How to add an AI chatbot to SAP Business ByDesign
Short answer
Build the chatbot as a separate web app that queries ByDesign through scoped OData or web service calls, answers only from what those calls return, and appears to users through a mashup, a portal page or a link. It can run on an open-weight model you host yourself.
A chatbot for ByDesign usually serves one of two audiences. Employees want to ask plain questions: what is the status of this order, who approved that request, what stock do we hold. Customers and suppliers want self-service: where is my delivery, what is the status of my invoice.
These two are very different engineering problems. The internal assistant can inherit employee identity and see a lot. The external one must never see more than the single party's own data, which means the identity and filtering rules matter more than the model.
This page covers how to ground a chatbot in ByDesign data, how users reach it, and how to keep answers accurate and bounded.
What teams use it for
Internal order desk assistant
Customer service staff ask for order status, delivery dates and open items instead of switching between work center views.
Procurement Q and A
Buyers ask which requests are pending, which suppliers have late orders and what the last price paid was.
Supplier self-service
Suppliers authenticate and ask about the payment status of their invoices and open orders, filtered to their own account.
Customer delivery status
A portal assistant answers 'when will my order arrive' using the shipment and delivery data of that customer only.
Policy and process help
A second retrieval source holds your approval policies and process documents so the bot answers how-to questions with citations.
SAP Business ByDesign integration surface
- Custom OData services for read queries, with filters that include the requesting party so external users are limited to their own records.
- SOAP web services through communication arrangements for lookups not covered by OData.
- Mashups and embedded URL components in ByDesign's UI, which let you place an external web app inside a work center view or a home page tile; availability and setup vary by release.
- Customer and supplier portals or your own website as the front end for external audiences, with ByDesign accessed only from your back end.
- Business users and roles for internal identity; an identity provider you control for external identity, mapped to business partner IDs.
How it works
- 1
Decide audience and boundary
Write down who may ask what. Internal and external bots should be separate deployments with separate credentials.
- 2
Build the query tools
Create a small, fixed set of backend functions for status, open items and stock. Parameters that identify the party come from the session, never from the model.
- 3
Add retrieval for documents
Index policy PDFs and process notes in a vector store with metadata for audience, so internal documents never appear in external answers.
- 4
Connect the model
Use tool calling so the model chooses a function and receives structured results, then writes an answer limited to those results.
- 5
Embed or link it
Place it in a mashup or embedded component for employees, or on your portal for external users, with single sign-on in front.
- 6
Add feedback and handoff
Each answer shows its sources and has a thumbs up or down. Offer a handoff to a named person for anything the bot cannot resolve.
Example flows
Supplier asks about payment
Trigger: A logged-in supplier types, 'When will invoice 4471 be paid?'
- The session supplies the supplier's business partner ID.
- The invoice tool queries OData with that ID and invoice number.
- It returns status, due date and any block reason.
- The bot answers in plain language and offers to email accounts payable.
Result: A routine query answered without a phone call and without exposing any other supplier's data.
Employee checks stock before promising a date
Trigger: A salesperson asks, 'Do we have 200 of material M-1180 in the main warehouse?'
- The model calls the stock tool with material and site.
- The tool returns on-hand, reserved and expected receipts.
- The bot states available quantity and the next receipt date.
- It notes the time the data was fetched.
Result: A date promised with confidence, with the data source visible.
Keeping data on-prem
The chatbot service, the document index and the language model can all sit in your own network even though ByDesign is SAP-hosted. A common design puts an open-weight model on local GPUs, a vector store for policy documents, and a back end that calls ByDesign outbound. Employees reach it through an embedded component or an internal URL; external users reach a thin portal layer in a DMZ that holds no ERP credentials. This keeps prompts, retrieved passages and conversation logs under your control, which matters for customers with export-controlled or contractual restrictions.
Security checklist
- Derive the party identity from authenticated session data and apply it as a mandatory filter in every query.
- Keep internal and external bots on separate deployments, credentials and document indexes.
- Log questions, tool calls and answers, and set a retention period that matches your policy.
- Mask or avoid personal data in prompts and avoid storing full conversations longer than needed.
- Test with adversarial prompts that try to read other parties' records or ignore instructions.
Ways to get it done
Use ByDesign's built-in search and assistant features
Effort: Low.
Enough if your users only need navigation and standard search; check what your release offers.
Build a chatbot in-house over OData
Effort: Three to six weeks including testing.
Fits teams with a web developer and a clear internal audience.
Netray-built, on-prem capable assistant
Effort: A scoped pilot of a few weeks, then extension.
Fits teams needing private models, supplier or customer self-service, and audit.
Pitfalls
- !Letting the model supply the customer or supplier ID instead of reading it from the session.
- !Answering from general knowledge when a lookup returns nothing; require an explicit 'not found' reply.
- !Mixing internal policy documents into an external bot's index.
- !Embedding a third-party page in the UI without checking cross-origin and framing settings.
- !Launching to customers before a month of internal use has exposed odd questions.
Frequently asked questions
Can the chatbot live inside the ByDesign screen?
Often yes, through mashups or embedded URL components, depending on your release and configuration. If not, a link or tile that opens the assistant in a browser tab works and is simpler to maintain.
How do we stop it exposing one customer's data to another?
Take the identity from the authenticated session and apply it as a mandatory filter in the back end, never as a model-chosen parameter. Test this specifically with prompts that try to override it.
Does it need a large model?
Not necessarily. Status and lookup questions are well handled by modest open-weight models with tool calling. Larger models help with messy phrasing and long documents. Evaluate on your own questions.
Can it take actions, like creating a service request?
It can, but start with read-only. When you add actions, create drafts or tickets that a person reviews, and confirm with the user before submitting anything.
Will it work in other languages?
Many current models handle common business languages reasonably well, but test with your actual terminology and ByDesign status texts, which may be localised differently from what a model expects.
More for SAP Business ByDesign
Want AI chatbot working on your SAP Business ByDesign?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment