How to add AI agents to Infor LN
Short answer
Add AI agents to Infor LN by giving a model a small set of tools backed by ION API, BODs and a read replica of the LN database, then wrapping any write in a human approval step. Start with read-only exception triage, prove the answers against LN sessions, and only then allow drafted transactions.
Infor LN, the descendant of Baan, runs a lot of complex discrete, project and engineer-to-order manufacturing. The data model is deep: item master, enterprise planning, project control, warehousing and service all share one company-numbered schema. Planners spend their day in sessions like planned orders, order planning exceptions and purchase order lines, chasing the same few kinds of problem.
An AI agent is not a chatbot. It is a loop: observe an LN event or exception, call tools to gather context, decide, and either report or propose a change. On LN the value is in the repetitive judgement work: which late purchase orders actually threaten a production order, which MRP messages can be ignored, which receipts are missing inspection.
The catch is that LN is often on-prem, heavily customised, and owned by a small team that does not want an experiment writing into live tables. So the design below keeps the agent outside LN, talks to it only through supported interfaces, and treats every write as a draft until a person approves it.
What teams use it for
Late purchase order triage
Each morning the agent reads open purchase order lines past or near their planned receipt date, joins them to the production orders and project activities they feed, and ranks which delays matter. It drafts supplier chase emails for the top few.
Order planning exception review
It reads planning exceptions, groups them by cause (supplier lead time, forecast change, BOM change), and suggests which to action, defer or ignore, with the reasoning shown for the planner to accept.
Draft nonconformance and inspection follow-up
When a receipt is blocked in quality, the agent gathers the item, supplier history and prior rejections and drafts a nonconformance summary for the quality engineer to submit.
Project cost watch
For engineer-to-order projects, the agent compares committed cost and actuals against budget by project element on a schedule and writes a short variance note when a threshold is crossed.
Sales order promise checking
Before a salesperson confirms a date, the agent checks on-hand, planned supply and open demand and explains whether the promise is realistic and what would have to move.
Master data hygiene sweeps
It scans for items with missing planning parameters, zero lead times or inactive suppliers still on approved lists, and raises a worklist rather than changing records.
Infor LN integration surface
- ION API gateway (Infor OS): the supported front door for REST calls and for exposing LN business logic to external callers. Available services and versions depend on your LN release and Infor OS setup, so confirm which are enabled.
- BODs through ION: LN publishes and consumes Business Object Documents such as sales orders, purchase orders and item masters. Good for event triggers (a PO changed) and for controlled writes, with document flows visible in ION Desk.
- LN business object layer (the DAL and business interfaces): code-level entry points that enforce LN rules on writes. Prefer these over direct table updates, which bypass validation.
- Relational database (typically SQL Server or Oracle; some estates still DB2): tables are named by package, module and number plus the company number, for example tdpur400 and tdsls400 style names with a company suffix. A read replica is the safest source for agent analytics.
- Infor Data Lake and Data Fabric, where licensed: replicated LN objects queryable with SQL, which keeps heavy agent reads off the transactional system.
How it works
- 1
Pick two or three read-only jobs
Choose workflows with clear ground truth, such as late PO triage. Write the questions a planner would ask and the LN sessions that currently answer them.
- 2
Build a read layer
Expose parameterised queries over a read replica or the Data Lake, one per question, rather than letting the model write free SQL against the live schema. Map company numbers explicitly.
- 3
Wrap queries as tools
Give each query a name, typed inputs and a short description. Add a document-lookup tool for specs and supplier files if relevant. Keep the toolset small so the model chooses well.
- 4
Add an orchestration loop with limits
Run the model with a step cap, a token cap and logging of every tool call and result. Use a model served on your own GPUs if data must not leave the site.
- 5
Evaluate against planners
Replay a few weeks of historical exceptions and compare the agent's call with what planners actually did. Fix tool gaps before widening scope.
- 6
Introduce drafted writes
For actions like a PO date change, have the agent produce a proposed BOD or business-object call, show it with the evidence, and route it to a named approver. Execute only after approval under that user's authorisation.
- 7
Operate it
Monitor tool errors, disagreement rate with planners, and cost per run. Review the audit trail periodically with the LN administrators.
Example flows
Morning late-supply briefing
Trigger: Scheduled run at the start of the planning shift.
- Agent queries open purchase order lines with a planned receipt in the past or the next few days.
- For each, it finds the production orders and project activities that consume the item.
- It checks whether on-hand or other supply covers the need.
- It ranks the lines by exposure and drafts a chase message for the top ones.
- It posts the briefing to the planner's channel with links to the LN sessions.
Result: The planner starts with a short, ranked list instead of scanning a full exception grid.
Approved reschedule
Trigger: A planner clicks 'propose fix' on a late-supply item.
- Agent builds a proposed change to the PO line promise date from the supplier reply.
- It shows the before and after, affected orders and its reasoning.
- An approver accepts or edits the proposal.
- The change is sent through the supported interface as that user.
- Agent re-reads the line to confirm LN accepted it.
Result: A reschedule is made with LN's own validation applied and a full record of who approved what.
Keeping data on-prem
Everything here can run inside your network. LN and its database stay where they are; the agent service sits beside ION or a DMZ-adjacent application server, reads from a replica, and calls a locally hosted open-weight model served through vLLM or Ollama. If ION runs on-prem you can keep event traffic internal, and if you use Infor-hosted cloud the agent can still sit in your own tenancy and reach only the approved ION API endpoints. Air-gapped sites can export snapshots to the agent environment on a schedule and accept the lag.
Security checklist
- Give the agent its own service identity with read-only rights scoped to the companies and tables it needs, never a developer or admin account.
- Execute writes under the approving person's identity so LN authorisations and audit fields stay meaningful.
- Log every prompt, tool call, result and approval to a store the agent cannot edit.
- Redact or exclude price, customer and export-controlled fields the workflow does not need before they reach the model.
- Treat supplier emails and uploaded documents as untrusted input to guard against prompt injection.
Ways to get it done
Infor's own AI and automation features through Infor OS
Effort: Low to moderate; limited by what is productised for your version.
Reasonable if you are on a current cloud-connected stack and your use cases match what Infor ships.
Build in-house on ION API and a read replica
Effort: Typically a few engineer-months for a first useful agent plus ongoing upkeep.
Teams with a strong LN technical lead and software engineers who can maintain tool code.
Partner build, such as Netray with an LN domain partner
Effort: A scoped pilot of several weeks, then incremental workflows.
Sites that want private models and a tested tool layer without staffing it.
Pitfalls
- !Letting the model write free SQL against live tables leads to slow queries, wrong company joins and accidental load on production.
- !Customised sessions and extension tables mean the agent's idea of an 'open order' may differ from yours; validate definitions with planners.
- !Direct table updates skip LN business logic and can corrupt integrity; always use supported write paths.
- !Starting with writes instead of read-only reporting destroys trust after the first wrong update.
- !Ignoring multi-company and multi-site setups produces confident answers about the wrong entity.
Frequently asked questions
Can an AI agent write back into Infor LN?
Yes, but through supported paths such as BODs via ION or LN's business object layer, never by updating tables directly. Start read-only, then add drafted changes that a named approver confirms. The write executes under that person's identity so LN validation and audit still apply.
Does the agent need Infor OS and ION?
Not strictly. A read-only agent can work from a database replica alone. ION API and BODs become important for event triggers and controlled writes, and they are the cleaner route for anything that changes data. Availability depends on your LN version and licences.
Which model should run the agent?
For data that must stay private, an open-weight model class such as Llama, Qwen or gpt-oss class served on your own GPUs is workable for tool-calling workflows. Smaller models need tighter tools and tests. Cloud models are an option where policy allows.
How is this different from a chatbot over LN?
A chatbot answers a question when asked. An agent watches for conditions, runs several steps, and produces a result or proposal without being prompted. Both can share the same tool layer, so many teams build the chat interface first and add scheduled agents second.
How do we keep it from acting on bad data?
Constrain it to named queries, show the evidence for every recommendation, cap steps, and require approval for writes. Measure disagreement with planners during a trial period and tighten tools wherever it diverges, rather than relying on prompt wording alone.
Want AI agents working on your Infor LN?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment