Only 13% of Enterprises Report Scaled AI ROI While 37% Already Deploy Agents, ETR Survey Finds

Roughly 13% of enterprises report sustained return on investment from AI at scale, whether they build in-house or buy from vendors, according to a new analysis of Enterprise Technology Research survey data published Saturday by SiliconAngle. The figure sits in sharp contrast to the pace of AI agent deployment: 37% of organizations already have agents deployed or in active testing, up 10 percentage points from last year, according to ETR data presented at RSAC 2026.

The numbers sketch what SiliconAngle calls the “agentic gap” — vendors are shipping agentic capabilities faster than enterprises can operationalize them.

The ROI Breakdown

ETR’s quarterly drill-down survey (N = 1,573) splits enterprise AI adoption into two tracks: building in-house and buying vendor solutions. In both cases, the largest single category is “ROI in pilots or limited use cases, but not yet at scale,” at roughly 33% for in-house builds and 39% for vendor purchases, per the SiliconAngle analysis.

On the in-house side, about 30% of respondents report “adoption, ROI not yet realized” — organizations that are actively building and experimenting but not seeing payback. The “no adoption or traction” segment is declining but remains meaningful, especially for in-house efforts.

The constraint is not enthusiasm. The share of organizations saying they don’t leverage AI in any capacity dropped from 10% in July 2025 to 6% now, according to the same ETR dataset. Productivity automation and task augmentation remain the dominant use case at around 70%, consistent across multiple quarters. The fastest-growing category is decision support through AI-driven analytics.

Security Controls Lag Agent Deployment

The governance side of the gap is even starker. Of the 37% of organizations already running AI agents, 20% admit to having no agent-specific security controls at all, according to ETR’s chief strategist Erik Bradley, speaking at RSAC 2026. Only 3% said they have broad controls in place.

When ETR asked respondents to identify the right governance model for agents, every answer option received roughly equal support at 23-24%, Bradley noted in his theCUBE interview. “There’s no clear view,” he said. “We are really way ahead in deployment than we should be from a governance perspective and control perspective.”

LLM and generative AI protection has overtaken cloud security as the top area where enterprises plan to increase spending for the first time, with a 10-point jump, according to ETR’s annual security study. Cloud security, the top category for the past two years, declined by roughly 5 points.

Macro Headwinds Compound the Problem

IT budget sentiment dropped from 4.6% expected growth in January to 3.6% now, according to the ETR survey data (N = 1,543). Geopolitical disruptions, oil prices, inflation risk, and potential Fed tightening are all cited as factors pulling spending expectations down. Overall cybersecurity spending remains resilient — only 5% of respondents plan to decrease security budgets — but the broader IT belt-tightening means enterprises have less room to fund the operational buildout that agent adoption requires.

SiliconAngle frames the core problem as architectural: enterprises are trying to bolt agentic workflows onto software stacks designed for a pre-agent era. The analysis proposes a four-layer model — frontier models, a “cognitive surface” for governance and integration, transactional systems of record, and edge inference — as the architecture that needs to emerge before pilots convert to production at scale, per the full analysis.

For builders selling into enterprises, the data points to a specific bottleneck: the governance and integration layer between the model and the business system. Agent capability is ahead of agent governability. Until the middle layer matures, enterprise sales cycles will continue to stall between pilot and production.