# How to add an AI chatbot to IFS Cloud > Source: https://netray.co/integrations/ifs-cloud-ai-chatbot > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer An AI chatbot for IFS Cloud answers questions from live IFS data and approved documents, uses the asker's identity so permissions apply, and can be embedded in a portal or opened alongside Aurena. It can run fully on-prem on an open-weight model. ## System IFS Cloud ## Capability ai-chatbot ## Intro Most IFS users do not want a general chatbot. They want to ask 'what is the status of this order', 'who supplies this part' or 'how do I raise a nonconformance' and get a trustworthy answer with a link back to the record. The difference between a helpful assistant and a risky one is grounding. There are two audiences. Internal users such as planners, buyers and technicians benefit from data lookup and process help. External users such as customers and suppliers may need narrower portal assistants limited to their own orders and documents. IFS provides its own assistant capabilities in recent releases, which may cover basic needs. A custom chatbot makes sense when you want your own policies, documents and model, or must keep everything private. ## Use Cases ### Buyer assistant **Description:** Ask which supplier lines are late, what the last price was, or who the alternate suppliers are, and get answers drawn from purchasing data. ### Technician help at the asset **Description:** A technician asks about a fault, gets prior work orders for the asset and the relevant manual extract, in a mobile friendly chat. ### Customer order portal bot **Description:** Customers ask about their own order status and shipment dates, with access restricted to their account. ### Process and policy answers **Description:** Retrieval over your procedures answers how-to questions, such as when a quality hold is required, with citations. ### New user onboarding **Description:** Staff learning IFS ask where to find a function and get a short explanation using your own naming and terminology. ## System Surface OData projections provide live structured data for status, stock and cost questions, queried through tool calls so answers are never from model memory. IFS IAM and OpenID Connect let the chatbot sign users in and pass delegated tokens, so IFS permissions determine what each person can retrieve. Documents in IFS Document Management, plus external shares such as SharePoint or file servers, can be indexed into a retrieval store for policy and manual questions. Aurena supports embedding and linking, so a chatbot can open in a separate panel or portal and deep link to the relevant IFS pages. Exact embedding options depend on your release and configuration. ## How It Works ### Choose the audience and the scope **Detail:** Start with one internal group and ten question types. Write down what the bot must refuse. ### Build the data tools **Detail:** Wrap a few projections as read-only tools and return compact results with record identifiers for linking. ### Index documents **Detail:** Chunk and embed approved documents with metadata for site, revision and access group, and refresh on a schedule. ### Orchestrate retrieval and tools **Detail:** Route each question to data tools, document search, or both, and instruct the model to answer only from the results with citations. ### Authenticate and embed **Detail:** Sign users in with your identity provider, pass delegated tokens to tools and place the chat in a portal or alongside Aurena. ### Evaluate and monitor **Detail:** Keep a test set of real questions with verified answers, track thumbs down and unanswered topics, and iterate weekly. ## Example Flows ### Late order question from a buyer **Trigger:** Buyer types 'which of my POs for supplier X are late'. **Steps:** - Bot identifies the supplier and calls the open PO tool. - Tool returns lines past promised date with identifiers. - Bot summarizes with links to each purchase order. - Buyer asks the bot to draft a chase, which is shown for review. **Result:** The buyer works from one conversation instead of several list pages. ### Customer asks about a shipment **Trigger:** Customer opens the portal chat and asks about order 4410. **Steps:** - Bot verifies the order belongs to the signed-in account. - It calls the shipment status tool. - It responds with the carrier status and expected date. - If unknown, it offers to raise a request to customer service. **Result:** Fewer status emails to the service desk. ## On Prem A chatbot can run completely inside your network. Host an open-weight chat model with vLLM or Ollama, an embedding model, and a vector store such as pgvector on your own servers. The orchestration service calls IFS projections internally and indexes documents from internal shares. Choose a model size by concurrency and language needs; an eight to thirty billion parameter class model often handles lookups and summaries, while longer document reasoning may need more GPU memory. No prompt, document or IFS value leaves the boundary, which matters for export-controlled estates. ## Security - Pass the user's identity to every tool so IFS permissions apply, and apply document access groups in retrieval. - For external users, restrict by customer or supplier account in the tool layer, not just in the prompt. - Block write actions in the chatbot unless routed to an approval step. - Defend against prompt injection from document text and free-text fields. - Retain chat logs according to policy and restrict who can read them. ## Options ### IFS built-in assistant features **Fit:** Basic help where licensing and release already include them. **Effort:** Low. ### In-house chatbot with a framework and your model choice **Fit:** Teams with developers and a defined audience. **Effort:** Weeks for a pilot, ongoing evaluation work. ### Netray or another partner, on-prem **Fit:** Regulated or portal facing chatbots with private models. **Effort:** Pilot scoped around one group, then expansion. ## Pitfalls - Letting the bot answer from model memory when a tool returns nothing. - Indexing superseded document revisions and serving stale procedures. - Applying customer isolation only in the prompt. - Launching to everyone before a test set exists. - Underestimating GPU memory for concurrent users on a local model. ## Faq ### Can the chatbot live inside Aurena? **Answer:** Often it sits beside Aurena in a browser panel or portal and deep links into records. Direct embedding depends on your release and configuration, so check with your IFS administrator. A separate page with single sign-on is the dependable starting point. ### Will the chatbot see data users should not? **Answer:** Not if tools use delegated tokens and document retrieval filters by access group. Relying on prompt instructions alone is unsafe. Test with accounts of different roles to confirm each only receives permitted records. ### How accurate will it be? **Answer:** For structured lookups, accuracy tracks the tools, since values come straight from IFS. For document questions, quality depends on indexing and chunking. Measure it with a test set of real questions rather than trusting demos. ### Can it run without internet access? **Answer:** Yes. With a local open-weight model, embedding model and vector store, the whole stack runs offline inside your network. You trade some top-end model quality for control, which is usually acceptable for lookups and summaries. ### Should it take actions like approving orders? **Answer:** Not initially. Let it prepare a request and hand it to the existing IFS approval flow. Direct approvals from a chat box need strong identity checks and audit, which is a later phase. ## Search terms ifs cloud chatbot, ai chatbot for ifs, ifs erp chatbot, chatbot on ifs data, on-prem chatbot for ifs cloud, ifs customer portal chatbot, ifs aurena ai assistant, private chatbot for erp, ifs rag chatbot --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json