Agentic ERP

SMEs shopping for an ERP face an awkward choice. Legacy ERPs carry deep business logic but are hard to use and have no meaningful AI; modern SaaS tools are easy to use but lack ERP depth such as double-entry accounting or inventory valuation; and building from scratch costs too much. On top of that, frontline teams talk to customers in LINE and voice notes, so the information never reaches a system. The brief: an ERP with real depth across sales, purchasing, stock and accounting, where AI turns chat and voice into records — without asking everyone to learn a new system first.

We didn’t reinvent the business logic. We studied Odoo’s open-source code to understand how double-entry accounting and inventory valuation work, then re-implemented it in TypeScript on our own architecture, with AI built in from the start.

1. 1. Plug-in modules that talk through events

   Each module declares its routes, permissions, events and AI capabilities in a manifest, and modules talk through domain events rather than calling each other’s code, so one can be added or removed without breaking another. Built on Next.js 16, Bun, Hono, Prisma and MongoDB.

2. 2. Tenant isolation at the data layer

   Every document carries a tenant_id enforced by Prisma middleware and the request context. Permissions take the form module.resource.action, and every create, update and delete is written to the audit log automatically.

3. 3. LINE and voice as inputs

   A LINE group connects with an activation code and waits for an admin’s approval in the web app before its messages are tracked and summarised into reports. Voice notes are stored in the organisation’s own Google Drive, transcribed and summarised with Gemini, shared through optionally password-protected links and exported to PDF.

4. 4. An orchestrator with specialist agents

   A separate AI service built on LangGraph.js and the Vercel AI SDK routes each request to one of seven specialist agents, using Gemini, Claude or OpenAI models. Every action the AI wants to take is scored on the risk of the change, the user’s role and the organisation’s credits, which decides whether it runs straight away or waits for a person to approve.

5. 5. Designed for non-technical teams

   The interface runs in Thai and English, a Lite mode hides advanced menus, and the data-analysis module accepts CSV, JSON, Excel and Parquet files for the AI to summarise and chart.

Agentic ERP runs in production as a multi-tenant service, with the AI orchestrator deployed alongside the web app as its own service.

• The whole business on one record

  Customers, projects, sales, purchasing, stock and accounting sit on one organisation’s data, so managers stop stitching reports together from separate tools.

• LINE chats become reports

  Approved LINE groups are summarised into reports, in Thai by default, and each organisation can add its own context or instructions for the AI.

• Calls become customer notes

  Voice notes come back transcribed, with brief, in-depth and conversation views, and can be linked to a customer or project at any time.

• AI asks before it acts

  The assistant answers from the data straight away; risky changes wait for a person’s approval.