive activity—not value.
Smaller Models Could Change the Economics Again
Sending every enterprise query to a giant frontier model is economically inefficient. A mature architecture relies on selective routing:
Use the least expensive capability that reliably satisfies the requirement.
- Frontier Models: Complex multi-step reasoning.
- Specialized Small Models: Focused domain tasks.
- Classical ML & Code: Structured data analysis and deterministic rules.
- High-Consequence Workflows: AI recommendation + deterministic guardrails + human approval.
The future belongs to orchestration: routing work between models, tools, rules, data, and humans based on cost, risk, and value.
The Labor Market May Already Be Reacting
While business investment and productivity rise, hiring remains subdued and labor's share of nonfarm business output fell to 52.9% in the second quarter of 2026—the lowest in BLS records back to 1947.
Companies expecting AI to alter future staffing may choose to preserve workforce flexibility today:
AI investment rises → Hiring becomes selective → Vacancies aren't backfilled → Workflows are redesigned → Productivity becomes clearer → Different skills are hired later
The Jobs May Change Before They Disappear
McKinsey reports that 46% of leaders identify skill gaps as a barrier to AI adoption, while over 20% of employees report minimal training. Meanwhile, demand for data scientists, engineers, and solution-oriented AI roles remains strong.
The immediate divide is not human vs. machine, but:
AI-enabled worker vs. non-AI-enabled worker
Workers who combine domain expertise with AI tools, data literacy, judgment, and workflow redesign become vastly more productive, while traditional task execution reorganizes.
More Autonomous AI Makes Governance More Important
Agentic AI systems that execute multi-step tasks independently increase the need for explicit governance regarding privacy, security, bias, IP, and transparency.
Greater capability → Greater autonomy → Greater potential value → Greater consequences when errors occur
The most sophisticated AI organizations will not simply maximize autonomy, but will strategically determine where autonomy creates value and where control and deterministic rules must remain intact.
The 1% May Be the Most Important Number in the Report
While 78% use AI and 92% plan to spend more, only 1% report full deployment maturity. Almost everyone is experimenting; almost no one has fully rewired their business model yet.
We are not at the end of the AI transformation—we have primarily finished installing the hardware. Redesigning business operations around it is the next multi-year phase.
What Happens When the Organization Finally Catches Up?
The capital investment arrived before the organizational transformation needed to extract value. As companies mature, the results will manifest in productivity, employment composition, wages, corporate margins, and operating costs.
As powerful AI models become commoditized and cheap, competitive advantage shifts from access to execution:
The real question is no longer whether companies can get access to AI, but whether they can turn increasingly abundant intelligence into measurable economic value.