Overview
Nama ERP has a built-in AI assistant: a chat window inside the system where anyone can ask a question in plain Arabic or English and get an answer from the company’s actual data — not a generic model guessing at how an ERP works.
The important part is not that it talks. It is that it is governed. The assistant has no capability of its own. It can do nothing except through tools your system administrator has defined, and every tool executes as a real user, with that user’s permissions and dimensions — legal entity, branch and the rest — fully applied.
So the assistant sees only what the person asking is allowed to see, and does only what that person is allowed to do. For a system holding your ledger, payroll and customer data, that boundary is the whole point.
Features
Answers from your data, not from training data
Ask “how many sales invoices this month?” or “who are the top five customers by balance?” and the answer comes from a query running against your database now. Open the assistant from inside a report and the report’s output becomes context it can read — ask it to summarise the results or pick out the highest values. Open it from a record screen and it can discuss the document in front of you.
It can act, not only read
Four kinds of tool can be defined, and together they cover reading and doing:
- Query tools — a SQL query the administrator writes, which the model may run with parameters it infers from the question.
- Report tools — any existing report definition becomes something the model can run and then read the output of.
- Entity-flow tools — a predefined flow the model can execute to perform an action: approve a document, post it, and so on.
- System tools — ready-made capabilities including record export and import, and searching the Nama ERP documentation.
Each tool carries its own access-control grid, so a tool that posts journal entries can be restricted to the people who should be posting journal entries.
Connect Claude and other AI clients directly
Nama ERP ships with a built-in MCP server. MCP — the Model Context Protocol — is the open standard AI assistants use to connect to external systems, and it means clients such as Claude Desktop and Claude Code can connect straight to your ERP and use the same tools, under the same controls, as the in-system assistant.
Authentication is per request, through API credentials that map to a specific user account with a validity window and the ability to be disabled. Everything the client does runs as that user: record security, dimensions and validation rules all apply exactly as if they were working from the system screens.
The in-system assistant and an external MCP client are two faces of one base — the same tool definitions, the same access check at execution time.
The model provider is your choice
Eleven providers are supported — OpenAI, Anthropic, Google Gemini, AWS Bedrock, DeepSeek, GCP, Mistral, Azure, Zhipu AI, Hugging Face, and any OpenAI-compatible endpoint, which covers self-hosted models.
That last one matters for organisations that will not send business data to a third party: you can point the module at a model running on your own infrastructure. Multiple chat models can be defined and switched per task — a fast one for simple questions, a stronger one for real analysis.
Semantic search over your records
Rather than matching text literally, the system converts records and documents into embeddings held in a vector store, so it can find the record closest to a meaning. That is what lets someone type an approximate customer name instead of a code and still land on the right record. You choose which entities are indexed.
Expert Mode, for when an answer looks wrong
Normal mode shows a conversation. Expert Mode shows the machinery: which tool the assistant called, with which parameters, and what came back, alongside the system instructions guiding the model.
This turns “the AI gave a strange answer” from a mystery into something a support engineer can actually diagnose — is it the tool’s description, the query, or the user’s permissions? Answers can also be rated, and the ratings are stored with the message to guide refinement of tool definitions over time.
Requirements
The AI module is licensed separately. It needs at least one chat model configured and at least one committed tool — without tools the assistant can converse but cannot reach system data, which is the deliberate default. Connecting an external MCP client additionally needs an API Credentials record.
The full setup and tool-authoring guide is published in our documentation.
