# How to add an AI chatbot to Dynamics 365 Finance and Supply Chain > Source: https://netray.co/integrations/dynamics-365-fscm-ai-chatbot > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer Ground the chatbot in two things: live ERP data through OData or custom services, and your own documents through retrieval. Deploy it where users already work, such as Teams or a portal, enforce FSCM security on every lookup, and decide whether the model is hosted or private. ## System Microsoft Dynamics 365 Finance and Supply Chain ## Capability ai-chatbot ## Intro FSCM is powerful and its forms are dense. New buyers do not know which workspace shows overdue receipts, shop supervisors do not want to navigate to production order details, and customers phone the sales desk to ask where a shipment is. A chatbot lets them ask in a sentence and get the answer or a link to the right form. Two audiences need different designs. An internal assistant can use staff identity and broad read access under their roles. An external assistant for suppliers or customers must be locked to that party's own records, and it should be far more conservative about what it will say. Microsoft's Copilot in Dynamics 365 already answers many in-app questions, so be clear about the gap you are filling: a branded portal assistant, a Teams bot for the floor, support for non-FSCM users, or a model you host yourself. ## Use Cases ### Supplier self-service **Description:** Suppliers ask whether a PO was received, when payment is due, or which lines are open, answered from their own vendor account records only. ### Customer order enquiries **Description:** Customers ask for order status, tracking and invoice copies without calling the sales desk. ### Where do I find it? navigation help **Description:** New staff ask how to release a production order or review a vendor invoice, and the bot answers from your SOPs and links to the form. ### Shop floor questions **Description:** Supervisors ask which component is short for a job, answered from BOM and inventory data with the warehouse named. ### Policy lookups **Description:** Staff ask about approval thresholds or tolerance rules, answered from policy documents with the source cited. ## System Surface - OData v4 data entities under the /data endpoint for live lookups of orders, inventory and vendor or customer records. - Custom X++ services under /api/services when an answer needs logic, such as available-to-promise style calculations. - Microsoft Entra ID for staff sign-in, and Entra External ID or a portal identity layer for supplier and customer access, mapped to account numbers. - Attachments and document handling for SOPs held outside FSCM, typically in SharePoint or a file share, which feed a retrieval index. - Power Virtual Agents or Copilot Studio style tooling in the Microsoft ecosystem, and the Dynamics 365 Customer Insights or Commerce portals where relevant, subject to licence and version. ## How It Works ### Define audience and scope **Detail:** Choose internal staff or external parties first. List the ten questions you want answered, and what the bot must refuse. ### Build ERP lookup tools **Detail:** Create bounded tools for the data answers: order status, receipts, inventory availability. The bot calls tools rather than guessing numbers. ### Index documents for retrieval **Detail:** Chunk SOPs, work instructions and policy PDFs, store embeddings, and have the bot cite the passage it used. ### Bind identity to data access **Detail:** Map the signed-in user to FSCM roles or to a vendor or customer account so tools filter automatically to what that person may see. ### Pick the surface **Detail:** Deploy into Teams, a web portal widget or the Dynamics environment as an embedded link, depending on where users already are. ### Handle escalation **Detail:** When confidence is low or the request is out of scope, hand over to a person with the conversation attached. ## Example Flows ### Supplier asks about payment **Trigger:** A vendor user asks, when will invoice 5521 be paid? **Steps:** - The portal identifies the vendor account from sign-in. - The bot calls a tool for vendor invoices filtered to that account. - It returns status and expected payment date from the payment proposal or terms. - If unposted, it explains the pending approval without exposing internal notes. **Result:** The supplier gets an answer in seconds and AP gets fewer calls. ### New buyer navigation help **Trigger:** A buyer asks how to confirm a purchase order and where it is in the system. **Steps:** - The bot retrieves the matching SOP section. - It summarises steps and names the form path. - It cites the document so the buyer can read the original. **Result:** Faster onboarding with answers tied to your own process documents. ## On Prem A chatbot can run entirely on-premises: the retrieval index, an open-weight model such as Qwen, Llama or Mistral served by vLLM or Ollama, and the lookup tools reaching FSCM over your network. For on-premises (local business data) FSCM the lookup tools use your AD FS based authentication and the entities available in your build. Be realistic about quality: smaller local models need tighter prompts, shorter tool lists and evaluation on your own questions, but for a bounded assistant they are usually adequate and keep every prompt inside your boundary. ## Security - Bind every lookup to the caller's identity; external users see only their own vendor or customer records. - Never allow the model to choose account numbers or legal entities from user text. - Do not expose cost, margin or internal comments to external audiences; return explicit field lists. - Log conversations with retention rules, and warn users not to type confidential data into free text. - Defend against prompt injection from retrieved documents and ERP text fields. ## Options ### Microsoft Copilot or Copilot Studio extended to FSCM **Fit:** Good for internal staff on standard scenarios when you prefer Microsoft-hosted tooling. **Effort:** Configuration plus licensing review, usually weeks. ### Build a custom assistant on your own stack **Fit:** Fits teams wanting a branded portal assistant, custom tools and control over the model. **Effort:** Four to eight weeks including evaluation and identity mapping. ### Partner build such as Netray with a private model **Fit:** For regulated manufacturers needing on-premises inference and document grounding over SOPs and ERP data together. **Effort:** Pilot on one audience and ten questions, then expand. ## Pitfalls - Letting the bot answer numbers from memory rather than tool results, producing confident stock figures that are wrong. - Launching to suppliers before identity mapping is solid, risking one supplier seeing another's records. - Indexing outdated SOPs, so the bot cites superseded procedures. - No escalation path, so unanswered users give up and call anyway. - Judging quality on demos instead of a held-out list of real user questions. ## Faq ### Do we need a chatbot if we have Copilot in Dynamics 365? **Answer:** Maybe not. Copilot handles in-application assistance for licensed users. A custom chatbot helps when you need external supplier or customer access, a Teams or portal experience, grounding in your own documents, or a model hosted on your own infrastructure. Compare the feature set for your release first. ### How does the bot get live ERP answers? **Answer:** Through tools that call OData data entities or custom services, filtered by the caller's identity. The model decides which tool to call and explains the result, but the numbers come from FSCM. Retrieval over documents handles policy and how-to questions. ### Can suppliers and customers use it safely? **Answer:** Yes if identity is mapped to their account and every tool filters on it server side. Return only whitelisted fields, hide internal notes and margins, and test with adversarial prompts. Start with a few suppliers and review conversation logs before opening up. ### Can it run without sending data to a cloud model? **Answer:** Yes. Use an open-weight model on your own GPUs with a local vector store, and let tools reach FSCM over your network. Quality depends on the model size and your evaluation, but for a bounded assistant over known documents and tools it is commonly workable. ### What does it cost to maintain? **Answer:** Mostly content and evaluation work: keeping documents current, reviewing failed conversations, and adjusting tools when forms or entities change after FSCM updates. Plan for a regular review after each service update and keep a test set of questions to rerun. ## Search terms dynamics 365 finance and supply chain chatbot, ai chatbot for dynamics 365 f&o, dynamics 365 supplier portal chatbot, d365 fscm assistant, chat with dynamics 365 data, dynamics 365 copilot alternative chatbot, on premises chatbot dynamics 365, dynamics 365 order status chatbot, dynamics 365 sop chatbot, natural language dynamics 365 supply chain --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json