Home  ›  Headless CRM  ›  AI Agents and MCP3 of 6

AI agents, MCP and headless CRM

What happens when AI assistants can do more than answer questions about your CRM?

AI assistants are becoming another way for people to interact with business software.

Instead of opening the CRM to find information or perform a routine task, an employee may be able to make a request through conversational AI or another supported AI application.

The assistant can then use authorised tools to retrieve information or request an action.

This is one of the most interesting developments in headless CRM.

What is Model Context Protocol?

Model Context Protocol, or MCP, is an open standard for connecting compatible AI applications to external tools and information.

An MCP server can expose specific capabilities to an AI application.

Those capabilities might include searching customer records, retrieving opportunity information or performing a supported CRM action.

The assistant can discover the available tools and use them within the permissions and constraints of the connected systems.

MCP doesn’t replace the CRM, its APIs or its security controls.

It provides a standard mechanism for making certain capabilities available to compatible AI applications.

Model Context Protocol
Compatible AI applicationDiscovers the available tools.
MCP serverExposes specific capabilities.
Search customer records
Retrieve opportunity information
Perform a supported CRM action
CRMIts APIs and security controls stay in place.

What does this look like in practice?

Imagine an account manager preparing for a customer meeting.

They ask an AI assistant:

“Give me a summary of this customer’s open opportunities, recent activity and unresolved service issues.”

Through appropriate connections, the assistant retrieves the relevant information from the CRM and potentially other authorised systems.

It organises the findings into a useful briefing.

The employee can review the information without navigating several applications.

Now imagine the account manager asks:

“Create a follow-up task and prepare an update for the opportunity.”

The assistant may be able to perform those actions if the relevant tools and permissions are available.

This is where the distinction between retrieving information and changing business records becomes important.

Reading information isn’t the same as taking action

Looking up a customer’s phone number generally presents a different risk from changing pricing, approving a refund or creating an order.

An organisation needs to define which actions an AI assistant can perform.

Some may be allowed automatically.

Others may require human approval.

Certain actions may be prohibited entirely.

Those boundaries should be enforced through appropriate technical permissions and business controls, not simply described in instructions given to an AI model.

The agent doesn’t own the business rules

Consider an employee asking an assistant to apply a discount to an order.

The assistant might interpret the request and invoke an authorised pricing or approval process.

It shouldn’t independently decide that a discount is permitted merely because the user requested it.

The business’s pricing policy still applies.

The same principle extends to spending limits, financial approvals, customer data and other sensitive actions.

A discount request
EmployeeAsks an assistant to apply a discount to an order.
AI assistantInterprets the request.
Authorised pricing or approval processThe business’s pricing policy still applies.

AI can participate in the process without becoming the authority governing it.

Can an AI assistant work through Slack?

Yes,

where an organisation has implemented the appropriate integrations and permissions.

An employee might ask a question in Slack, receive a customer summary from an AI assistant and initiate a follow-up action from the same conversation.

The assistant doesn’t have to exist solely within a standalone conversational AI interface.

The wider idea is that employees can interact with business information through the tools that suit their work.

Do all CRM processes need an agent?

No.

AI is particularly useful when a process involves interpreting natural language, documents or complex information.

Predictable tasks such as validating a required field, checking a numerical limit or applying a fixed business rule are often better handled by conventional software.

The right solution might combine AI with integrations, workflows and process orchestration.

AI is particularly useful for interpreting
Natural language
Documents
Complex information
Often better handled by conventional software
Validating a required field
Checking a numerical limit
Applying a fixed business rule

The opportunity isn’t giving an AI agent unrestricted access to your CRM. It’s allowing AI to perform useful, authorised work within the rules of your business.

Thinking about what this means for your business?

Talk to Kongo about your requirements, the systems you already operate and what you’re trying to improve.

Talk to Kongo