Analysis
Enterprises with full visibility into agentic AI operating costs are 5x more likely to report established ROI. The problem isn't just that costs are seller-made — it's that the enterprises that track them are the only ones winning.
The deployment bill for agentic AI is not a single invoice. It is a stack of costs that most enterprises cannot see, let alone measure. According to a CIO.com synthesis of vendor TCO research, AI service bills account for roughly 30 percent of total cost. The remaining 70 percent is the hidden AI stack: infrastructure, integration, governance, and operational overhead that rarely appears in any single budget line. For enterprises deploying agentic systems at scale, this visibility gap is not a reporting byproduct. It is the ROI barrier.
The data converges on a single structural claim: enterprises that cannot see their full operating costs cannot prove their returns. The surveys — from IBM, CrewAI, Salesforce, Deloitte, Accenture, and BCG — do not agree on methodology, sample size, or success definition. But they agree on the direction. The measurement gap is the reason most agentic AI deployments cannot produce a credible ROI number, and the enterprises that have closed that gap are the only ones reporting returns.
The AT&T Exhibit
The strongest first-party exhibit comes from AT&T. Andy Markus, Chief Data and AI Officer, stated at a recent industry event: “Last year we had five-X cash ROI on our AI investment.” The claim is company-specific, not a cross-vendor synthesis. AT&T invested in visibility infrastructure — the ability to track operating costs across its AI stack — and got returns. The 5x figure is not a benchmark. It is a case study in what happens when a company can actually see the bill.
What the Surveys Actually Say
The IBM Institute for Business Value study, commissioned by Salesforce and surveying 1,222 Salesforce customers, reports that only 33 percent of AI initiatives are meeting ROI targets. Sixty-two percent are concerned about unpredictable costs. Sixty-four percent report unclear total cost of ownership. Seventy-two percent have failed to scale beyond pilot. These are not failure rates for bad technology. They are failure rates for bad measurement.
CrewAI’s survey of 500 C-level executives, vendor-commissioned and fetch-verified, reveals a striking contradiction: 100 percent plan to expand agentic AI deployments, yet time-to-value and ROI ranked last in platform priorities at 2 percent. Only 41 percent report year-one positive ROI, with a median payback period of 6.7 months. The gap between expansion intent and ROI confidence is the visibility problem in one data point.
Gartner’s August 2026 forecast projects a 5x-plus increase in inference costs through 2028, which it calls the “Inference Paradox.” The firm also predicts that 40 percent or more of AI projects will be canceled by the end of 2027. The cancellation rate is not a technology failure. It is the consequence of deploying without a cost baseline.
Salesforce’s own survey of 2,025 decision-makers, vendor-commissioned, reports that unified data infrastructure reduces time-to-ROI from 8.8 months to 7.3 months. The 1.5-month delta is the cost of the measurement gap. Deloitte’s survey of 3,235 leaders, fetch-verified, reports 42 percent strategy preparedness (not adoption — correctly labeled), 21 percent mature governance, and 74 percent hoping for revenue growth while only 20 percent are achieving it. Accenture’s survey of 3,000 C-suite executives reports 82 percent investing in AI but only 23 percent achieving sustained value, down from 32 percent. BCG’s survey of 2,360 executives reports 90 percent of CEOs expect measurable returns while companies are doubling AI spend.
The pattern is consistent across every survey: the gap between investment and return is not a technology problem. It is a measurement problem.
The 30/70 Split
The CIO.com TCO article, search-confirmed but Cloudflare-blocked at primary fetch, reports that AI service bills account for approximately 30 percent of total cost. The remaining 70 percent is the hidden AI stack — infrastructure provisioning, model fine-tuning, data pipeline maintenance, governance tooling, and operational staffing. A separate CIO.com article on cost overruns reports that 85 percent of enterprises misestimate costs by 10 percent or more, with a quarter off by 50 percent or more.
This is the structural finding: the enterprises that report established ROI are the ones that can see the full 100 percent. Everyone else is spending blind.
What This Means for Deployers
The implications are straightforward. First, model evaluation can no longer stop at benchmarks and pricing. It must include a cost-visibility audit: What does the full operating stack cost? Where are the hidden lines? Second, the measurement gap is not a reporting byproduct that will resolve itself as the market matures. It is the primary barrier to proving ROI, and it requires deliberate infrastructure investment to close. Third, the 5x figure from AT&T is not a universal benchmark — it is a specific company’s return on a specific investment in visibility infrastructure. But it demonstrates the structural relationship: visibility enables measurement, measurement enables proof, and proof enables continued investment.
The deployment bill is not getting simpler. As agentic systems move from pilot to production, the cost stack will grow deeper, not shallower. The enterprises that build visibility infrastructure now will be the ones reporting returns in 12 months. The ones that do not will be the ones canceling projects — not because the technology failed, but because they could not prove it was working.
Follow-up links: Post 132033 (The Deployment Bill), Post 131797 (Microsoft zero-numbers).