We gave our ERP's AI assistant the same permissions as the staff. Here's how it works with Claude over MCP The ERPoida team built an ERP whose AI assistant runs under the same permission checks as the human user rather than a separate service account, with every tool call routed through a registry that verifies the caller's access before execution. The system exposes an MCP server so users can connect Claude or ChatGPT to their workspace, and it also acts as an MCP client for other servers, all under the same access controls. The team split storage into a flexible metadata-driven tier for CRM and HR and strict relational tables for ledger, invoices, inventory and payroll, with rules enforced in the database itself. Most "AI in the ERP" demos have the same problem. The assistant talks to the database through a service account, so it can see and change far more than the person typing the prompt. That's fine in a demo. It isn't fine when the data is payroll, invoices and customer records. When we started building ERPoida https://erpoida.com , we made one rule early: the AI never gets its own permissions. It works as the user, under the same checks, and every action is logged. This post walks through how that works and what it means once you connect Claude or ChatGPT over MCP. ERPoida has three levels of access control: The AI assistant doesn't sit next to this. It sits inside it. Every tool the assistant can call goes through a registry, and the registry checks the caller's permissions before the tool runs. If a sales rep asks the assistant to "show me everyone's salary," the request fails the same way it would fail in the UI. The refusal is logged too. That gives you three useful properties: The second thing we cared about is that every business runs a bit differently. In most ERPs, adapting the system means a consultant, custom code or both. ERPoida is metadata-driven. Objects, fields and views are definitions stored as data, so: Because the AI uses the same tools as the user, it can also do this work. You can ask it to add a field, set up a view or create an automation, and it does that inside your permissions, not around them. Flexible data is great for CRM and projects. It's a bad idea for money. So we split storage into two tiers. Things like CRM, HR and helpdesk live in a flexible store. The ledger, invoices, inventory and payroll live in strict relational tables, and the rules are enforced in the database itself : This matters even more with AI in the loop. You don't want "the model was confident" to be the only thing between you and a broken ledger. ERPoida exposes an MCP server, so you can connect Claude or ChatGPT to your workspace and work with your data in plain language: The assistant signs in as you. It can only do what you can do. If you can't approve payroll, neither can Claude. ERPoida is also an MCP client . You can connect other MCP servers you already run, and their tools become available to ERPoida's assistant and automations, again under the same permission checks. The setup guide is here: Connect Claude or ChatGPT https://erpoida.com/help/mcp . Automations are saved instructions that run on a schedule or when a record changes, for example when a deal moves to "Won" or an invoice becomes overdue. Each run is logged, so you can see what happened and why. ERPoida covers CRM, sales, accounting, payments, inventory, purchasing, manufacturing, people HR , payroll, projects, helpdesk, point of sale, marketing and contacts. The interface is available in English, German and Arabic, and the finance side includes German localisation. We'd love feedback from people who have fought with ERPs before, especially on the permission model and on what you'd want an agent to do inside one. Written by the ERPoida team, with help from AI tools for drafting and editing.