# How to connect AI to your ERP and CRM: API, MCP

> Connectors, APIs, MCP: how to connect AI to your ERP and CRM, with which access rights, and in what order to proceed with your IT department.

URL: https://aiorigin.io/en/ressources/insights/integrer-agent-ia-erp-crm-pme-2026

- Insights
- Method

# Connecting AI to your ERP and CRM: connectors, APIs, MCP

Without a connection to your software, AI works blind and your teams copy and paste. There are three ways to connect it to the ERP and CRM: connectors, APIs and MCP. The choice is made with the IT department, one use case at a time.

Updated on 16 October 2026 · Read · 5 min

Contents

01Why connect AI to the ERP and CRM?
02Three ways to connect AI to your software
03What is MCP, in practice?
04How to go about it: the steps with the IT department
05Watch points: rights, writes and traceability
06How AI Origin supports you

In short
To connect AI to an ERP or CRM, you use an off-the-shelf connector, a custom build on the software’s API, or an MCP server, an open protocol that exposes defined actions to several AI tools. In every case: AI acts with the user’s rights, starts read-only, and every write is approved by a person and logged.

## Why connect AI to the ERP and CRM?

Without a connection, AI works blind. Your teams copy extracts from the ERP or CRM into an assistant, paste the answer somewhere else, and nothing is logged. It is slow, and it is a source of data leaks.

Once connected, AI can read the relevant information when it is needed and, where authorized, prepare changes that the user approves:

- “Which of this customer’s orders are late, and why?”
- “Prepare the summary of my meeting and update the opportunity.”
- “Which invoices do not match the purchase order?”

This is also the point that most often blocks the move from pilot to real use: according to BCG, 72% of companies cite integrating AI with their systems as a major challenge. So the question is not whether to connect, but how to do it properly, with the IT department. On scaling up, see our guide from pilots to production.

## Three ways to connect AI to your software

There are three main routes, often combined within the same company. They cover operational data (customers, orders, invoices); for documents (contracts, procedures, meeting notes), you would rather use a knowledge base queried by AI, such as our Company Brain.

RoutePrincipleWhen to use it

Off-the-shelf connectorThe AI tool vendor or the software vendor provides a standard connectionWidely used commercial software, simple read needs

APIA custom build connects AI to the software’s programming interfacesSpecific needs, software with a documented API

MCP serverThe software is exposed to AI through an open protocol, with defined actionsSeveral AI tools to connect, fine-grained access control

### Off-the-shelf connectors

Microsoft Copilot, Claude, ChatGPT and Dust offer connectors to widely used software. This is the fastest route. Its limits: the scope is whatever the vendor has planned, and fine-grained rights management is not always possible.

### APIs

Most commercial ERPs and CRMs expose an API. A custom build lets you choose exactly what AI reads and what it can prepare. It takes longer, but it is more precise, and it is often the only route for specialized software.

### MCP

The Model Context Protocol is an open standard for connecting AI tools to software and data. A single MCP server can serve several assistants, so you do not have to rebuild the connection every time you change tools.

## What is MCP, in practice?

MCP works like a standard plug between your software and AI:

- an MCP server sits in front of your ERP or CRM; it describes what AI can do: search for a customer, read an order, create a draft meeting summary;
- the AI tool (Claude, Copilot or another compatible tool) connects to this server and uses these actions when the user’s request calls for them;
- the IT department decides which actions exist, for whom, and keeps a record of every call.

### An example

A salesperson asks: “Prepare my meeting with this customer tomorrow.” The assistant calls the customer search action, reads the open opportunities and the latest orders in the ERP, then writes a briefing sheet. After the meeting, it proposes a summary and an update to the opportunity stage; the salesperson reviews and approves, and only then does the server write to the CRM. Every step is logged.

### Why it is useful

For a mid-sized company, the value comes down to three points. Control: only the defined actions are possible, nothing else. Reuse: the same server serves several AI tools, with no exclusive tie to any vendor. Traceability: every read and every write is logged. The protocol specification is public, and more and more software vendors provide their own MCP server.

## How to go about it: the steps with the IT department

- Start from the use cases. List the questions your teams ask the ERP or CRM and the data they re-enter by hand. Each use case tells you which data is needed.
- Take inventory of your software. For each one: is there a connector, a documented API, an MCP server provided by the vendor? Where is the data hosted?
- Define the rights. AI acts with the rights of the user querying it, never more. Through AI, a salesperson does not see what they cannot see in the CRM.
- Start read-only. The first use cases read and prepare. Writing comes next, one action at a time, with user approval.
- Test on a limited scope. One team, one module, a few weeks. Check the accuracy of the answers and the relevance of the rights.
- Log and monitor. Every call from AI to the software is logged. The IT department tracks volumes, errors and unusual access.
- Expand. Add actions, modules and teams, keeping the same rule: the IT department approves each access.

## Watch points: rights, writes and traceability

Access rights. The main risk is an overly powerful technical account that would give AI access to everything. Prefer a connection that carries over the user’s identity and rights.

Writing to software. A change in the ERP can have knock-on effects: order, inventory, invoice. Any write prepared by AI is approved by a person, and sensitive actions (payment approval, data deletion) stay out of AI’s reach.

Misleading content. An email or document read by AI may contain hidden instructions. The possible actions must be limited, so that a misleading instruction cannot trigger serious consequences. ANSSI, the French national cybersecurity agency, has published security recommendations for generative AI systems.

Personal data. The CRM contains personal data: the GDPR applies, with a clear purpose and an entry in your register of processing activities.

Hosting. Check where data travels between the software, the MCP server and the model: in France or on your premises if needed, with no model training on your data. See our approach to security.

## How AI Origin supports you

When we map your operations, we list with your IT department the use cases, the software involved and the best connection route for each. We then bring your teams on board through hands-on sessions, on the chosen AI tool, once it is connected to their software.

When a tool does not connect, often an in-house tool or an older ERP, the AI Factory builds the MCP server or rebuilds the business tool, ready for AI, in a few weeks. The code and the data stay with you: you take back the keys. See our use cases on the in-house ERP and the custom CRM.

## Frequently asked questions

How do you connect ChatGPT, Claude or Copilot to an ERP?

Through a connector provided by the vendor when one exists, through a custom build on the ERP’s API, or through an MCP server placed in front of the ERP. The choice depends on the software, the rights to manage and the number of AI tools to connect. The IT department approves each access.

What is an MCP server?

A component placed in front of a piece of software that describes to AI the authorized actions: search, read, prepare a change. It follows the Model Context Protocol, an open standard, and can serve several compatible AI tools.

Can AI modify CRM data?

Yes, if you allow it, but one action at a time and with user approval. You start read-only, then open up writing for specific actions, such as creating a meeting summary or updating an opportunity stage.

Is MCP secure?

The protocol allows fine-grained control, but security depends on how it is implemented: rights carried over from the user, limited actions, logging, controlled hosting. ANSSI’s recommendations on generative AI are a good basis to work from.

Do you need to change your ERP to use AI?

No, in most cases. An ERP that exposes an API can be connected through an MCP server. If the tool is too closed, you can build a connection layer or rebuild the business tool concerned, without replacing everything.

## Texts and studies cited in this article

Every figure on this site is sourced. See all our sources

- BCG, The Widening AI Value Gap, September 2025, more than 1,250 companies. media-publications.bcg.com (new tab)
- Model Context Protocol, Spécification et documentation du protocole MCP. modelcontextprotocol.io (new tab)
- ANSSI, Recommandations de sécurité pour un système d’IA générative, April 2024. cyber.gouv.fr (new tab)

## Further reading

Your in-house ERP or CRM, rebuilt AI-ready Custom CRM, AI-ready Security and privacy: SOC 2 Type II, GDPR

## Let’s talk about your operations.

Thirty minutes to review current usage, and tell you where we would start.

contact@aiorigin.io

Thank you, your request has arrived.
We will reply to set up a time.
