How to add an AI chatbot to DELMIAworks (IQMS)
Short answer
An AI chatbot for DELMIAworks answers questions in plain language from live ERP data and plant documents, inside a web page, Teams or a customer portal. Build it on curated read-only views, add retrieval over SOPs, and run the model on your own servers if data must stay on-prem.
DELMIAworks screens are dense. A supervisor wanting to know whether a job will ship Friday may need the order form, the work order, the press schedule and the material view. A chatbot collapses that into one question, and it can also serve people who never log into the ERP: sales, customers and executives.
The difference between a chatbot that earns trust and one that gets switched off is grounding. Every answer must come from a query or document the user can see, and the bot must say when it does not know.
Here is how to design that for an IQMS-based plant, including the on-prem path.
What teams use it for
Supervisor job questions
Ask 'Will the housing job on press 9 finish by Thursday?' and get status, quantity remaining, cycle rate and a clear answer.
Customer order portal assistant
Customers ask where their order is and see ship dates and tracking drawn from their own orders only.
Sales quick answers
Reps ask about stock, lead time and last price for an item without calling customer service.
Operator SOP help
An operator asks how to do a color change on a given press and gets the controlled work instruction with a link.
Buyer questions on suppliers
Ask for on-time delivery history and open lines for a supplier, with the PO list attached.
Executive daily brief
A scheduled message summarizing shipments, late orders and scrap by plant, with the ability to ask follow-ups.
DELMIAworks (IQMS) integration surface
- Oracle read views: the main source of grounded answers for orders, inventory, purchasing, work orders and MES events.
- Vendor web services or a customer portal module: where available, these provide a supported path for external users to see their own data.
- Document sources: controlled SOPs, work instructions and quality procedures from file shares or a document system, indexed for retrieval.
- Authentication: Active Directory or single sign-on for staff, and a separate portal identity for customers, mapped to customer IDs.
- Exact tables and modules vary by version and licensed options, so map them for your site.
How it works
- 1
Define who asks what
List the top thirty questions by role. Staff, customers and executives need different data and different tone.
- 2
Build the data tools
Wrap read-only views in parameterized functions the chatbot can call. Avoid letting the model write SQL.
- 3
Index documents
Chunk SOPs and procedures, store embeddings in an on-prem vector store, and keep document version and approval status as metadata.
- 4
Compose the bot
A router decides between data tools and document retrieval, and the answer always shows source tables or document titles.
- 5
Apply identity and scope
Pass the user's role into every tool so customers see only their orders and staff see their plants.
- 6
Choose the surface
Embed a chat panel in an intranet page, a customer portal or Teams, rather than modifying the ERP client.
- 7
Measure and tune
Track unanswered questions and thumbs-down answers weekly and fix the tools behind them.
Example flows
Customer asks about a shipment
Trigger: Customer types 'When will PO 55102 ship?' in the portal.
- Bot identifies the customer from login.
- It calls the order tool with the PO filtered to that customer.
- It reads promise date, quantity shipped and the next planned ship.
- It answers with the date and offers to notify on change.
Result: Customer service receives fewer status calls and the customer sees only their own data.
Operator asks for a procedure
Trigger: Operator on press 6 asks how to purge before a color change.
- Bot searches the approved SOP index for the press and process.
- It returns the current approved procedure step by step.
- It links the full document and shows revision date.
Result: The operator follows the controlled instruction rather than an outdated printout.
Keeping data on-prem
A chatbot for an on-prem DELMIAworks plant can be fully self-contained: an open-weight model served with vLLM or Ollama on a GPU server, an embedding model and vector store on the same network, and tools that connect to Oracle read views. Nothing needs to leave the plant. For customer-facing access, publish only the chat front end through a DMZ or reverse proxy, with the model and database behind it, and limit tools to the customer-scoped ones. Smaller models are often adequate for status lookups, with larger ones reserved for document questions.
Security checklist
- Resolve every user to a role or customer ID and apply it in each tool query.
- Keep customer-facing and internal tools in separate toolsets.
- Show sources in answers and refuse to answer when no source supports it.
- Filter out cost, margin and supplier pricing from non-finance roles.
- Log conversations with retention rules, and tell users they are logged.
Ways to get it done
Vendor portal or built-in features
Effort: Configuration only.
If your release has a customer portal and basic alerts, they may meet simple status needs.
Build in-house
Effort: A month or more for a reliable first version.
Possible with a developer who knows retrieval and your schema.
Partner such as Netray
Effort: A few weeks to a pilot.
For on-prem models, role scoping and evaluation handled together.
Pitfalls
- !Letting the bot answer from general knowledge when ERP data is missing, producing confident wrong dates.
- !Exposing internal fields in a customer-facing bot because the tools were shared.
- !Indexing superseded SOPs and returning obsolete procedures.
- !No fallback to a human, so frustrated users abandon the bot.
- !Skipping latency work: answers over thirty seconds will not be used.
Frequently asked questions
Can the chatbot live inside the DELMIAworks screen?
Embedding in the desktop client is usually impractical. Place the chat in a web page, Teams or a portal next to the ERP, and deep link into DELMIAworks records where your setup allows it.
How do I stop it making things up?
Ground every answer in a tool result or document passage, display the source, and instruct the bot to say it does not know when none is found. Test with questions that have no answer.
Can customers use it?
Yes, with a separate customer-scoped toolset, authenticated identity and strict filtering by customer ID. Keep it in a separate deployment from internal tools.
What hardware does an on-prem bot need?
A single modern GPU server can run a mid-size open-weight model for a plant's chat load, and smaller models run on less. Size it from concurrent users and acceptable response time, then test with real questions.
Will it work in other languages?
Many open models handle major languages, but quality varies. Test your languages on real plant vocabulary, such as part descriptions and defect names, before committing.
More for DELMIAworks (IQMS)
Want AI chatbot working on your DELMIAworks (IQMS)?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment