Should you connect AI to Epicor Kinetic through MCP?

MCP makes it easy to connect an AI assistant to Epicor Kinetic. Here's what that actually exposes, how to reduce the risk, and how it differs from a bounded integration.

·Mark Betts
Should you connect AI to Epicor Kinetic through MCP?

MCP (Model Context Protocol) is a standard way of connecting AI assistants to other systems. There are already MCP servers for Epicor Kinetic, including open-source ones, and they're genuinely useful. You can ask an assistant about orders, stock or customers and it goes and finds the answer.

If you're an IT manager deciding whether to allow it, this is my honest take. It's not "MCP is dangerous". It's that you should know what you're turning on.

What MCP actually does

An MCP server gives an AI assistant a set of tools, such as "run this query" or "call this business object". The AI decides which tools to use, when, and with what inputs, based on what you asked and whatever data it has seen so far.

That flexibility is the point, and it's also the risk. You're not running a fixed process. You're letting a model choose actions on the day.

Where the risk is

  • What it can see. The assistant can reach whatever the server's account can reach. If that's a broad service account, a question about one order can pull data from anywhere that account has access to.
  • Whose permissions it uses. Some servers run as one shared account for everyone. Better ones pass through the user's own identity, so the AI can only see what that person could see in Kinetic anyway.
  • What it reads. Customer emails, notes and documents can contain text that looks like instructions. An assistant reading them can be steered by them. This is called prompt injection, and no prompt fully prevents it.
  • What else it's connected to. An assistant with Kinetic access and the ability to send email or post to the web can move data out of the business, whether or not anyone meant it to.
  • Where it runs. A local MCP server runs as code on your machine, with your privileges. You're trusting that code.

How to reduce it

  • Run the server yourself, and preferably use one whose code you can read.
  • Keep it read-only, and give its account the least access it needs.
  • Limit what it can see to specific BAQs or tables.
  • Pass through the user's own identity, not a shared account.
  • Don't connect it alongside tools that can send data outside the business.
  • Require approval before tool calls, and log everything.

Done like that, MCP is a reasonable way to give people a natural-language way into Kinetic data.

How this compares with lowki.ai

lowki.ai has data flowing through it too. It reads email, talks to Kinetic and sends document content to an AI model. Any IT manager will know that, so I won't pretend otherwise.

The difference is that its behaviour is bounded. It does a fixed, known set of operations every time. In lowki.ai today, the AI reads documents into a fixed structure and has no tools. It reads Kinetic through BAQs you can see and change, and it creates records through Kinetic's own business objects, so your rules and validation still apply. Anything it's unsure of is flagged for a person to check, and we recommend holding new orders so every one is reviewed before it's processed.

MCP gives the AI choices. lowki.ai's order entry deliberately doesn't. They're different tools for different jobs: MCP for open-ended questions, a fixed pipeline for the transactions you post into your ERP every day.

The question to ask

Whatever you connect to Kinetic, ask: what can this do that I didn't explicitly ask for? If the honest answer is "quite a lot", put the controls above in place before you switch it on.

Related: Built on Kinetic, not beside it · Why we didn't use an AI agent for order entry · Security and architecture · Sales order entry · Accounts payable

About the author: Mark Betts

Business Technology Manager, Thomas Dudley Ltd

Mark is Business Technology Manager at Thomas Dudley, a UK manufacturing group running Epicor Kinetic, where he leads the business technology team that supports the whole group, from infrastructure and desktop support through to business systems and ERP development. He has over 20 years of commercial IT experience, all of it inside Thomas Dudley, developing and deploying business systems improvements, new routes to market and efficiency gains across the group. He works on both sides of the table: hands-on with AI, C#, SQL and Kinetic, and translating what a business needs into systems that actually deliver it. An on-prem advocate.

Mark Betts on LinkedIn