sign in
What Is an AI Assistant in ERP, and How Is It Different from ChatGPT?
ChatGPT or generic AI
AI assistant integrated into ERP
How It Works Technically
- 1
Understanding the question
When a user writes, “Which products have fewer than 10 units in stock?”, the AI analyses the request and identifies the query type, the relevant parameters, and the data that must be returned. In this case, it recognises inventory as the query type, a quantity below 10 as the filter, and products and available quantities as the expected output.
- 2
Searching the catalogue of predefined queries
The system maintains a catalogue of standard, preconfigured queries. If the user’s question matches an existing report, the AI runs it directly. Predefined queries are tested, optimised, and controlled, which reduces the risk of incorrect results caused by ambiguous wording.
- 3
Custom query as a fallback
If the question does not match any predefined report, the AI can generate a custom query based on the system’s data structure. This is part of agent mode. It is more flexible, but it also carries a higher risk of ambiguous interpretation. It is used only when the report catalogue does not contain a suitable answer.
What AI Can Do in an ERP System in Practice
Inventory queries
- “How many boxes of milk are available in the Bucharest warehouse?” → inventory filtered by product and warehouse
- “Which products will expire within the next 14 days?” → inventory filtered by batch or lot and expiry date
- “Which products are below the minimum stock level?” → comparison between current inventory and configured minimum levels
Orders and sales
- “How many open orders do we have from Berezka?” → orders filtered by open status and customer
- “Show me this week’s sales for store 3” → sales report filtered by period and store
- “What is the total value of undelivered orders?” → total value of orders with an open delivery status
Searching master data
- “Find the supplier that sells sunflower oil. I cannot remember the exact name” → semantic or fuzzy search across suppliers and products
- “Show me products in the beverages category priced above 20 RON” → combined filtering by category and price
“Show me products in the beverages category priced above 20 RON” → combined filtering by category and price
An important feature of an AI assistant in ERP is semantic search. Unlike conventional search, which looks for an exact text match, semantic search can understand that “sun oil” and “sunflower oil” refer to the same type of product.
This works through a vector database. Each product, customer, or supplier in the ERP system is converted into a numerical vector representing its meaning. When a user enters a search query, that query is also converted into a vector, and the system identifies the records that are closest in meaning.
An important configuration requirement is that products and customers must be vectorised regularly, either daily or whenever records are changed. This process should not run only during the initial implementation. Otherwise, newly added products will not appear in semantic search results.
Agent Mode: AI That Performs Actions, Not Just Answers Questions
The advanced version of the AI assistant can work as an agent. This means it can perform a sequence of steps to complete a more complex task.
For example, a user may write: “Create a purchase order for products that are below the minimum stock level.” The agent then: (1) checks current inventory, (2) identifies products below the configured minimum, (3) calculates the required order quantity based on historical data, and (4) creates a draft purchase order.
This functionality is not a magic button. Business rules such as minimum stock levels, preferred suppliers, and minimum order quantities must already be configured correctly in the ERP system. The AI executes the logic; it does not invent it. Any action with a real operational or financial impact must be confirmed by a user.
Important Limitations to Understand
AI does not make the final decision; it executes. It can create a draft purchase order, but it should not send the order to a supplier without human confirmation.
Some AI functions may need to be disabled. Not every company needs fuzzy search or agent functionality. If the product catalogue is small and well structured, conventional search may be more precise. AI features can therefore be enabled or disabled selectively.
The quality of the answer depends on the quality of the data. If the inventory recorded in the ERP system does not reflect the physical stock, the AI will return an inaccurate answer. AI does not automatically correct incorrect data; it reports the data available in the system.
