We’re entering the phase where agentic AI stops being a research demo and starts being an operational headache. The question on the table isn’t whether these systems can be built—it’s whether they can be governed once they’re running loose in enterprise environments.
Now, the numbers tell a pretty clear story about where the market is heading. Gartner predicts that 40% of enterprise applications will integrate AI agents by the end of 2026, up from less than 5% in 2025 [1]. That’s a massive jump in a very short window, and it means agentic AI is moving from pilot projects to production infrastructure at a pace that would make most IT security teams wince. And wince they should: 74% of organizations are ramping up their AI investments this year [2], but 55% of IT security leaders lack confidence in their current guardrails for deploying agentic AI solutions [2]. Meanwhile, 79% of organizations are grappling with compliance challenges related to these systems [2].
So what’s going on? Basically, you’ve got a deployment wave running ahead of a governance wave, and the gap between them is where the risk lives.
Lonnie Ross, Digital Experience Marketing Lead at BigID, put the structural shift well: “Agentic AI is changing how organizations think about governance. Unlike traditional AI systems, these agents operate autonomously, persist across sessions, and interact directly with sensitive data and enterprise systems” [3]. That’s the crux of it. Traditional AI governance was built for models that sat in relatively contained environments—think recommendation engines or classification tools that processed inputs and returned outputs. Agentic systems don’t stay in their lane. They initiate actions, maintain state across interactions, and touch sensitive systems in ways that make traditional access controls look quaint.
The visibility problem compounds this. ITECS notes that 68% of employees use AI tools without IT approval, creating what they call a “Shadow AI visibility gap” [1]. Now, that’s a statistic cited without clear attribution to an original survey, so I’d treat it as directional rather than definitive. But the phenomenon it describes—unauthorized AI tooling proliferating faster than central IT can catalog it—tracks with what practitioners report anecdotally. When agents can be spun up by individual employees or line-of-business teams, the traditional enterprise security perimeter essentially dissolves.
But tied up with the technical challenges is a regulatory story that’s just getting started. Satyadhar Joshi, an independent researcher who submitted comments to the Office of Science and Technology Policy, wrote that “the rapid advancement of artificial intelligence (AI), particularly agentic AI systems capable of autonomous decision-making, has exposed significant gaps in existing federal regulatory frameworks” [4]. The OSTP is actively soliciting input on regulatory reform for agentic AI, which tells you something about where the U.S. government thinks this is headed. They’re not regulating a hypothetical future technology—they’re trying to catch up to one that’s already deploying.
Now, staying on governance for a minute, KPMG frames this as a foundational shift in how organizations need to think about AI risk [5]. The move from model-centric to data-centric governance—where the agent’s access to and manipulation of data becomes the primary control surface—is emerging as a consensus framework. BigID’s Ross makes this explicit: “data governance is becoming the foundation for AI governance” [3].
What does this all mean for the space? I tend to think we’re looking at a classic infrastructure-buildout pattern. The first phase is capability demonstration—can we make agents that work? The second phase is integration—can we embed them in workflows? We’re now entering the third phase, which is where the real money gets made and lost: operational reliability and compliance at scale. The companies that solve governance don’t just de-risk their own deployments; they become the picks-and-shovels infrastructure that everyone else has to buy.
Some expect this to create a regulatory moat for incumbents who can afford compliance teams. Then again, there’s also the potential that governance tooling becomes a democratized layer—open standards and automated compliance that levels the playing field. The honest answer is that it’s too early to call.
Sources
- Agentic AI Governance Framework 2026 | Shadow AI Guide | ITECS — ITECS (https://itecsonline.com/post/agentic-ai-governance-2026-guide)
- 4 Best Practices for Robust Agentic AI Governance | TEKsystems — TEKsystems (https://www.teksystems.com/en-hk/insights/article/agentic-ai-governance)
- Agentic AI Governance Trends for 2025 and Beyond | BigID — BigID (https://bigid.com/blog/agentic-ai-governance-trends/)
- OSTP-TECH-2025-0067-0401_attachment_1.pdf — U.S. Government (OSTP) (https://downloads.regulations.gov/OSTP-TECH-2025-0067-0401/attachment_1.pdf)
- AI governance for the agentic AI era — KPMG (https://kpmg.com/us/en/articles/2025/ai-governance-for-the-agentic-ai-era.html)