The average enterprise now runs 12 AI agents. Within two years, that number is expected to reach 20. And right now, half of those agents are not connected to anything.

That is the headline finding from the Salesforce 2026 Connectivity Benchmark Report, an annual survey of 1,050 enterprise IT leaders across nine countries, conducted with Vanson Bourne and Deloitte Digital. The report, released in February 2026, was summarised and contextualised by Belitsoft’s Chief Innovation Officer Dmitry Baraishuk in a widely circulated April 6 piece that captures where enterprise AI adoption actually stands heading into the second quarter of 2026.

The summary verdict from Baraishuk: “In 2025, everyone talked about AI agents. In 2026, they’re actually using them.” He might have added: they are just not using them well yet.

The Siloed Agent Problem

Fifty percent of enterprise AI agents currently operate in isolation — no data sharing, no coordination, no handoffs between systems. This is not a minor inefficiency. It is the central challenge holding back agentic AI ROI at scale.

The logic is straightforward. An AI agent that handles customer support inquiries is useful. An AI agent that handles customer support, is connected to the CRM, can see open orders in the ERP, can check inventory in the supply chain system, and can flag billing anomalies to the finance team is transformative. The first version is a productivity tool. The second version is an autonomous workforce member.

The reason 50% of agents stay siloed is the same reason enterprise software has always fragmented: integration is hard, and the average enterprise app count has now grown to 957 — up from 897 the previous year — with only 27% of those applications actually integrated with each other. You cannot orchestrate agents across systems that are not connected in the first place.

What the Numbers Say

The Connectivity Benchmark Report has been tracking enterprise integration trends for eleven years. The 2026 edition reveals how quickly the AI agent story has moved from aspiration to operational reality — and where the gaps are showing up.

Key findings:

• 83% of organisations report that most or all of their teams have adopted AI agents
• Only 11% of planned agentic use cases from the prior year reached production — the gap between “we’re deploying agents” and “we have agents in production” is enormous
• 86% of IT leaders are concerned that agents will add more complexity than value without proper integration infrastructure
• Agentic AI as a top IT priority grew 31.5% year-over-year, from 13% to 17.1% of global IT decision-makers
• The AI agents market itself is valued at $8 billion in 2025 and projected to grow at a 46.6% compound annual growth rate through 2034

That gap between the 83% adoption rate and the 11% production rate is the story inside the story. Almost every enterprise says it is running agents. One in nine has actually moved them into production in the use cases it planned for. The rest are still in pilot, still integrating, or still running agents as disconnected point solutions that do not compound into something bigger.

Why 12 Agents That Don’t Talk to Each Other Is Worse Than Two That Do

There is a certain point at which adding more disconnected AI tools increases operational complexity without adding capability. That point is somewhere around where most enterprises are now.

Twelve isolated agents means twelve separate contexts to manage, twelve separate data streams that do not share information, and twelve separate points of failure that your team has to monitor. You do not get the benefit of an agent that has seen what every other agent in your operation has seen. You get twelve single-purpose tools that happen to use AI.

The compounding value of agentic AI comes from orchestration. An agent that can see everything — customer history, inventory, finance, support tickets, sales pipeline — and coordinate with other agents to take action across those domains is qualitatively different from twelve agents that each own one slice of the data. The difference is not 12x. It is closer to the difference between a team of people who never talk to each other and a team that operates as a unit.

This is why 86% of IT leaders are worried about agents adding complexity rather than value. They have seen this movie before with traditional software. Every new SaaS tool promised ROI. The integration cost to connect those tools often consumed the ROI.

The Integration-First Approach

The businesses that are getting the most out of agentic AI in 2026 are not the ones that deployed the most agents. They are the ones that built the integration infrastructure first.

That means connecting the systems agents need to read and write before deploying agents into those workflows. It means establishing a coordination layer — some mechanism for agents to hand off context, share state, and escalate to human review when they encounter decisions they should not make autonomously. It means treating agent deployment as a workforce design problem, not a software procurement problem — the difference between AI automation and an actual AI workforce.

The Connectivity Benchmark’s finding that only 27% of enterprise apps are integrated explains the 50% siloed-agent figure almost entirely. You cannot have connected agents without connected systems. Businesses that have not done the data infrastructure work yet will hit the same wall at the agent layer that they have been hitting at the analytics layer for years.

What This Means for Business

If your organisation is somewhere in the 83% that has deployed agents, the question worth asking is not “do we have AI agents” but “what can our agents actually see, and what can they actually do.”

An agent that can only access the data in one system is a sophisticated search tool. An agent that can read across your customer, operational, and financial data — and take action across those domains — is closer to what the case for agentic AI actually promises.

The 67% growth in agent count expected over the next two years (from 12 to 20 per enterprise) will make the integration problem significantly worse if it arrives without a coordination strategy. Twenty siloed agents is not better than twelve siloed agents. It is twelve siloed agents with eight more problems added.

The businesses that treat this moment as a reason to invest in orchestration — connecting their agents, building shared data access, establishing governance for autonomous action — will have a compounding structural advantage over those that keep adding point solutions to an already fragmented stack. If you are not sure which parts of your business are ready for AI agents, that is the right place to start.

---

Enterprise DNA put together a free field guide on exactly this: the full Claude ecosystem, Claude Code, and how to roll agents out without breaking things. Get the guide.