Est.

AI Adoption Plateaus and How Enterprises Restart Momentum

Most enterprises have AI tools but lack the governance and infrastructure to scale them safely.

· 9 min read · Updated
Cover illustration for “AI Adoption Plateaus and How Enterprises Restart Momentum”
AI Change Management · September 30, 2026 · 9 min read · 2,002 words

The enterprise AI plateau has nothing to do with a shortage of tools or budget. It comes from a gap between having AI somewhere in the building and having the foundation to actually scale it. Nearly nine in ten organizations now use AI in at least one business function, up from just over half in 2024. That number sounds like a finished story, but it isn't one. It covers everything from one team running a chatbot to company-wide transformation, and most organizations sitting under that figure are closer to the chatbot end.

The depth numbers tell the real story. Only about one in eight CEOs report both a revenue gain and a cost reduction from AI, and only a small slice of organizations say AI is core to how the business runs. McKinsey research found nearly two-thirds of enterprises have experimented with AI agents, but fewer than one in ten have scaled them to deliver real value, and that ratio holds even inside a single business function, not just across the company. Gartner's June 2025 forecast projects that more than four in ten agentic AI projects will be cancelled by 2027, caused not by the technology failing to work but by unclear return on investment paired with weak risk controls.

That question is settled now. The question that matters today is whether an organization has redesigned its workflows around AI and can point to repeatable, provable value. Most cannot answer yes to that. Adoption was the easy part. Depth is where almost everyone is stuck.

Three structural failures that create the plateau

Diagram: The Scaling Gap: AI Adoption vs. Actual Value. Visualizes: Show the stark contrast between AI adoption breadth and depth using four concrete statistics from the article.

The plateau comes from three structural causes: a strategy gap, a governance vacuum, and an infrastructure deficit. Companies that treat the plateau as a technology problem, buying another tool or running another pilot, end up applying fixes that don't touch any of the three.

Start with the strategy gap. Broad usage is cheap to produce. License counts climb, dashboards light up green, and leadership can report strong adoption numbers in a board deck. Turning that surface-level usage into scaled, repeatable value takes workflow redesign, clear ownership of outcomes, and governance built in from the start. An organization can hand every employee a seat at an AI tool and still have nothing that scales, because access was never the hard part.

Then there's the governance vacuum. Cisco's 2025 AI Readiness Index found that a large majority of organizations plan to deploy autonomous agents, but only one in three say their infrastructure is actually ready for it. Cisco's finding shows intent running far ahead of capability. When governance isn't built before agents start acting on real systems, every attempt to expand runs into the same obstacle: security steps in to slow things down, compliance asks for audit trails that were never built, and IT has no central view of what agents are doing or which systems they can reach. The outcome is a familiar, wasteful pattern: one team's pilot gets rebuilt from scratch by the next team, because there's no shared foundation anyone can reuse.

Last is the infrastructure deficit, and it's the one most often mistaken for a modeling problem. Most organizations are stuck in the experiment or pilot phase because the connective tissue around the AI models, identity, access, credential handling, audit, was never built. These three failures don't appear on a product roadmap; they appear the moment a pilot tries to become something permanent.

Why agentic AI makes the infrastructure deficit dangerous

AI that answers questions is a convenience problem when it goes wrong. AI that acts, calling tools, moving data, touching production systems, becomes a security and compliance liability the moment it's deployed without governed infrastructure underneath it.

The piece of infrastructure doing most of that connecting today is the Model Context Protocol, known as MCP, which Anthropic open-sourced in 2024. MCP has become the standard way enterprises link AI agents to databases, APIs, file systems, and cloud services, and by early 2026 its SDKs were seeing tens of millions of downloads a month, with a large share of technical leaders reporting it's already running in production. That's a fast, real adoption curve for a protocol most executives have never heard of. It's also a protocol whose security posture has not kept pace with how widely it's being used.

Researchers in early 2026 found MCP servers sitting on the public internet with no authentication at all. Those servers were technically compliant with the spec as it existed at the time, and completely exposed in practice. Multiple CVEs were disclosed against MCP implementations across several language SDKs during 2026, and researchers estimate hundreds of thousands of vulnerable servers are exposed globally. One specific attack type, called tool poisoning, lets a malicious MCP server manipulate what an agent actually does. Benchmarked against production systems, the MCPTox benchmark published on arXiv found tool poisoning succeeded more than sixty percent of the time across major LLM agents, and against some models it succeeded on nearly three-quarters of attempts.

The clearest real-world case came from Wiz Research, which disclosed two flaws in the Amazon Q Developer extension, affecting VS Code, JetBrains, Eclipse, and Visual Studio, publicly. A malicious repository could use a crafted workspace MCP configuration file to achieve arbitrary code execution and steal cloud credentials, handing an attacker direct access to cloud infrastructure through what looked like an ordinary developer workflow. That incident makes the supply chain risk concrete. A compromised MCP server doesn't just break a build somewhere, it can exfiltrate production data, because by design these servers have direct access to the enterprise systems they're connected to. And when an MCP server reads customer records and sends them somewhere outside the company, most current deployments log none of it: no record of which agent called which tool, what data came back, or whether any of it matched what the agent was supposed to be doing. Gartner Sr. Director Analyst Aaron Lord has projected that the share of enterprise GenAI applications suffering at least one major security incident a year will rise substantially by 2029, and ties that rise directly to the attack surface MCP has opened up.

Diagram: Tool Poisoning: How Often Attacks Succeed. Visualizes: Visualize the attack success rate of tool poisoning against AI agents, as measured by the MCPTox benchmark published on arXiv.

The governance exposure most security teams haven't mapped yet

Security teams at most enterprises don't have the visibility to know whether attackers are actively exploiting MCP vulnerabilities, and don't have the controls in place to respond if they did.

The 2026 CISO AI Risk Report, surveying 235 security leaders at large enterprises, found that a large majority lack full visibility into their own AI identities, and most don't enforce access policies for those identities at all. At most large enterprises, AI systems already have access to core business platforms, ERP, CRM, financial systems, but only a small fraction of those organizations govern that access effectively. That gap is current state at most large companies right now.

Shadow MCP usage makes the exposure worse. Employees who find useful MCP servers on their own install and connect them without going through any approval process, which leaves IT with integrations it can't inventory and security with activity it can't audit. Every major compliance framework, SOC 2, HIPAA, GDPR, requires audit trails for system access, and most MCP deployments today have none. That's a compliance gap that stays invisible right up until a regulator or an auditor asks for the records, and at that point there's nothing to produce. The EU AI Act's enforcement deadline for high-risk AI systems, set for December 2027, arrived before any agent-specific guidance did, so regulators applying the law's existing provisions to agentic deployments will be making interpretive calls, and those calls may not land in the reviewed organization's favor.

Credential management is the structural piece still unsolved. MCP authorization was built around a person sitting in a browser approving access, but a growing share of the callers making requests today are agents running as cloud workloads, acting on behalf of users who aren't present, or handing off narrower authority to subagents underneath them. Practitioners handling this correctly give each agent specialization its own service account. A customer support agent and a financial reporting agent should never share an API key or a database credential, because separating them is what makes attribution possible, audit investigation workable, and the blast radius of any single compromise limited. Skip that separation, and one stolen credential gives an attacker lateral movement across an entire fleet of agents spanning multiple business systems. The industries furthest behind on turning pilots into production, government, public sector, healthcare, are also the ones carrying the heaviest regulatory load. The gap and the exposure are compounding in exactly the places that can least afford it.

What the new spec resolves

The organization that maintains MCP has moved to fix the authentication gap at the protocol level. A better spec doesn't rewrite what's already deployed, and it doesn't close the distance between what's technically required now and what most organizations have actually built.

The spec change, dated July 28, 2026, brings MCP authorization in line with OAuth 2.1 and OpenID Connect. MCP servers are now formally defined as OAuth 2.1 resource servers and must implement OAuth 2.0 Protected Resource Metadata, so a client can automatically discover which authorization server it's supposed to talk to. MCP clients are also required to implement Resource Indicators, defined in RFC 8707, to specify exactly which MCP server a given token is meant for. That closes a real and meaningful hole: a malicious server grabbing a token that was actually issued for a different, legitimate server.

None of that touches the installed base already running. Because authentication under the older MCP posture was optional rather than required, a large number of servers already deployed remain unauthenticated and were fully compliant under the rules that applied when they were built. Updating a spec doesn't retroactively patch a server sitting in production right now.

The harder problem the new spec leaves open is delegation between agents. The authorization model still assumes a human in a browser clicking approve. Agentic workflows increasingly don't have that human in the loop at all: agents act on their own, hand work off to subagents, or run as cloud workloads with no session for anyone to attach to. There's no standard way yet for an agent to pass a narrower slice of its authority to a subagent while keeping a complete, auditable record of which agent did what, on whose behalf, and under which policy. That's the gap the organizations ahead of the curve are having to solve on their own, because the spec hasn't caught up to it yet.

What governance-first deployments look like in practice

The enterprises that have broken through the plateau didn't do it by finding a better model or a smarter prompt. They made a structural decision early on to treat governance, identity, and access as the foundation adoption gets built on.

Uber is a clear example. It reached scale by building three governance layers before it let adoption expand any further. The first is an LLM gateway that handles PII redaction, access control, and audit logging across every model interaction in the company. The second is an MCP gateway and registry that governs every connection between an agent and a tool across Uber's large catalog of internal services. The third is an agent identity system that extends Uber's existing Zero Trust infrastructure to cover multi-agent workflows, with cryptographically attested lineage recorded on every tool call an agent makes.

ServiceNow took a parallel path, making AI governance the centerpiece of its Knowledge 2026 event. The specific shape of that announcement differs from Uber's, but the underlying decision is the same one: governance comes before scale, not after it.

What connects both examples is sequencing. Neither company bolted security onto a finished agent deployment after the fact. That ordering is the actual difference between an organization stuck reporting 88% adoption with nothing to show for it, and one that can point to measurable, repeatable value from AI across the business. The tools were never the bottleneck. The foundation underneath them was.

Sources

  1. AI Adoption in Enterprise Statistics and Trends 2026 - Second Talent
  2. AI Adoption Statistics 2026: Business & Enterprise Data
  3. AI Agent Adoption Statistics 2026: Enterprise AI Usage
  4. 70 Enterprise AI Statistics for 2026: Adoption, ROI & Trends

More in AI Change Management