There are already plenty of articles about MCP. Most start with a complex definition, a technical architecture diagram, or a long history lesson. These can be useful but often leave a basic question unanswered: How can MCP actually help me?

We wrote this guide to address that. We’ll start with examples of how MCP can be used across sales, marketing, support, data, and engineering. Then we’ll explain how MCP works, why it was created, and what you should look out for when using it.

What is MCP?

MCP tackles a basic problem. Although AI models are becoming more powerful every day, their usefulness is limited by the information and systems they can access. An AI agent may understand the question you’re asking and how to find the answer, but still not be able to help you because it can’t access other systems and the data they contain.

Suppose you ask Claude which sales rep is assigned to a certain account in Salesforce. Claude understands the question, but it cannot give you the answer unless it has a way to look inside your Salesforce account. The same problem comes up when you ask an agent to do something. Claude may know how to create a new opportunity in Salesforce, for example, but it cannot make the change unless it has a connection to use Salesforce.

Model Context Protocol (MCP) gives AI agents a standard way to access other systems and take action on them. In practical terms, it allows AI agents to query systems and execute operations on them based on plain language directions.

What Can You Do With MCP?

The easiest way to understand MCP is to look at what you can do with it. Instead of gathering information yourself, switching between another system and your AI agent, and carrying out each step manually, you can simply tell an AI agent to work through the available connected system.

Here are a few examples of what this can look like across different roles:

RevOps: Investigate Pipeline Risk

Pipeline reviews often mean checking opportunities one by one to see which deals have stalled and whether the sales rep has followed up.

With MCP, a RevOps leader could ask:

The AI agent can review the opportunity and activity data available in Salesforce, apply the criteria, update the relevant records, and create the follow-up tasks.

Marketing Operations: Build and Activate an Audience

A targeted campaign requires filtering contacts based on who they are and what they’ve done, then sending the right message to everyone who qualifies.

A product marketer could use MCP to ask:

The AI agent can use the customer, email engagement, and form submission data available in HubSpot to find the right contacts. It can then send the webinar invitation to the relevant people, all from a single request.

Customer Support: Add Business Context to Prioritization

A support queue can tell you which tickets are open. But it may not show when a customer has experienced several serious problems in a short time frame, which can be an early sign that the relationship is at risk. To understand the severity of the pattern, someone has to review the customer’s recent ticket history, compare the priority and status of each issue, and determine which problems remain unresolved.

Customer support could use the following prompt with MCP:

The AI agent can review each company’s ticket history in Zendesk, find the customers with recurring serious issues, and escalate their unresolved tickets. The support team can then step in before these issues cause significant damage to the customer relationship.

Engineering: Investigate Failed Workflows

When several application errors appear at once, an engineer has to decide which ones matter most. That usually means opening each issue, checking how often it occurred, seeing how many users it affected, and reviewing the available error details.

With MCP, an engineer could ask:

The AI agent can review the issue and event data in Sentry, identify the errors that meet the criteria, and organize the relevant details into a clear summary. The engineer still decides what caused the problem and how to fix it, but they no longer have to open and compare every issue manually.

Data and Analytics: Answer Ad Hoc Questions Quickly

Some business questions matter in the moment, but not often enough to justify building a dashboard. Answering them still takes work. Someone has to find the right tables in the data warehouse, write a query, check the results, and explain what they mean.

The operations team could use MCP to find out:

The AI agent can find the relevant data in Snowflake, write and run the query, and summarize what it finds. The data team still owns the models, definitions, and access controls that make the answer reliable. MCP simply gives other teams a more direct way to use what the data team has already built.

How Does MCP Expand What AI Agents Can Do?

On its own, AI agents can only work with publicly available information and what you give it. Its capabilities are limited to what’s built into the application.

So what changes when an AI application has access to MCP?

MCP Gives AI Agents Context and Capability

There are two main reasons an AI application cannot do significant operational work on its own: it doesn’t know what’s happening in your company, and it cannot act in your systems.

MCP can help with both.

Context

Context is the information an AI agent needs to understand a specific company, customer, or task. A general-purpose AI model does not know which opportunities changed stages this morning, which customers opened high-priority support cases, or which campaigns exceeded their targets. That information lives in your company’s systems and changes constantly.

MCP lets AI agents retrieve current company information from a connected system. Its responses can then be based on your company’s actual data instead of general knowledge or information you put in the prompt.

Capability

Knowing what’s happening is only half the problem. The AI agent also needs a way to do something about it.

A general-purpose AI model can explain how to update a CRM record, configure a data sync, or launch a marketing campaign. But explaining the steps is not the same as carrying them out.

MCP allows AI agents to carry out approved actions in connected systems. This is how it moves from telling you how to complete a task to actually doing it for you.

MCP Brings the Controls for Multiple Systems Into One Interface

Most teams use a lot of software. Each tool has its own interface, so getting work done often means switching back and forth between applications.

MCP lets you control many of those tools from a single AI application like ChatGPT or Claude. You can ask the AI agent to complete a task in Salesforce and then handle a different task in HubSpot, all from a familiar interface.

The software still plays a critical role: it stores the data, upholds security, and performs the work. What changes is the interface you use to control it.

MCP Reduces Technical Barriers to Operational Expertise

Accessing some data requires technical knowledge, such as writing SQL or navigating a complicated database schema. A business user may know exactly what they want to learn from a database but still need an analyst to retrieve and format it.

MCP can help close that gap. The user can describe what they want in plain language and the AI agent can use the connected system to get it for them.

This does not eliminate the need for data and operations teams. They still design the data models, define the business logic, maintain data quality, and control access. But it can give non-technical teams a more direct path to the data they need.

How Does MCP Work?

It is clear that MCP can provide many meaningful benefits. But how exactly does it work?

There are a few main parts to MCP:

• MCP Host: the AI agent, such as ChatGPT or Claude, where you make the request
• MCP Server: a program that lets an AI application access information and request actions from a service
• MCP Client: a component within the AI agent that connects to an MCP server and manages the communication
• Connected System: the system that contains the relevant information or performs the necessary actions (ex: CRM, database, support platform, or data tool)

Through an MCP connection, an AI agent can list the information a system provides and the actions it supports. It chooses the relevant capability, sends the request through the MCP client, and shows you the result.

MCP is an open standard, which means its rules are public and not controlled by a single vendor. A company can make a service available through MCP, and many compatible AI applications can use it.

Why Was MCP Created?

MCP was created to solve a growing integration problem between AI agents and external systems.

As AI agents became more powerful, companies wanted to connect them to their own data and tools. At first, this meant building a connection between each AI agent and each external system.

That approach can work when there are a handful of connections. As you add more AI applications and business tools, though, the number of connections quickly grows out of control. Each connection has to be built, secured, and maintained individually.

Anthropic introduced MCP in November 2024 to solve this problem. Instead of requiring every AI application and external service to communicate in a different way, MCP gave them a shared language.

In December 2025, Anthropic contributed MCP to the Linux Foundation’s Agentic AI Foundation. This gave MCP a vendor-neutral home where it could develop under open governance instead of being controlled by a single company.

Adoption has spread quickly across AI platforms and development tools. MCP is now supported by products such as ChatGPT, Claude, Gemini, Microsoft Copilot, Visual Studio Code, and many more.

How Is MCP Different From a Traditional Integration?

Traditional integrations and MCP can both retrieve information and trigger actions. But they organize the connection differently.

A traditional integration is built around a workflow defined in advance. Developers decide what triggers it, what data it accepts, how that data should be mapped, and what happens next. The integration can process different records, but it still follows the same basic path each time.

An MCP server does not define one fixed workflow. Instead, it describes the information and actions a system makes available, along with the inputs each action accepts. When a user makes a request, the AI agent can inspect those capabilities, choose the ones that fit the task, and supply the appropriate inputs.

This makes MCP useful for work that changes from one request to the next. Traditional integrations are still better for stable, repeatable processes that should run the same way every time.

What MCP Does Not Solve

MCP can make AI agents significantly more powerful, but it also creates new vulnerabilities. It’s important to understand the potential pitfalls of MCP before giving AI agents access to sensitive data or consequential actions.

MCP Does Not Compensate for Bad Data or Broken Business Logic

MCP makes it easier for AI agents to access company data, but it does not make that data complete, consistent, or correct. If the data is outdated or poorly maintained, the AI agent’s response will reflect that.

The same idea applies to business logic. MCP cannot decide how your company should define an active customer, calculate qualified pipeline, or identify an at-risk account. Those definitions must already be clearly established.

MCP makes established data and operational capabilities more accessible. The accuracy of the results still depends on the quality of the data and logic behind them.

MCP Does Not Improve the AI’s Reasoning

MCP can give an AI agent access to more relevant and current information, but it doesn’t change the model’s underlying intelligence. Just like in general use cases, the AI agent can misunderstand a request or reach the wrong conclusion.

Of course, the risk increases when an agent can take action. High-impact workflows should always have safeguards such as limited permissions, testing, and human review before an action is carried out.

MCP Struggles to Combine Data Across Systems

MCP works well when an AI agent needs to get information or take action in a single system. It is less efficient when it needs to join data from several systems.

If account data is in Salesforce, support history is in Zendesk, and product usage is in a database, the AI agent has to request data from each system separately. It must wait for each response before it can combine the results. The more systems and records involved, the slower, more complex, and more brittle the task becomes.

This can work for small, focused requests. But for larger joins, it is usually better to index the data into a storage system.

MCP Does Not Distinguish Between What Is Possible and What Is Appropriate

Being able to do something is not the same as having the authority to do it. A user may be able to update a sensitive field, launch a bulk action, or delete a workflow, but understand that the action falls outside their responsibilities or requires approval.

AI agents do not inherently understand these unwritten boundaries. If the connected account is able to take an action, the agent will treat it as an option. The risk is even greater when an MCP connection uses a shared account with broader permissions than those of the individual user.

Organizations need controls beyond basic authentication, such as least-privilege access, separate read and write permissions, and additional approval requirements for consequential actions.

MCP enforces the permissions it is given, but it cannot determine whether those permissions accurately reflect what a user should do.

Supercharge Your MCP Strategy with Polytomic

Thanks to MCP, users can simply tell AI agents their desired outcome and let agents gather information and take action in the connected systems their company uses.

As MCP adoption grows, teams will be able to transform how they investigate problems, coordinate workflows, and execute work across systems.

The Polytomic MCP server gives AI agents access to the systems and data available through a company’s Polytomic environment. Instead of setting up a separate MCP connection for each supported service, teams can use Polytomic as a single access point.

With the Polytomic MCP server, users can:

• Configure and Manage Data Syncs: Create and modify data syncs through natural-language instructions instead of configuring every workflow manually
• Query Databases and Data Warehouses: Generate queries, run them, and view results directly in the AI application’s interface
• Access Connected Business Systems: Work with applications connected to Polytomic across sales, marketing, support, finance, and more

Book a demo today to see how the Polytomic MCP Server can help streamline your organization’s operations.