How to add an AI chatbot to Baan IV and Baan V
Short answer
Build a web chat that answers from two sources: live Baan data through read-only queries and your own documentation through retrieval. Host it on a web page or Teams, and run the model on your own hardware if the data is sensitive.
Baan's user interface is a product of its time. New staff struggle to learn session codes, and customers email the order desk for information that sits in the system. A chatbot can take those questions off people, if it answers correctly and says so when it does not know.
The difference between a useful chatbot and a risky one is grounding. It must fetch order or stock facts from Baan at the time of the question, quote documents for procedure answers, and refuse to guess.
This page covers what to build, who to build it for first, and how to keep it inside your network.
What teams use it for
Internal help for session navigation
Ask how to enter a goods receipt or find a production order and get steps drawn from your own procedure documents, naming the session to open.
Order desk assistant
Order status and delivery dates for a customer reference, with the result shown for the agent to confirm before sending.
Customer portal answers
Customers ask where their order is and get a limited answer based on their own account only.
Shop floor questions
Operators ask for the routing step, tooling or work instruction for a job, with documents retrieved from the file share.
New hire onboarding
A guide to your company's way of using Baan, including local codes for order types and reason codes.
Baan IV / Baan V integration surface
- Read-only database access or a replica for facts such as orders, stock and item data.
- Document sources: procedure documents, session help exports and work instructions on file shares, indexed for retrieval.
- Baan users and roles: not directly usable for web login, so map your directory (such as Active Directory) to allowed data scopes.
- ODBC: for tools running on Windows servers beside the chatbot.
- Print-to-file and exchange schemes: for refreshing document sets and reference data on a schedule.
How it works
- 1
Pick the first audience
Start with internal users such as the order desk. External customers add identity and legal complexity.
- 2
Index the documents
Chunk and embed procedures and work instructions into a local vector store with source links.
- 3
Add data tools
Create a few read-only functions for order status, stock and item lookups on the replica.
- 4
Write the answer rules
The bot cites sources, shows the as-of time for data, and says it cannot answer if retrieval finds nothing.
- 5
Serve the model locally if needed
An open-weight model with vLLM or Ollama can run the chat on a GPU server in your network.
- 6
Collect feedback
Add thumbs up or down and review wrong answers weekly to fix documents and tools.
Example flows
How do I receive a purchase order
Trigger: New warehouse clerk asks in the chat.
- Bot retrieves your goods receipt procedure.
- Replies with steps and the session name used at your site.
- Links to the document for screenshots.
Result: Fewer calls to senior staff.
Customer asks where order 4821 is
Trigger: Message on the portal chat.
- Bot checks the logged-in customer's identity.
- Calls the order status tool scoped to that customer.
- Replies with status and promised date, or hands off to a person if the data is unclear.
Result: A quick answer without the order desk.
Keeping data on-prem
Because Baan itself is typically on-prem, a self-contained chatbot is natural: a web front end, a vector store, a read-only replica and an open-weight model all inside your network. Embeddings and the chat model can run on the same GPU server for a modest user count. For external customers, place only the web tier in a DMZ and keep the model and data tier internal, passing narrow requests between them.
Security checklist
- Scope each question to what the user could see in Baan.
- No write actions in version one.
- Treat documents as untrusted input to avoid prompt injection from file content.
- Retain chat logs for review and delete according to policy.
- Separate internal and customer-facing bots, with different tools and data.
Ways to get it done
Generic document chatbot
Effort: Days to a couple of weeks.
Procedure questions only, with no live Baan data.
Custom chatbot with data tools
Effort: A few weeks including testing.
Order desk and planning use where live facts matter.
Partner deployment with Netray
Effort: A scoped pilot.
On-prem model, customer-facing portal or multi-company Baan.
Pitfalls
- !A bot that answers from model memory about Baan rather than your setup.
- !Stale documents describing a retired process.
- !Customer bot with a broader data scope than the customer's own account.
- !Not showing the as-of time, so users trust a replica that is hours behind.
Frequently asked questions
Can a chatbot sit inside Baan?
Not inside the Baan client in a clean way. Most sites run a web page or a Teams bot alongside Baan, with a link or shortcut from the desktop. Embedding in the older client is rarely worth it.
Will it make mistakes?
Yes, sometimes. Grounding with live queries and cited documents reduces errors, and refusal rules help. Plan a review loop and keep humans in the path for customer commitments.
Can customers use it?
Yes, with strict identity checks and narrow tools. Start internal first, then open a limited customer view once you trust the answers and have tested access controls.
Does it need a GPU?
For on-prem open-weight models, yes, usually one or two server GPUs for a small team. Hosted models avoid that but send questions and context outside your network.
Is the vendor's own AI enough?
Baan is out of mainstream support and has no vendor AI. If you are on a later Infor LN release, check what Infor provides first, as it may cover some needs without a build.
Want AI chatbot working on your Baan IV / Baan V?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment