# IFS Applications 10 On-Prem AI Architecture: Keep Data on Site

> Source: https://netray.co/erp-help/private-ai-layer-ifs-applications-10-on-prem-architecture
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: ERP Help

## Short answer

IFS Applications 10 on-premises runs on an Oracle database behind an IFS middleware server, and it exposes data through logical unit views and APIs, IFS Connect messaging, web services and Aurena projections. A private AI layer reads a read-only Oracle copy, answers with a local open-weight model, and writes back only through supported IFS interfaces after human approval.

## Kind

advanced

## System

ifs

## Erp

IFS Applications 10 (on-prem)

## Module

AI integration / Private architecture

## Query

IFS Applications 10 AI integration on-prem

## Alternate Queries

- how to add AI to IFS Applications on-premises
- IFS private LLM air-gapped
- chat with IFS data Oracle
- IFS Connect OData AI assistant

## Applies To

IFS Applications 9 and 10 on-premises with Oracle. Aurena projection and OData availability depends on update level; check your release notes. IFS Cloud is a different architecture.

## Steps

### Add a private AI layer to IFS Applications 10 in seven steps

**Items:**

- Confirm release and update level, the Oracle edition and licences, and which sites and companies are in scope.
- Create a read-only Oracle copy for AI use: a physical standby opened read-only, or replicated tables or materialized views, after checking licence terms for the chosen feature.
- Build views over logical unit views such as SHOP_ORD, INVENTORY_PART_IN_STOCK, CUSTOMER_ORDER_LINE and PURCHASE_ORDER_LINE_PART, exposing only needed columns.
- Use IFS web services, IFS Connect or Aurena projections for record-level reads and for drafted transactions that a person approves.
- Index documents managed in IFS Document Management and file shares, with their access rules carried into the index.
- Serve an open-weight model with vLLM or Ollama on local GPUs and a gateway that logs prompts, tool calls and answers.
- Run phased rollout: read-only questions, then drafts, then approved actions through supported interfaces.

## Sections

### Data access paths in IFS Applications 10

**Body:**

IFS organizes data in logical units (LUs). Each LU has a table, a view that adds computed and translated columns, and a PL/SQL API package with the business rules. Familiar views include SHOP_ORD for shop orders, INVENTORY_PART_IN_STOCK for stock, CUSTOMER_ORDER_LINE for sales, and PURCHASE_ORDER_LINE_PART for purchasing. For reads, query the views, not the underlying tables, so state names and descriptions are translated correctly. Confirm view and column names in your dictionary.

For record-level access with business logic, IFS offers several routes: web services published from LU APIs, IFS Connect for message-based integration (XML and SOAP, with REST in newer updates), BizAPI in older releases, and Aurena projections that expose OData for the same data the client uses. Which are available depends on update level.

Writes must go through these supported routes, which call the LU APIs and apply validation, state transitions and events. Direct DML on IFS tables breaks state handling and is unsupported.

**Code:**

-- Read-only query on the Oracle copy (verify columns for your update)
SELECT order_no, release_no, sequence_no, part_no, contract, state
FROM   shop_ord
WHERE  contract = 'SITE1'
AND    objstate IN ('Released', 'Started')
FETCH FIRST 50 ROWS ONLY;

SELECT contract, part_no, location_no, qty_onhand, qty_reserved
FROM   inventory_part_in_stock
WHERE  contract = 'SITE1' AND part_no LIKE 'A1%';

### Reference architecture, including air-gapped

**Body:**

Components: Oracle read-only copy; change data capture from redo using Oracle tooling such as GoldenGate or LogMiner based capture, or timestamp-based incremental pulls on views; retrieval index; tool server with named read actions; model server; audit log. The AI stack never connects to the production database.

The tool server exposes fixed, parameterised queries (for example "shop orders late by site") and a few record-level calls to web services or projections. The model chooses a tool and parameters; the gateway validates them and applies row and time limits.

For air-gapped plants, bring weights and images in on approved media, deploy the retrieval index and model server on the same network as IFS, and block all outbound routes. IFS Applications 10 does not need the internet to operate, so the AI layer need not either.

**Code:**

{
  "name": "ifs_late_shop_orders",
  "description": "Shop orders past need date for a site",
  "source": "shop_ord",
  "params": {"contract": "string", "days_late_min": "int"},
  "max_rows": 100,
  "mode": "read-only"
}

### Permissions and write-back

**Body:**

IFS controls access through permission sets, per-user allowed sites and companies, and data access rules in some areas. The assistant must reflect these. Resolve the asker to an IFS user, then apply their allowed sites and companies as mandatory filters on every query against the copy. For anything the copy cannot express, call a web service or projection under that user so IFS evaluates the rights itself.

Write-back should be limited and approved: the assistant drafts a purchase requisition, a shop order operation report or a service request through a supported interface, shows the payload, and submits it after a named approver. The IFS audit trail then shows a person, and state-machine rules still apply.

### Sizing and phased rollout

**Body:**

For 100 to 1,000 named users plan one or two data center GPUs for a 14B to 32B class model, and two 80 GB GPUs if you want a 70B class model for harder reasoning. Document indexes for maintenance manuals, drawings and procedures commonly run tens to low hundreds of gigabytes. Treat these as ranges to validate.

Phase 1: read-only questions on shop orders, stock, purchasing and service orders. Phase 2: drafts, such as shortage notes, maintenance summaries and supplier emails. Phase 3: approved actions through supported interfaces for a narrow set of transactions, with approvers and rollback rules.

### Honest limits

**Body:**

IFS includes its own search, reporting and, in newer products, AI features. If you are moving to IFS Cloud, use the vendor path for new capabilities. For Applications 10 on-prem, a private layer is justified when you need cross-module questions, documents, or strict data residency.

Licensing for standby databases and replication features is your Oracle contract matter. Customisations, custom fields and custom LUs need mapping, and totals should come from SQL rather than the model. Do not assume OData projections exist on every update level.

## Pitfalls

- Querying base tables instead of IFS views, losing state translations and computed columns.
- Running AI queries against the production Oracle instance during peak load.
- Writing with DML instead of supported APIs, bypassing state machines and events.
- Ignoring allowed-site and company rules so users see other sites.
- Assuming IFS Cloud interfaces exist in Applications 10.
- Forgetting to check Oracle licensing for standby or replication features.

## Ai Angle

ERPray can connect read-only to ERP data, show the query behind answers and respect roles, with on-prem options. For IFS Applications estates, Netray builds the Oracle copy views, tool server and approval flow with your IFS partner, using the interfaces your integrations already use.

## Faq

### Can I add AI to IFS Applications 10 without the cloud?

**Answer:** Yes. Keep Oracle and IFS where they are, create a read-only copy, serve an open-weight model on local GPUs, and run the gateway inside your network. An air-gapped design has no outbound route.

### Should I query IFS tables or views?

**Answer:** Query the logical unit views, such as SHOP_ORD or INVENTORY_PART_IN_STOCK, because they add computed columns and translated state names. Use web services or projections for record-level reads that need business logic.

### How do IFS permissions carry over?

**Answer:** Map the asker to an IFS user, apply their allowed sites and companies as filters, and call web services or projections under their identity for anything views cannot express.

### Is this the same as IFS Cloud?

**Answer:** No. IFS Cloud has a different architecture and vendor-provided AI path. This design targets IFS Applications 9 and 10 on-premises on Oracle.

## Search terms

IFS Applications 10 AI integration, IFS on-prem AI, IFS private LLM, IFS Oracle read replica AI, IFS Connect AI assistant, chat with IFS data, IFS air-gapped AI, IFS Applications 10 OData

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json