An AI agent that cannot look anything up or do anything is just another chat assistant. Its value appears when it can read a customer record in the CRM, check stock in the ERP, find an exchange in the mailbox or create a task in the project management tool. That is also when the serious questions begin: what does the agent have access to, with which rights, and what happens if it gets it wrong?
This article explains how an agent is actually connected to your business tools, which technical options exist today, and the rules we apply to keep that connection under control.
What “connecting an agent” means
A language model does not connect to your CRM by itself. It is given tools: specific functions, described in plain language, that it can ask to call. For example: “search for a customer by name”, “read a customer’s order history”, “create a note on a record”.
When the agent decides to use a tool, it is your code that actually performs the action, with its own credentials and its own checks, then returns the result to the model. The model never touches your systems directly: it makes requests, and a software layer decides whether to execute them. That layer is what makes the difference between a safe agent and a dangerous one.
The technical options
- Tool APIs: most modern CRMs, ERPs and email systems offer an API. A small number of tools tailored to the use case is built on top of it. This is the most common and most controllable approach.
- The MCP protocol: the Model Context Protocol is an open standard, launched by Anthropic in late 2024 and now supported by many tools and model providers. It defines a common way to expose tools and data to an agent. An “MCP server” for your CRM can then be reused by different agents or assistants, instead of rewriting the integration each time.
- Automation platforms: some workflow platforms offer ready-made connectors to hundreds of applications. Handy for moving fast, provided you check exactly which rights each connector gets.
- Direct database access: possible, but to be reserved for tightly controlled cases, read-only and on limited views. An agent that can write freely to a production database is a risk nothing justifies.
MCP simplifies the connection, but it does not handle security for you: an MCP server that exposes too many tools, or with rights that are too broad, raises exactly the same problems as a poorly designed in-house integration. And an MCP server downloaded from the internet must be assessed like any third-party software before being plugged into your data.
Connecting an agent is not opening your tools to it. It is giving it a handful of precise actions, and keeping control of everything else.
Rights kept to the minimum
- A dedicated service account: the agent never uses an employee’s account. It has its own account, identifiable in the logs, that can be suspended at any time without affecting anyone.
- Least privilege: an agent that prepares customer summaries only needs to read certain records. It has no reason to see HR data or change pricing terms.
- Separating read and write: read tools are many and broad, write tools are few and targeted. Creating a note, yes. Deleting a contact or changing an amount, no, or only with human approval.
- Business tools rather than generic ones: “add a note to an opportunity” is far safer than “run a query on the CRM”. The more precise the tool, the easier it is to control.
Technical safeguards
- Parameter validation: every agent request is checked by code before execution (format, allowed values, scope).
- Volume limits: calls per minute, records that can be changed per task, maximum number of steps. A runaway loop must not be able to send a thousand requests to your ERP.
- Full logging: every tool call is recorded, with the request, parameters, result and originating task. This is essential for understanding unexpected behavior and for answering an audit.
- Human confirmation for sensitive or irreversible actions: sending outside the company, changing financial data, deleting.
- Protection against booby-trapped content: an agent that reads emails or external documents may encounter malicious instructions. Its rights must be designed with that in mind.
Test before connecting production
An agent is developed in a test environment: a copy of the CRM or a sandbox provided by the vendor, with synthetic data. There you check that the agent picks the right tools, respects the limits, and handles error cases properly: customer not found, API unavailable, inconsistent data. Moving to production then happens gradually, first read-only, then with writes approved by a person, then with wider autonomy where the system has proven itself.
A concrete example
A services company wants an agent to help its account managers prepare monthly check-ins with clients. The agent has five tools: read a customer record in the CRM, read the latest invoices in the ERP, list open support tickets, search recent email exchanges with the client, and create a preparation note on the CRM record.
All read tools go through a read-only service account, limited to the clients of the account manager making the request. The only write tool creates a note, never anything else, and the note is flagged as generated by the agent. Every call is logged. The result: preparation that used to take a good part of a morning is done in a few minutes of review, without the agent being able to change a contract, send a message or view another portfolio’s accounts.
Our recommendation
Before connecting an agent, write down the exact list of actions it needs, and only those. For each one, ask three questions: read or write? What is the worst case if the agent misuses it? Does it need human approval? This list, more than the choice between API and MCP, decides whether your agent will be a reliable colleague or a permanent risk.