# AI for Dynamics 365 Business Central in Discrete Manufacturing > Source: https://netray.co/erp-ai/dynamics-365-business-central-discrete-manufacturing-ai > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: AI for ERP ## Short answer AI for Business Central in discrete manufacturing connects a private language model to production orders, production BOMs, routings, item ledger entries and the planning worksheet, so planners and plant managers can ask plain-English questions and get answers grounded in live data with the query shown. It runs next to your tenant or on-prem server, reads through API pages and OData, and drafts changes for human approval instead of writing silently. ## Kind industry ## Erp Family microsoft ## Erp Products - Dynamics 365 Business Central - Business Central On-Premises ## Industries - Discrete Manufacturing - Fabricated Products - Machinery Components ## Region Global ## Persona Operations Manager ## Eyebrow Business Central + discrete mfg AI ## Intro Business Central is a common landing spot for discrete manufacturers that outgrew QuickBooks or Sage 50 but do not need a full Dynamics 365 Finance and Supply Chain footprint. A 40 to 250 person shop with job-style and repetitive work typically runs production orders, multi-level production BOMs, routings with work centers and machine centers, and the planning worksheet, often with a handful of AL extensions bolted on for labeling, EDI or quality. The daily friction is rarely the engine itself. It is that the answers planners need sit across several pages: a released production order, its components and their reservations, the capacity ledger entries against the routing, the purchase lines that are supposed to cover the shortage, and the cost variance that will show up after Adjust Cost - Item Entries runs. Each is a separate list, filter and export, and the person who knows the path is usually one planner. Microsoft's own Copilot features in Business Central are useful for sales-side text, item descriptions and some analysis, and for many shops that is enough. Where they stop is grounded, auditable questions over your customized production data, your own flushing methods and routing conventions, and any requirement that item costs, customer drawings and supplier terms stay inside a boundary you control. This page describes what AI on Business Central looks like for a discrete manufacturer: the objects it reads, the integration surface that is actually available, where a private model genuinely helps, and where standard pages, Power BI or the planning worksheet are simply the better tool. ## Pain Points ### Shortages are found late and by hand **Detail:** Component availability is spread across Prod. Order Component, reservation entries, open purchase lines and transfer orders. Finding why a released order cannot start usually means opening four pages and asking a buyer. ### Planning worksheet output is trusted or ignored, rarely understood **Detail:** After Calculate Regenerative Plan, planners face hundreds of action messages. Without clear reasons tied to reordering policy, lot accumulation period and safety lead time, they either accept everything or override from memory. ### Routings and actual run times drift apart **Detail:** Standard run times in routing lines age while capacity ledger entries show what really happened. Nobody has time to compare them across thousands of operations, so quotes and schedules stay optimistic. ### Cost variances surface at month end **Detail:** Production variances from output and consumption journals post after Adjust Cost - Item Entries and are read in value entries. Explaining a bad number to finance means a manual trace from the item up through BOM levels. ### Customization limits generic tools **Detail:** Extensions add fields to Production Order, Item and routing tables. Off-the-shelf analytics tools see the standard schema only, so questions about your own flags and statuses cannot be answered without custom work. ## Use Cases ### Shortage explanation for released production orders **Description:** A planner asks why a released order cannot start. The assistant lists missing components, the supplying purchase lines with promised dates, existing reservations competing for the same stock and the nearest alternative item, with the underlying filters shown. **Erp Objects:** Production Order, Prod. Order Component, Reservation Entry, Purchase Line, Item Ledger Entry **Outcome:** Replaces a multi-page hunt with one grounded answer; planners spend their time on the decision, not the trace. ### Planning worksheet action-message triage **Description:** The agent reads the planning worksheet after a regenerative plan, groups suggested lines by cause (demand change, supply slip, reorder policy) and flags the few that need human judgment. Nothing is carried out until a planner approves it. **Erp Objects:** Requisition Line, Planning Component, Item planning parameters, Sales Line, Production Forecast Entry **Outcome:** Reduces review of a long worksheet to the exceptions that matter; accepted lines are traceable to a stated reason. ### Routing time versus actual comparison **Description:** The assistant compares standard run and setup times in routing lines with capacity ledger entries per work center and item family, and proposes where the routing is stale. **Erp Objects:** Routing Line, Capacity Ledger Entry, Work Center, Machine Center **Outcome:** Gives estimators and schedulers evidence for routing updates; quote accuracy improves where the data was previously ignored. ### Production variance explanation **Description:** Given an item or order, the agent walks output, consumption and capacity entries and the associated value entries to explain a material, capacity or overhead variance in plain language for finance. **Erp Objects:** Value Entry, Item Ledger Entry, Capacity Ledger Entry, Production Order **Outcome:** Cuts the month-end back-and-forth between operations and cost accounting from days to a reviewed explanation. ### Supplier follow-up drafting **Description:** The agent finds purchase lines past promised date that gate released orders and drafts follow-up emails for the buyer to review and send. **Erp Objects:** Purchase Header, Purchase Line, Vendor, Prod. Order Component **Outcome:** A morning scan of overdue lines becomes a queue of drafts a buyer approves in minutes. ### Engineering change impact check **Description:** Before a production BOM or routing version is certified, the assistant lists open orders, planned orders and stock affected by the change and the cost roll-up difference. **Erp Objects:** Production BOM Header, Production BOM Version, Routing Version, Prod. Order Line, Item **Outcome:** Fewer surprises where a revision hits work already on the floor. ### Plant manager daily briefing **Description:** A scheduled read-only summary of late orders, constrained work centers and material risks, with each statement linked to the page or filter that supports it. **Erp Objects:** Production Order, Work Center Calendar Entry, Capacity Ledger Entry, Sales Line **Outcome:** Replaces a standing morning meeting prep exercise with a reviewed brief; managers still verify before acting. ## Architecture **Overview:** The AI layer sits beside Business Central, not inside it. It reads through published API pages, OData web services or, on-prem, a read replica of the SQL database, retrieves and reasons with an open-weight model hosted on your hardware or private cloud, and returns answers with the exact query shown. Any change is staged as a draft for a user to post. **Layers:** #### Business Central data surface **Detail:** Standard API v2.0 covers sales, purchasing and items. Most manufacturing objects need custom API pages or OData-published pages and queries written in AL, scoped read-only. #### Access and identity **Detail:** A Microsoft Entra app registration with a dedicated, minimal permission set, or per-user delegation so answers respect each user's company and permission scope. #### Retrieval and semantic layer **Detail:** A curated map of production order, BOM, routing and ledger concepts, including your extension fields, so the model chooses real columns instead of guessing. #### Open-weight model serving **Detail:** Llama, Qwen, Mistral or gpt-oss class models served with vLLM or Ollama on your GPUs or sovereign cloud, sized by user concurrency. #### Approval and audit **Detail:** Writes become drafts (journal lines, purchase lines, worksheet changes) a human posts. Every prompt, query and answer is logged against the user. ## Deployment Options ### Private model beside a SaaS tenant **Fit:** Shops on Business Central online that want model traffic off public AI endpoints **Detail:** The model runs in your Azure subscription or on-site GPU box and calls tenant API pages with service-to-service OAuth. Data leaves the tenant only to your own infrastructure. ### Fully on-prem with Business Central On-Premises **Fit:** Plants with network restrictions or customer NDAs **Detail:** Everything runs inside your network. A read replica or read-only SQL login avoids load on the production database, and the model server needs no outbound internet. ### Hybrid with a sovereign cloud region **Fit:** Multi-site groups with data residency rules **Detail:** Retrieval indexes and model sit in a regional private cloud, with tenant access through scoped APIs and logged calls. ## Compliance ### Answers cite the underlying records, supporting document control and traceability expectations; the AI does not replace a controlled procedure. **Framework:** ISO 9001 and customer quality clauses ### A customer running the stack in a US-person-only, controlled environment keeps technical data out of public model services. Compliance remains the customer's obligation. **Framework:** ITAR / EAR for defense-adjacent work ### Read-only by default, drafts only for postings, and a prompt log that supports review of anything that touched cost or inventory. **Framework:** SOX-style financial controls ### Documents attached to items and orders stay on infrastructure you specify; the index respects the same access scope as the source. **Framework:** Customer NDAs and drawing control ### Minimization of personal fields in the index, retention limits on logs and region-pinned hosting. **Framework:** GDPR where staff and supplier data apply ## Integration Notes - Standard API v2.0 is a poor fit for production data; publish AL API pages for Production Order, Prod. Order Line, Prod. Order Component and Routing Line with read-only permissions. - Use OData-published pages or queries for ledger data such as Capacity Ledger Entry and Value Entry, with filters on posting date to bound payload size. - Authenticate service calls with an Entra app using client credentials, and give it a custom permission set rather than SUPER. - On Business Central On-Premises, point retrieval at a SQL read replica; table names include the company and extension GUID suffix, so a mapping layer is needed. - Throttle requests: the online service limits concurrent and per-minute calls, so cache slow-changing master data like items, routings and work centers. - Treat extension fields explicitly; add them to the semantic map or the model will ignore them. ## Product Fit ### ERPray **Why:** Natural-language questions, dashboards and read-only agents over production orders, planning and costing, with the query shown and Business Central permissions honored. ### Custom build **Why:** For bespoke agents such as planning-worksheet triage, routing recalibration, or integration with MES, EDI or quality add-ons. ## Engagement ### Discovery and data map **Duration:** 1 to 2 weeks **Deliverables:** - Inventory of extensions and custom fields - Priority question list from planners and finance - Access and security plan ### Data surface build **Duration:** 2 to 3 weeks **Deliverables:** - Read-only API pages or OData endpoints - Semantic map of manufacturing objects - Test set of known-answer questions ### Pilot with planners **Duration:** 3 to 4 weeks **Deliverables:** - Working assistant for shortages and worksheet triage - Accuracy review against known answers - Prompt and query logging ### Harden and extend **Duration:** 2 to 4 weeks **Deliverables:** - Approval workflow for drafts - Operations runbook - Backlog of next agents ## Buyer Questions - Which Business Central objects will be exposed, and through what read-only interface? - How are our extension fields and custom statuses included? - Does each answer show the query or filter it used, so planners can verify it? - Which model runs where, and does any prompt or data leave our environment? - How do you handle API throttling in Business Central online? - What can the assistant write, and who approves it? - How are answers tested after each Business Central release wave? ## Faq ### Is Microsoft Copilot in Business Central not enough for a manufacturer? **Answer:** It can be, for sales text, item descriptions and basic analysis. The gap appears for grounded, auditable answers over production orders, BOMs and routings, especially with extension fields, and when policy forbids sending cost and drawing data to external services. Try Copilot and standard pages first; add a private layer only for the questions they cannot answer. ### Can the AI release or change production orders? **Answer:** By default it cannot. It reads, explains and drafts. A proposed change, such as a reschedule or a purchase line, is staged for a planner to review and post using Business Central's normal permissions, so the audit trail and costing behave exactly as they do today. ### How does it get production data if the standard API is limited? **Answer:** We publish custom API pages or OData endpoints in a small AL extension, scoped read-only to the manufacturing objects that matter. On-prem installs can also use a SQL read replica. Both approaches avoid scraping the user interface and keep access explicit and reviewable. ### Will it handle our extension fields and local conventions? **Answer:** Yes, if they are mapped. During discovery we inventory table extensions and page customizations and add them to the semantic layer with plain-language definitions. Fields not described are effectively invisible to the model, which is why the mapping step matters more than model choice. ### Does it help with the planning worksheet or replace it? **Answer:** It does not replace it. The planning engine still calculates the plan. The AI summarizes action messages, explains causes and highlights exceptions so planners review less. Poor planning parameters on items remain the biggest lever, and an assessment of reordering policies is often a better first step. ### What hardware is needed for a mid-size plant? **Answer:** A single server with one or two data-center GPUs usually serves a 7B to 32B parameter open-weight model for tens of concurrent users. Exact sizing depends on question volume and context length. Our GPU sizing tool gives a starting range, and a pilot confirms it. ### What happens when Business Central upgrades? **Answer:** Online tenants get two release waves a year. We keep a regression set of known questions and rerun it after sandbox preview, checking custom API pages for compatibility. Breaking changes show up before production, not after. ## Search terms business central discrete manufacturing ai, dynamics 365 business central manufacturing ai, ai for business central production orders, business central planning worksheet ai, business central copilot alternative manufacturing, private llm business central, business central on-premises ai, business central natural language query, business central routing capacity ledger analysis, ai for dynamics 365 manufacturers, business central production variance explained, business central ai agent manufacturing --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json