Enterprises already run 12 AI agents on average, and half of them cannot talk to each other. That is the headline from Salesforce’s 2026 Connectivity Benchmark Report, based on interviews with 1,050 IT leaders across nine countries. The report makes one thing clear: the bottleneck in enterprise AI is no longer building agents. It is connecting them.
Multi-agent adoption is projected to grow 67% by 2027. But 50% of today’s agents operate in isolated silos, 27% of enterprise APIs remain ungoverned, and only 54% of organizations have any centralized governance framework for their AI agents. The Salesforce data paints a picture of an enterprise landscape that adopted agents faster than it built the infrastructure to coordinate them.
The Numbers: 12 Agents, 957 Apps, and a Coordination Problem
The 2026 Connectivity Benchmark is Salesforce’s 11th annual report, produced in collaboration with Vanson Bourne and Deloitte Digital. The survey ran between October and November 2025 across the US, UK, France, Netherlands, Germany, Australia, Hong Kong, Japan, and Singapore.
The core findings:
- 83% of organizations report that most or all teams and functions have adopted AI agents
- The average enterprise runs 12 AI agents, projected to grow to roughly 20 by 2027
- 50% of those agents operate in silos, disconnected from each other and from core business systems
- 86% of IT leaders worry that agents will introduce more complexity than value without proper integration
- The average enterprise runs 957 applications, and only 27% are connected via APIs
That last number is the one that explains everything else. You cannot orchestrate 12 agents across 957 applications when three-quarters of those applications have no API integration. The agents work. The plumbing does not.
One regulated medical-device company ended up with over 900 agents built in silos within a single year. That is not an anomaly. It is what happens when agent creation outpaces governance.
Protocol Adoption Is Real But Uneven
The report includes some of the first large-scale data on enterprise adoption of agent communication protocols:
- 40% of respondents are using Agent-to-Agent (A2A) protocols
- 39% are using the Model Context Protocol (MCP)
- 68% find it challenging to stay current with emerging agent standards
Those A2A and MCP numbers are higher than most industry observers expected. But 68% struggling to keep up with standards suggests that adoption is often reactive rather than strategic. Organizations are deploying protocols because their vendors push them, not because they have a coherent interoperability plan.
The Silo Problem Is Not a Tech Problem
The easiest reading of “50% of agents operate in silos” is that enterprises need better integration middleware. That is partly true. But the SalesforceDevops.net analysis of the report goes further: the hardest problem is not connection. It is context.
Agents can call APIs. They can send messages to other agents. What they cannot do reliably is understand what another agent means. When a customer service agent escalates a ticket to a billing agent, the billing agent receives a data payload, not business context. It knows the ticket ID and the customer ID. It does not know that this customer has called three times this week, that they are on a legacy pricing plan, or that the service agent already offered a partial credit.
This is the semantic context gap. Enterprises lack what Salesforce calls a “semantic layer” that harmonizes meaning, resolves identity, and enforces permissions before agents act. Without it, multi-agent systems degrade into glorified API chains where each agent starts from scratch.
Why Governance Is Half-Built
Only 54% of organizations have a centralized governance framework for AI agents. The other 46% are governing agents through a mix of team-level policies, informal agreements, or nothing at all.
The governance gap has concrete consequences:
- 27% of APIs are ungoverned: Nobody tracks who calls them, how often, or with what permissions. When an agent chains five API calls, one ungoverned endpoint makes the entire workflow unauditable.
- Shadow AI risk: Without centralized discovery, teams spin up agents that duplicate existing ones or access data they should not touch. The report notes that disconnected agents create “redundant automations and the potential risk of shadow AI.”
- 96% of IT leaders agree that AI agent success depends on seamless integration across all systems. Only a fraction have achieved it.
Salesforce’s Answer: MuleSoft Agent Fabric
Salesforce is not publishing this data out of academic curiosity. The Connectivity Report is, among other things, a case for MuleSoft Agent Fabric, their platform for discovering, connecting, and governing agents across vendors.
Agent Fabric’s January 2026 release introduced Agent Scanners, now generally available, that automatically discover agents across Salesforce Agentforce, Amazon Bedrock, Google Vertex AI, and Microsoft Copilot Studio. The pitch: you cannot govern agents you do not know about, so start by finding them.
The platform supports both MCP and A2A protocols, positioning MuleSoft as the integration layer that sits between agents regardless of which vendor built them. MuleSoft’s blog describes this as “an operating system for the agentic enterprise.”
Whether you buy the product pitch or not, the architectural argument is sound. Multi-agent systems need a coordination layer that handles:
- Discovery: Which agents exist, what they do, and who owns them
- Identity propagation: When Agent A calls Agent B, the downstream system needs to know the original user’s permissions
- Semantic context: Shared business meaning that travels with the data, not just the data itself
- Audit trails: Full lineage of which agent did what, when, and why
These are real requirements regardless of whether you solve them with MuleSoft, a competitor, or a custom build.
What This Means for Enterprise AI Strategy
The Connectivity Report’s data points to three practical takeaways for teams planning their 2026-2027 agent strategy.
1. Audit Before You Build
The report found that organizations do not know how many agents they have. Before deploying agent number 13, run a discovery scan. Map every agent, its data sources, its API dependencies, and its governance status. The medical-device company with 900 ungoverned agents is the cautionary tale.
2. API-First Is Non-Negotiable
94% of IT leaders in the survey agree that AI agents will require IT architecture to become more API-driven. With 73% of enterprise apps still unconnected, this is a multi-year infrastructure investment, not a quick fix. Prioritize APIs for the systems your highest-value agents need first.
3. Context Beats Connection
Getting agents to exchange data is table stakes. The harder problem, and the one that determines whether your multi-agent system actually works, is sharing business context. A customer context layer that travels with every agent interaction will matter more than adding another integration connector.
Frequently Asked Questions
How many AI agents does the average enterprise run in 2026?
According to Salesforce’s 2026 Connectivity Benchmark Report, the average enterprise runs 12 AI agents. This number is projected to grow 67% by 2027, reaching approximately 20 agents per enterprise. 83% of organizations report that most or all teams have adopted AI agents.
What percentage of AI agents operate in silos?
The Salesforce report found that 50% of enterprise AI agents operate in isolated silos, disconnected from other agents and core business systems. This creates redundant automations and shadow AI risks. Only 54% of organizations have a centralized governance framework for their AI agents.
What is the adoption rate of A2A and MCP protocols in enterprises?
40% of surveyed IT leaders are using Agent-to-Agent (A2A) protocols and 39% are using the Model Context Protocol (MCP). However, 68% report difficulty keeping up with emerging agent communication standards, suggesting adoption is often vendor-driven rather than strategic.
What is MuleSoft Agent Fabric?
MuleSoft Agent Fabric is Salesforce’s platform for discovering, connecting, and governing AI agents across multiple vendors. Its Agent Scanners automatically find agents across Salesforce Agentforce, Amazon Bedrock, Google Vertex AI, and Microsoft Copilot Studio. It supports both MCP and A2A protocols and positions itself as a coordination layer for multi-vendor agent environments.
Why do enterprise AI agents fail to coordinate effectively?
The primary barrier is not technical connectivity but semantic context. Agents can exchange data via APIs, but they lack a shared understanding of business meaning. Without a semantic layer that harmonizes meaning across systems, multi-agent workflows degrade into disconnected API chains where each agent lacks the context to make intelligent decisions.