# How to add an AI chatbot to Epicor Kinetic > Source: https://netray.co/integrations/epicor-kinetic-ai-chatbot > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer Build an assistant that answers from Kinetic data through reviewed BAQs and read-only API calls, then surface it in a web page, Teams or a Kinetic-adjacent portal. Ground every answer in returned records, respect the asker's role, and run the model on-prem when data must stay inside. ## System Epicor Kinetic ## Capability ai-chatbot ## Intro A chatbot on top of Kinetic is useful when it removes the navigation, not when it chats. People do not want to learn which of Part Tracker, Job Tracker or Supplier Tracker holds a fact. They want to ask and get an answer with the record reference. There are two quite different audiences. Internal staff need breadth and speed. Customers and suppliers need a narrow, safe slice such as order status and documents, with strict identity controls. The mistake to avoid is a general-purpose bot that sometimes guesses. In an ERP, a wrong lead time or stock count is worse than no answer. ## Use Cases ### Internal 'ask the ERP' assistant **Description:** Planners ask for availability, open demand and supply for a part, with a link to the record and the as-of time. ### Customer order status portal bot **Description:** A customer logs into your portal and asks where their order is. The bot sees only that customer's orders. ### Supplier self-service **Description:** Suppliers ask which of their PO releases are open and what dates are needed, and confirm dates that feed back as a draft for the buyer. ### Shop floor help **Description:** Operators ask about the current operation, the routing note and the last reported quantity for a job on a kiosk or tablet. ### Policy and procedure Q&A **Description:** A retrieval layer over work instructions, quality manuals and return policies answers staff questions with the source document cited. ## System Surface - BAQs through the BaqSvc OData endpoint for the structured answers the bot will cite. - REST API v2 reads for single records, wrapped behind the bot's tool layer rather than exposed to the model. - Epicor attachments and your document stores for the retrieval side. Index them with a vector store alongside permissions metadata. - Identity: Kinetic users and groups for staff, and your portal's identity provider for customers and suppliers. Map each to allowed data. - Epicor Application Studio can embed a custom panel or link out to a hosted chat page, depending on version and your customisation approach. ## How It Works ### Pick the audience and the thirty most common questions **Detail:** Mine helpdesk tickets and email. The questions define your tools and your test set. ### Build the grounded tool layer **Detail:** Each tool is a BAQ or REST read returning compact JSON with IDs and as-of timestamps. ### Add document retrieval **Detail:** Index manuals and SOPs, tagging each chunk with who may see it. Retrieve only what the asker may access. ### Choose the model and serving **Detail:** Open-weight models on vLLM or Ollama if data stays inside, or a hosted API for non-sensitive scopes. ### Build the front end **Detail:** A web widget in your portal, a Teams bot or a kiosk page. Show the source records beside every answer. ### Evaluate and monitor **Detail:** Run the test set weekly, review thumbs-down answers, and track the share answered with a cited record. ## Example Flows ### Customer asks where order 4821 is **Trigger:** A logged-in customer types the question in the portal. **Steps:** - The bot resolves the customer identity from the session. - It calls the order status tool with a server-enforced customer filter. - It reads lines, promise dates and shipment records. - It answers and offers to email the packing list. **Result:** The customer gets an answer without a call to customer service. ### Planner checks a part **Trigger:** A planner asks, 'Can we cover 40 of P-1187 by Friday?' **Steps:** - The bot calls availability and open supply tools. - It subtracts allocations and shows supply with dates. - It answers yes or no with the numbers, and notes the PO that makes the difference. **Result:** A decision in seconds, traceable to Kinetic records. ## On Prem A fully on-prem chatbot is realistic for Kinetic. Host the application, the retrieval index and an open-weight model within your network. The bot service talks to Kinetic's REST API over your internal LAN, and users reach it through your intranet or a VPN. For customer-facing bots, place only the thin web front end in the DMZ and keep the model and ERP calls behind it. Smaller models can handle structured lookup well, so GPU needs are modest unless you add heavy document Q&A. ## Security - Enforce the asker's identity and data scope on the server, never in the prompt. - Use read-only access, and send any action, such as a date confirmation, to a human queue. - Hide internal-only fields like cost and margin from customer and supplier modes. - Add prompt-injection defences for retrieved documents and inbound text. - Retain conversation logs per policy, with redaction of personal data. ## Options ### Epicor's built-in assistant features **Fit:** General help and navigation inside the product, if licensed. **Effort:** Low, limited to what Epicor ships. ### In-house chatbot on your stack **Fit:** Teams with API developers and a clear first audience. **Effort:** Six to ten weeks for an internal pilot. ### Netray with private models **Fit:** On-prem, regulated or customer-facing scopes needing strict controls. **Effort:** Scoped pilot in weeks, then expand by audience. ## Pitfalls - Letting the bot answer from memory when no record is found. - Mixing internal and customer-facing data in one tool set. - Skipping an evaluation set, so quality drifts unnoticed after a Kinetic upgrade. - Indexing documents without permissions, which leaks restricted drawings. ## Faq ### Can the chatbot live inside Kinetic? **Answer:** Often you can link to or embed a hosted page from a custom Kinetic panel, depending on version and how you customise. Many teams find it simpler to place the bot in Teams or an intranet page and deep link to Kinetic records. ### How do I stop it from making up answers? **Answer:** Force answers to come from tool output, require record references, and return 'I could not find that' when tools return nothing. Evaluate on a fixed question set after every change. ### Can customers use it safely? **Answer:** Yes with strict controls: authenticate them through your portal, enforce customer filters in the tool layer, expose only status and documents, and send any request that changes data to a person. ### What hardware does an on-prem bot need? **Answer:** A mid-size open-weight model for structured lookup runs on one modern server GPU. Heavier document Q&A, many concurrent users or larger models need more. Size it from a pilot rather than guessing. ### Will it work with our Kinetic customisations? **Answer:** Yes if the data is reachable via BAQs or the API, including user-defined fields. Customisations that change business logic should be tested, since BPM directives can alter results for some calls. ## Search terms epicor kinetic chatbot, ai chatbot for epicor, epicor erp chatbot, epicor kinetic virtual assistant, epicor customer portal chatbot, on-prem chatbot epicor, epicor kinetic natural language query, chat with epicor data, epicor kinetic ai assistant, private chatbot for erp epicor --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json