MCP: The Enterprise AI USB-C Port
There's a lot of talk about AI agents capable of operating within enterprise tools. But how they connect is rarely explained. The answer lies in three letters: MCP, for Model Context Protocol. This standard, launched in late 2024 by Anthropic and subsequently adopted by the entire ecosystem, plays the same role for AI as the USB-C port does for our devices: a universal connector that replaces the multitude of proprietary cables.
The problem that MCP solves
Before MCP, connecting AI to a business tool was a custom integration project. Each vendor had to develop its own connector for each model, and each company had to maintain these integrations with each update. The result: assistants that were brilliant at conversation, but blind and helpless when it came to accessing the CRM, reading a shared folder, or updating a document.
MCP reverses the logic. The tool's publisher (or the company itself) exposes an "MCP server" that describes what the AI can read and do: list tickets, search for a contact, create a task, publish content. Any compatible assistant can then connect to it, with the same authentication and permission rules. One connector, all models.
Three concrete examples
The salesperson and their CRM. Connected to the CRM via MCP, the assistant prepares a client meeting in seconds: reviewing past interactions, current opportunities, recent invoices, and company news through a web search. After the meeting, they write the summary and update the client's record. What used to take thirty minutes navigating between four tools becomes a single conversation.
Communication and the website. An MCP server connected to WordPress allows you to ask AI to correct a typo on a page, publish an article, or adjust a meta description, without ever opening the administration interface. The benefit isn't just time saved: it's the elimination of the delay between the decision and its execution.
The executive's assistant. Calendar, email, meeting notes, document database: when these resources are accessible via MCP, the assistant can prepare a summary of the day each morning, identify scheduling conflicts, remind everyone of commitments, and find the right document at the right time. This is the invisible infrastructure behind the professional personal assistants we discussed in our series on Claude. Cowork.
What this changes for AI strategy
For a decision-maker, MCP transforms three questions.
The question of tool choice. The criterion "Does this tool have an MCP server?" is becoming as crucial as "Does this tool have an API?" was fifteen years ago. Business software without an MCP connector will be software your agents cannot use.
The question of application heritage. Companies that have invested in clean, well-structured data gain a decisive advantage: MCPs only transport data; they don't clean it. A poorly maintained CRM connected to AI confidently produces inaccurate summaries.
The question of the scope of action. MCP allows for precise control over what AI can do: read-only access to accounting data, write access to drafts but not to published documents, and no access to HR data. This granularity presents a governance opportunity, provided it is considered from the outset rather than granting all access by default.
Les points de vigilance
Universality has a downside: the more access your agents have, the larger the risk surface. Three precautions are essential. First, apply the principle of least privilege: each connector should only expose what is necessary for the use case. Second, keep a human in the loop for irreversible actions: sending messages, publishing, and deleting. Finally, track agent actions just as you track user actions, because an agent acting within your tools is a user like any other from a security perspective.
Where to start
The simplest approach is to start with a low-risk, high-value use case: preparing meetings from the calendar and notes, or searching for documents in an internal database. Major assistants on the market already offer directories of ready-to-use connectors for common tools. For internal tools, developing an MCP server is a project that takes a few days for a technical team, not several months.
The first week, without a project team
For an individual user or a small team, the quickest route is through ready-made connectors. Open your assistant's connector directory, activate the one for your calendar or document space, grant the requested permissions, and then ask a question that actually requires the data: "Prepare my three appointments for tomorrow using the calendar and recent exchanges." The difference compared to a disconnected assistant is noticeable from the very first conversation.
Three rules ensure a safe start. Begin in read-only mode: for the first week, the AI consults the data but doesn't modify anything. Activate only the connectors relevant to the intended use case, adding more later if necessary, because every unused access point represents a risk without benefit. And initially train on a non-sensitive dataset, giving you time to understand how the assistant actually interacts with the access granted to it.
For the next step, here are four questions to ask your business software vendors at your next renewal: Do you offer an official MCP server? What permissions does it expose, and with what level of granularity? Does authentication rely on our existing accounts? Are AI actions logged like those of a user? The answers will quickly reveal which of your vendors will be ready for your agents.
The conclusion is self-evident: the value of AI agents will not come from their intelligence alone, but from what they are connected to. MCP is the socket. It remains to be seen, with discernment, what you plug into it.
To go further: delve deeper with RAG at the heart of AI systems et 2026, from promise to actual architecture.