There is a piece of AI infrastructure that most marketing leaders have never heard of, and it is quietly becoming the standard that determines whether your favorite tools can actually talk to AI assistants. It is called the Model Context Protocol, or MCP, and its adoption numbers this year are staggering: roughly 97 million monthly SDK downloads, up from about 100,000 in its first month, a 970-fold increase in under two years.
What MCP actually is, without the jargon
Anthropic released MCP as an open standard in late 2024, meant to solve a specific problem: every AI assistant that wanted to connect to an external tool (your CRM, your analytics platform, your ad accounts) used to require custom, one-off integration work. MCP is closer to a universal adapter, letting any MCP-compatible AI application plug into any MCP-compatible data source or tool without bespoke engineering for each connection.
OpenAI, Google, Microsoft, and Salesforce have all shipped support for it. A large and growing share of Fortune 500 companies have deployed it in some form.
Why this matters for marketing specifically
Your existing tools are becoming AI-accessible faster than you might expect. Pre-built MCP connectors are emerging for Google Ads, Meta, LinkedIn, Salesforce, and HubSpot, meaning an AI assistant can soon query and act on live data from those platforms without your team building custom integrations to make it possible.
“Read-only first” is the right adoption pattern. The safest way to start is with an AI assistant that can read your analytics, search your knowledge base, and draft content, before you grant it any write access, like launching a campaign or updating a CRM field. Require human approval on anything customer-facing or budget-related, at least until you have a track record to build trust from.
It changes what “integrated marketing stack” means. The old integration model required your martech vendors to build direct API connections to each other. MCP means an AI layer can sit on top of tools that were never designed to talk to each other directly, as long as each one exposes an MCP server.
Why “universal adapter” is the part worth understanding, even non-technically
The analogy is worth sitting with a bit longer, because it explains why this adoption curve looks the way it does. Before a common protocol existed, connecting an AI assistant to, say, your CRM meant someone building a bespoke integration specific to that one AI tool and that one CRM — engineering effort that had to be repeated for every new tool-to-data-source pairing. That’s the same problem email had before SMTP, or the web had before HTTP: without a shared protocol, every connection is a custom project. MCP being an open standard rather than one company’s proprietary connector is exactly why adoption spread across competing AI vendors (OpenAI, Google, Microsoft all shipping support for a protocol Anthropic originated) — none of them wanted to be the AI assistant that couldn’t talk to the tools everyone already used, and none of them wanted to build and maintain hundreds of one-off integrations either.
What “read-only first” looks like in an actual rollout
The read-only-first guidance is right, but it’s worth being concrete about what a staged rollout actually looks like, because “start read-only” can otherwise mean anything. A reasonable sequence: first, connect the AI assistant to your analytics and reporting data only — it can query and summarize, but touch nothing. Once you have weeks of that running without incident and a clear sense of how it behaves, extend it to drafting actions that land in a queue for human approval — a campaign brief, a CRM field update suggestion — but nothing executes without a person clicking approve. Only after that stage has built a track record does any direct write access make sense, and even then, scoped narrowly (a specific field, a specific campaign type) rather than broad account access. Skipping stages because the read-only phase “seemed fine” is the most common way these rollouts go wrong — the read-only phase tells you whether the assistant understands your data, not whether it should be trusted to act on it.
What to actually do with this
Ask your core martech vendors, CRM, email platform, ad platforms, whether they have an MCP server or one on their roadmap. The vendors that do are positioning themselves to work well inside the agentic workflows your team is likely to build over the next year. The vendors that do not may become the integration bottleneck in your stack sooner than you expect.