Enterprise AI adoption plateaus as ROI scrutiny grows

By late 2026, the clearest signal was not runaway AI expansion but stall speed. Marketing School, citing a credit-card-spend-based tracker, said the “RAMP AI index shows AI adoption held flat at 45% of businesses in November driven by slight declines in the finance technology sectors,” with RAMP described as “the credit card company” monitoring payments for ChatGPT, API, Anthropic and Gemini subscriptions; even vendor momentum looked uneven, as “Open AI actually declined by 1%. Enthropic up8% Google up 7%,” suggesting churn and saturation rather than broad-based acceleration.

That plateau mattered because it coincided with a change in what enterprises were optimizing for: not more tools, but measurable business value. On SiliconANGLE theCUBE, Dave Vellante argued that “2026 is the year that enterprises actually do start reporting AI ROI at scale,” adding that “the ETR data and the Cube data… is pointing us in a new direction with signs of productivity increasing,” while noting “the percent of customers reporting that they don't measure AI ROI has now dropped 10 percentage points in the past year,” evidence of a pivot away from tool-first adoption toward operational integration.

The pattern across enterprise cases is that AI pays off when it is threaded through a workflow, not merely switched on as a tool. On LaunchPod, Zuora said AI cut “the four week minimum journey of scripting, recording, editing, publishing” to “four hours” by embedding an AI rough-cut and localization flow into work already being done, then extended value downstream when AI could “search through the video and pinpoint the exact video clip,” shrinking retrieval from “four to five hours to less than five minutes” for multilingual training content. A similar workflow-first pattern showed up when, in “just 6 weeks we were able to roll out cursor AI to create custom demos and custom content a whole lot faster… [which] allowed for iterations in the field,” underscoring that value came from speeding a field-content loop rather than from the model alone.

What makes those gains durable is operating-model design: guardrails, human checks, and clear decision rights that connect outputs to business context. Zuora described a “well established AI governance process… a three stage process” that permits experimentation “within the guardrails of security and compliance,” then asks whether it is “moving the needle,” while at Davos an Ernst and Young risk workflow fell from “46 hour” to “15 minutes” before leaders “came back and added some verification and double checking and it ended up at six hours,” making speed usable. Snowflake similarly emphasizes “making sure that the people” are aligned around how AI is used, reinforcing that adoption depends on governance and human operating models, not infrastructure alone.

The turn toward process-first AI was driven by a pattern of tool-first disappointments that exposed operational weakness, not model weakness. On with Kara Swisher captured the distinction bluntly: companies “failed to realize” the gap between “capability and reliability,” because a chatbot can be “about 80% good” for drafting while “the five nines of reliability… 99.999% reliable” needed in production is far harder; the same discussion pointed to Humane Tech PIN and Rabbit R1 as “pretty capable” assistants that could order food correctly, except when “maybe 10% of the times” they sent it to the wrong address, a “catastrophic failure” in real use.

What changed next was the diagnosis: enterprises increasingly concluded that AI was being dropped into workflows that were never designed to support it. On with Kara Swisher compared the problem to electrification—“after we invented electricity, it took 40 years for us to figure out how to electrify factories,” because “just putting in electricity in the existing factories was not enough”—while Cloud Wars Live said the reason “80% of initiatives are not seeing the value that are anticipated” is not simply the models, but basic questions such as “are you choosing the wrong initiatives?,” using the right data, and whether “the processes” were “defined correctly.”

The clearest sign that the market has moved past experimentation is that companies are now reporting workflow-level metrics, not just model usage. CFO Thought Leader reported Pigment’s “new platform” can fill out “80% of an RFP based on prior questions, prior responses,” with sales-consulting productivity rising sharply, while Investinq said Bank of America’s Erica has handled more than 3 billion customer interactions, producing a 19% revenue lift through cross-selling at a scale framed as the equivalent of 11,000 full-time employees working continuously.

Those gains are also showing up in execution speed, throughput, and cost structure once firms hand real work to agents. Investinq said UWM’s 24/7 agent Mia made 400,000 outbound calls that led to 14,000 closed loans and helped drive its best loan-origination quarter since 2021; XPO cut empty miles 12%, reduced unnecessary rerouting 80%, and saw purchased transportation costs fall 53% year over year after AI enabled 93% insourcing of linehaul; Infinite Runway said C.H. Robinson’s 30-plus agentic tools lifted quote response rates from 65% to 100% and cut response times from 17 minutes to 32 seconds.

Skeptics of the AI ROI story point first to how slippery the evidence remains: Big Technology reported a Wharton study finding that “74% of businesses that measure the ROI from their generative AI efforts are already seeing a positive return,” yet that headline sits on uneven ground because the study is self reported and based on responses from more than 800 people at large U.S. companies. The variation is stark enough to weaken broad claims: only 54% in retail reported positive ROI, versus 88% in tech and telecom and 83% in banking and finance, while 81% of VPs and above saw returns compared with 69% of managers and directors, who were twice as likely to say it was still too early or stuck in pilot mode.

The deeper skeptic case is that productivity claims still outrun proof of durable financial impact: The AI Productivity Plateau found that 91% of teams report higher productivity from AI, but only 12% are extremely confident in the metrics behind that claim, a 79-point gap between perceived and proven results. That gap shows up in expectations and spending too, with 49% of leaders believing effective adoption could produce gains of 51% or more while only 23% report gains at that level today, and even among teams that explicitly report a value plateau, 88% still plan to increase AI spending over the next six months.