Here is a number that should stop you: 68.
That is the percentage-point distance between how many enterprises have adopted AI agents in some form (79%) and how many are actually running them in production (11%), according to IBM Institute for Business Value research published this quarter. It represents, by one measure, the largest deployment backlog in enterprise technology history.
For comparison, when cloud computing was at a comparable moment in its adoption curve, the pilot-to-production gap was roughly 20 points. The gap for AI agents is more than three times wider. Something structural is going wrong — and it is not the AI models themselves.
The Graveyard of Proof-of-Concepts
Most organizations experimenting with AI agents go through the same arc. A team builds an impressive internal demo: an agent that summarizes customer support tickets, or one that cross-references inventory data with supplier emails to flag shortages before they happen. The demo works. Executives are excited. Budget is approved.
Then the agent hits the real environment.
It encounters data that lives in three different systems, none of which speak to each other. It needs to call an internal API that requires credentials no one thought to provision. It produces an answer that is plausible but wrong, because the underlying data is six weeks stale and no one built a freshness check. By month four, the agent is quietly decommissioned and the team moves on to the next pilot.
This is not a failure of AI. It is a failure of infrastructure — specifically, the middleware layer that sits between an AI agent and the data it needs to act on. And it is costing organizations enormous amounts of time and money. IBM estimates that AI-enabled workloads are expected to rise from 3% of enterprise computing in 2024 to 25% by 2026. If most of those workloads follow the pilot-to-graveyard path, the gap between promise and delivery will become a serious strategic liability.
The Real Bottleneck Is Data, Not Models
Think of the electric grid. When the first electrical generating stations came online in the 1880s, Thomas Edison did not have a problem building generators — the technology worked. What he had was a wiring problem. Every building that wanted electricity needed its own custom connection to the generating station, its own voltage conversion, its own safety circuit. Scaling was brutally expensive because there was no standard interface between the source of power and the things that consumed it. The invention that changed everything was not a better generator. It was standardized alternating current transmission — a protocol that let any device plug into any grid.
AI agents in 2026 face the same problem. The models are extraordinarily capable. The problem is that every agent, accessing every data source, requires a custom integration — a hand-built wire between the AI and the tool it needs to use. A company with ten internal systems and three AI agents needs thirty custom integrations, each one brittle and each one requiring maintenance every time a system is updated.
IBM's research is blunt about this: the organizations that are failing to move agents from pilot to production are not being defeated by the intelligence of the models. They are being defeated by the data behind the models. Ninety percent of enterprise data is unstructured — sitting in emails, PDFs, Slack threads, meeting recordings — and only a tiny fraction of it is accessible to AI agents in a form they can reliably use.
A Protocol for Agent Plumbing
In November 2024, Anthropic released the Model Context Protocol (MCP) — an open standard that functions, essentially, as the alternating current for AI agent integrations. Instead of building a custom connector between an AI agent and each data source or tool it needs to reach, developers build one MCP server per tool. Any MCP-compliant AI client can then connect to any MCP server with no additional custom work.
The adoption numbers tell the story of a genuine inflection point. MCP launched with roughly 2 million monthly SDK downloads. When OpenAI adopted the standard in April 2025, downloads jumped to 22 million. Microsoft integrated it into Copilot Studio in July 2025 (45 million). AWS Bedrock added support in November 2025 (68 million). By early 2026, MCP had crossed 97 million monthly SDK downloads and 5,800-plus community-built server integrations, with every major AI provider — Anthropic, OpenAI, Google, Microsoft, Amazon — committed to the standard.
The math of what this changes is significant. Before MCP, connecting an AI agent to ten business tools required ten custom integrations. Each one had to be maintained separately as APIs changed. With MCP, each tool needs one server, and that server works with every compliant agent. Integration costs, by some industry estimates, drop 60 to 70 percent.
Pinterest's Production Playbook
The clearest proof that this is not theoretical came on April 1, 2026, when Pinterest Engineering published a detailed account of how they built a production-grade MCP ecosystem across their engineering organization.
Pinterest's implementation is worth examining closely because it reveals what "production-ready" actually means in practice. It is not just deploying a few MCP servers. Pinterest built three things in parallel:
Domain-specific MCP servers. Each major engineering domain — logs, bug tracking, deployment pipelines — got its own dedicated MCP server, purpose-built for the kind of queries agents in that domain need to make.
A central registry. A single source of truth for all approved MCP servers, their capabilities, and their permission requirements. Before an agent calls a tool, it consults the registry to confirm the server is approved and the requesting agent has the correct permissions. This is governance built into the protocol layer, not bolted on afterward.
Human-in-the-loop approval. For actions with significant consequences, the system surfaces a confirmation request to a human before the agent proceeds. Autonomy where safe; human oversight where it matters.
The results as of January 2026: 66,000 tool invocations per month across 844 active users, saving approximately 7,000 engineering hours per month. That is not a pilot. That is production.
The Emerging Protocol Stack
MCP handles the vertical integration problem — connecting an agent to its tools. But what happens when multiple agents need to collaborate on a task that spans organizational boundaries?
In April 2025, Google announced the Agent-to-Agent (A2A) protocol, designed to handle horizontal coordination between agents. Where MCP is the wiring inside a building, A2A is the street grid that lets buildings communicate with each other. Google explicitly positioned A2A as complementary to MCP, not a replacement.
By late 2025, both protocols had been placed under open standard governance by the Linux Foundation, with Anthropic, Google, Microsoft, and Salesforce jointly committing to interoperability. The first joint interoperability specification is expected in Q3 2026.
For organizations building AI agent infrastructure today, this means the two protocols are converging into a coherent stack: MCP for tool access, A2A for agent coordination. Companies that build on both standards now will not need to rearchitect when the joint spec lands.
What Separates the 11% from the 89%
The data on successful deployments points to four consistent factors.
Narrow scope, deep integration. Agents that succeed in production are not trying to automate everything. They pick one well-defined workflow — invoice processing, customer support routing, code review summarization — and integrate deeply with the specific data sources that workflow requires. Breadth comes later, through orchestration of specialized agents.
Data readiness before agent readiness. The organizations making progress on deployment are spending as much time on their data infrastructure as on the agents themselves. This means establishing clear ownership of internal data, building pipelines to keep that data current, and designing the access controls that let agents reach it safely.
Governance at the protocol layer. The Pinterest case makes this point forcefully. Governance cannot be an afterthought added to an already-deployed agent. It needs to be built into the middleware — into the registry, the permission model, the approval workflows.
Redesigned workflows, not layered automation. Deloitte's 2026 research on agentic AI strategy found that high-performing organizations are three times more likely to scale agents than their peers, and the differentiator is not model sophistication. It is the willingness to redesign the underlying workflow rather than simply placing an agent on top of a process that was designed for humans.
Implications for SMBs
The good news for smaller organizations is that the infrastructure layer is maturing fast. MCP, in particular, changes the economics substantially. An SMB that commits to the standard today can connect its AI agents to its CRM, its support ticketing system, its internal documentation, and its accounting software using community-built MCP servers — many of which already exist in the 5,800-plus server ecosystem — without building custom integrations from scratch.
The critical question is not which AI model to buy. It is which workflows to target first and whether the data behind those workflows is clean, current, and accessible. A business that has its customer data scattered across three CRMs and a decade of PDFs will not get value from an AI agent regardless of how capable the underlying model is. The preparation work — data consolidation, access control design, workflow mapping — is less exciting than deploying a chatbot, but it is what determines whether the investment pays off.
Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. The organizations on the right side of that shift will not be the ones that deployed the most agents in 2025. They will be the ones that built the right infrastructure — the wiring — before they turned on the power.
Conclusion
The 68-point gap between AI agent adoption and AI agent production is not a story about technology failing. It is a story about infrastructure arriving later than the technology it supports — a pattern that has played out in every major platform transition, from electrification to cloud computing.
The infrastructure is now arriving. MCP and A2A represent genuine standardization moments, the kind that historically compress pilot-to-production timelines from years to months. Pinterest's published case study shows what the production-ready pattern looks like. The IBM and Deloitte data shows what distinguishes organizations that close the gap from those that do not.
For decision-makers at SMBs, the practical question is: do you want to be part of the 11% or the 89%? The answer depends far less on which AI tools you choose and far more on whether you are willing to treat your data infrastructure as a first-class investment before you treat your agent deployments as a finished product.
The wiring comes first.