# Private AI Layer for Infor VISUAL On-Prem: Reference Architecture > Source: https://netray.co/erp-help/private-ai-layer-infor-visual-on-prem-architecture > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: ERP Help ## Short answer Infor VISUAL runs on SQL Server (or Oracle), so the quickest private AI path is a read-only reporting copy plus a local document index, a tool server with named read actions, and an open-weight model on your own GPUs. Use the VISUAL API Toolkit for writes only, with human approval, and mirror VISUAL user security in the tool server. ## Kind advanced ## System infor-visual ## Erp Infor VISUAL (VISUAL Manufacturing) ## Module AI integration / SQL Server and VISUAL API Toolkit ## Query how to add AI to Infor VISUAL on-prem ## Alternate Queries - Infor VISUAL AI integration - VISUAL Manufacturing chatbot SQL Server - VISUAL API Toolkit AI assistant - chat with VISUAL ERP data on-prem ## Applies To Infor VISUAL (VISUAL Manufacturing) 8.x-era and later on-premises, SQL Server (Oracle in some estates) ## Steps ### How to build a private AI layer on Infor VISUAL **Items:** - List the questions to answer: open orders, work order status, inventory, purchasing, quality, plus documents such as drawings and certs. - Create a read-only reporting copy of the VISUAL database (AlwaysOn secondary, log shipping or scheduled restore) and a SELECT-only service account. - Map the tables and views you need and note site and entity filters, because VISUAL can run multiple sites or entities. - Index drawings, work instructions and attachments with a local embedding model, with permission metadata. - Expose named read actions through an MCP or tool server; each tool maps to one parameterised query. - Serve an open-weight model on local GPUs and log every prompt, query and answer. - Phase the rollout: read-only questions, drafts, then approved write actions through the VISUAL API Toolkit. ## Sections ### Data access paths in VISUAL **Body:** VISUAL stores data in a relational schema with readable table names: PART and PART_SITE for items, CUSTOMER_ORDER and CUST_ORDER_LINE for sales, PURCHASE_ORDER and PUR_ORDER_LINE for purchasing, WORK_ORDER for production, INVENTORY_TRANS for stock movements, CUSTOMER and VENDOR for partners. That readability makes SQL on a reporting copy the fastest first win; confirm column names in your own schema. The supported programmatic layer for writes is the VISUAL API Toolkit, which applies business rules instead of raw table writes. Do not let AI write with UPDATE statements; transactions in VISUAL touch several tables. Drawings, certs and notes live in a mix of VISUAL attachments, linked file paths and shares. Index the file paths and the files. **Code:** -- Reporting copy, read-only (verify column names in your schema) SELECT TOP 50 co.ID, co.CUSTOMER_ID, co.ORDER_DATE FROM CUSTOMER_ORDER co WHERE co.STATUS = 'R' ORDER BY co.ORDER_DATE DESC; ### Reference architecture **Body:** Keep five components on your network: a reporting replica, a document index, a tool server, a model server and an audit store. The tool server holds the database credentials; the model never does. For small plants, a single GPU server beside the database host can run the whole assistant, but keep the model on separate hardware from the production SQL Server. Air-gapped installs load weights and containers through a transfer station. Use small, well-described tools such as open orders by customer, work order status by job, stock by part and warehouse, so the model rarely builds SQL itself. ### Permissions and write-back **Body:** VISUAL has its own user and security setup that controls which functions and data a user can reach, and sites or entities can partition data. The assistant must honour those: map each person to their VISUAL user, scope queries by site or entity, and keep sensitive areas such as cost and pricing behind explicit grants in the tool server. Write-back: only through the VISUAL API Toolkit or the VISUAL application, with a named user approving a drafted action such as a PO line change. Treat any direct SQL write as out of scope. ### Sizing, limits and honest caveats **Body:** VISUAL estates often have 25 to 250 named users; one inference GPU normally handles the interactive load, with a document index in the single-digit to tens of GB. Use the calculator for your own numbers. Limits: customised schemas and user-defined fields need mapping; older versions may differ in table or column details; costs and MRP results should be read from the system, not recalculated by the model. If the need is only standard reports, VISUAL's reporting tools may be enough before adding a model. ## Pitfalls - Pointing the assistant at the production SQL Server and slowing shop-floor transactions. - Returning cost and pricing data to users who lack that security in VISUAL. - Ignoring multi-site or entity filters so one plant sees another's orders. - Writing with SQL UPDATE statements that bypass VISUAL business rules. - Assuming column names without checking your version and customisations. ## Ai Angle ERPray and DataRay can sit on a reporting copy of your VISUAL database, show the SQL behind each answer, respect user security, and run on your own GPUs. Write actions are drafts until a named user approves them. ## Faq ### Can VISUAL use AI without the cloud? **Answer:** Yes. A read-only SQL reporting copy, a local document index and an open-weight model served on your GPUs run entirely on your network, including air-gapped. ### Is there an API for VISUAL? **Answer:** The VISUAL API Toolkit exists for supported transactions. For reading, a SQL reporting copy is usually faster and safer for production. ### How are VISUAL permissions mirrored? **Answer:** Map users to VISUAL users and their site or entity access in the tool server, and apply those filters to every query and retrieval step. ### Do I need a big GPU? **Answer:** Usually not. A single modern GPU in the 24 to 80 GB class covers the interactive load of a typical VISUAL plant with a 8B to 32B model. ## Search terms Infor VISUAL AI, VISUAL Manufacturing AI, VISUAL ERP chatbot, VISUAL API Toolkit, Infor VISUAL SQL Server, private LLM VISUAL ERP, VISUAL on-prem AI, chat with VISUAL data --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json