The average enterprise runs 957 applications. Only 27% of them are connected to anything else. That single statistic, from the 2026 MuleSoft Connectivity Benchmark Report, explains why most AI agent projects stall. The bottleneck is not the model. It is not the prompt. It is decades of enterprise plumbing that was never designed for software that acts on its own.
Organizations already operate an average of 12 AI agents, and 83% report that most teams have adopted them. But 50% of those agents run in isolated silos, disconnected from the systems they need to be useful. Gartner projects that over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs and unclear value. The infrastructure gap is the primary reason.
The Fundamental Architecture Mismatch
Enterprise systems were built for humans. Agents are not humans. This sounds obvious, but the implications run deep.
Traditional enterprise software operates in deterministic mode: same input, same output, every time. An ERP processes a purchase order the same way whether it is Monday or Friday. AI agents are non-deterministic. They reason, adapt, and produce different outputs depending on context. As one CIO.com analysis notes, this is not a bug to fix. It is a fundamental architectural incompatibility.
The mismatch shows up in three concrete ways.
Batch vs. Real-Time
Most enterprise systems of record, your SAP, Oracle, or mainframe backends, still operate in batch mode. They process data in scheduled cycles: nightly reconciliation, weekly reporting, monthly close. AI agents expect real-time data access. An agent tasked with adjusting inventory levels needs current stock numbers, not last night’s snapshot. When agents hit batch-mode walls, they either fail silently or operate on stale data, both of which erode trust fast.
Structured Formats vs. Natural Language
AI agents communicate in plain English (or plain German, or any other language). Enterprise systems expect XML, JSON, SOAP, or proprietary API calls. One enterprise architect told CIO.com: “You’d need another agent whose sole work is to translate English into API.” That is not a joke. It is an actual pattern emerging in production deployments, and it adds latency, cost, and another failure point.
Identity and Permissions
Traditional systems assume a human will log in, authenticate once, and operate within a session. Agents need machine-to-machine authentication that preserves identity context across multi-step workflows. When Agent A calls Agent B, which calls a database, the database needs to know the original user’s permissions, not Agent B’s service account. Most legacy IAM systems cannot propagate identity through agent chains.
The Numbers Behind the Gap
The MuleSoft survey of 1,050 IT leaders paints a detailed picture of the infrastructure deficit.
96% of organizations face barriers to using their data for AI. Not some organizations. Not most. Virtually all of them.
27% of APIs remain ungoverned, meaning nobody tracks who calls them, how often, or with what permissions. For AI agents that chain multiple API calls together, one ungoverned endpoint in the sequence means the entire workflow is unauditable.
89% see room for API management improvement, which is a polite way of saying their current API layer cannot support autonomous software.
40% cite outdated IT architecture and data silos as their top blocker. Not budget. Not talent. Architecture.
These numbers align with what developers report from the ground. The Langbase State of AI Agents survey of 3,400+ developers found that 70% experience problems integrating agents with existing systems, 85% lack reusable multi-agent infrastructure, and 75% deal with fragmented tooling.
The Dynatrace Pulse of Agentic AI 2026 report adds another layer: roughly 50% of agentic AI projects remain stuck in proof-of-concept or pilot stage. They work in the demo. They break in production. The difference is almost always infrastructure.
Why the Gap Keeps Growing
The infrastructure deficit is not static. It is widening. Three forces are pushing it apart.
Application Proliferation
The average enterprise now runs 957 applications, up from 897 a year earlier. Every new SaaS tool, every department-specific platform, every shadow IT purchase adds another unintegrated endpoint. The integration rate (27%) has not budged, which means each new app increases the absolute number of disconnected systems.
Agent Sprawl
Agents multiply faster than infrastructure teams can connect them. John Wei, writing for CIO.com, describes how his organization grew from a handful of pilot agents to nearly 2,000 AI agent instances spanning more than 40 agent types. His conclusion: “Agents are not the hard part. Scaling them is.”
IDC projects that deployed AI agents will exceed 1 billion worldwide by 2029, a fourfold increase from 2025. Most enterprises are not ready for the 12 they have now, let alone the hundreds they will run in three years.
The Compliance Clock
For European companies, the EU AI Act adds a hard deadline. By August 2, 2026, high-risk AI systems must demonstrate full data lineage tracking, human-in-the-loop checkpoints, and risk classification. Non-compliance penalties reach up to 35 million EUR or 7% of global turnover. You cannot demonstrate data lineage through a system where 73% of apps are not connected. Compliance requires integration, and integration requires infrastructure that most enterprises do not have.
Closing the Gap: What Actually Works
The solution is not a rip-and-replace of every legacy system. That would take years and cost tens of millions. Enterprises that are making progress use incremental strategies that are 2-3x more cost-efficient than full-stack modernization.
The Model Context Protocol (MCP)
MCP, originally developed by Anthropic and now donated to the Linux Foundation’s Agentic AI Foundation, is emerging as the standard for connecting agents to data. With 97 million monthly SDK downloads and backing from Anthropic, OpenAI, Google, and Microsoft, it provides a universal interface between agents and enterprise tools.
Organizations implementing MCP report 40-60% faster agent deployment times. The MuleSoft survey found that 39% of IT leaders already use or plan to use MCP. The protocol handles the structured-format translation problem: agents speak MCP, and MCP servers speak to your SAP, Salesforce, or custom database.
The challenge remains authentication. Enterprises expect agent connections to flow through existing identity providers with full visibility and policy control. MCP’s auth layer is still maturing.
Agent Fabric and Integration Platforms
MuleSoft’s Agent Fabric takes a different approach: instead of standardizing the protocol, it standardizes the management layer. Agent Scanners automatically detect and register AI agents across platforms (Salesforce Agentforce, Google Cloud Vertex AI, Amazon Bedrock). The Agent Registry and Agent Visualizer provide a consolidated view of every agent, what it connects to, and what permissions it holds.
Andrew Comstock, SVP/GM at MuleSoft, puts it bluntly: “The true success of an Agentic Enterprise isn’t found in the sheer number of agents deployed but the overall effectiveness of those agents.”
The Strangler Fig Pattern
For legacy systems that cannot be replaced overnight, the strangler fig pattern offers a proven migration path. You wrap the legacy system in a modern API layer, route new agent traffic through the wrapper, and gradually replace the underlying system piece by piece. European manufacturers using this approach report 30-50% reduction in migration timelines compared to traditional lift-and-shift projects.
Platform-First Architecture
The organizations furthest ahead treat agentic AI as a platform problem, not a collection of agent deployments. That means shared infrastructure for security, policy enforcement, observability, and lifecycle management. Consistent design patterns across all agents. A governance layer that scales with the number of agents rather than requiring manual review for each one.
As Intuit’s CDO told CIO.com: “Your platform needs to be opened up so the LLM can reason and interact.” That single sentence captures the infrastructure gap better than any statistic.
The Budget Reality
Closing the infrastructure gap is not cheap, but failing to close it is more expensive.
Implementation costs for enterprise AI agent infrastructure range from $220K-$400K for a single department to $700K-$2M+ for enterprise-wide deployments, with $80K-$500K in annual operational costs. IT leaders are already responding: 19% of IT budgets are now allocated to agentic transformation over the next 12 months, and 74% of organizations expect their agentic AI budgets to increase.
The ROI data is encouraging for those who invest. Enterprises that have closed the infrastructure gap report a 3-year ROI of 405% with a 4.7-month payback period. But those numbers only apply to organizations where agents can actually reach the data and systems they need. For the 73% of enterprises where most apps remain unconnected, every AI agent deployment is fighting the infrastructure before it can deliver value.
Alois Reitbauer, Chief Technology Strategist at Dynatrace, summarizes the state of play: “Organizations are not slowing adoption because they question the value of AI, but because scaling autonomous systems safely requires confidence” that the underlying systems will behave as intended.
The confidence will come from the infrastructure. Not the other way around.
Frequently Asked Questions
Why are enterprise systems not ready for AI agents?
Enterprise systems were built for deterministic, human-operated workflows. AI agents are non-deterministic, require real-time data access, communicate in natural language rather than structured formats, and need machine-to-machine authentication. Only 27% of enterprise apps are currently integrated via APIs, and 96% of organizations report barriers to using their data for AI.
What percentage of agentic AI projects will fail?
Gartner projects that over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Separately, Forrester predicts less than 15% of firms will actually turn on the agentic features in their automation suites. About 50% of current agentic AI projects remain stuck in proof-of-concept or pilot stage.
How much does it cost to make enterprise infrastructure AI-agent ready?
Implementation costs range from $220K-$400K for a single department to $700K-$2M+ for enterprise-wide deployments, plus $80K-$500K in annual operational costs. IT leaders are allocating 19% of budgets to agentic transformation. Organizations that close the infrastructure gap report 405% three-year ROI with a 4.7-month payback period.
What is MCP and how does it help with AI agent integration?
The Model Context Protocol (MCP) is an open standard originally developed by Anthropic and now maintained by the Linux Foundation’s Agentic AI Foundation. It provides a universal interface between AI agents and enterprise data sources. With 97 million monthly SDK downloads and support from Anthropic, OpenAI, Google, and Microsoft, organizations implementing MCP report 40-60% faster agent deployment times.
How do AI agents connect to legacy systems?
The most effective approach is the strangler fig pattern: wrap the legacy system in a modern API layer, route agent traffic through the wrapper, and gradually replace the underlying system. This is 2-3x more cost-efficient than full-stack modernization. Other approaches include enterprise integration platforms like MuleSoft Agent Fabric, MCP servers, and dedicated translation agents that convert natural language into structured API calls.