Embedded Files
SAP S4/HANA AI based automation in Sales and Service
SAP S4/HANA AI based automation in Sales and Service
YUnit is collaborating with the Andorran AI Software partner EO Nexus Technologies in establishing sales and service related knowledge and interaction processes.
Enterprise software has a usability problem that we’ve all accepted as normal. SAP S/4HANA is extraordinarily powerful — it runs the supply chains, finance, and operations of thousands of global businesses. But interacting with it has always required training, patience, and a tolerance for complexity.
We set out to change that. What started as an internal proof-of-concept has become a production-ready architecture that lets customers manage their orders through natural conversation — while the ERP stays completely intact beneath the surface.
Anyone who has tried to modify a Sales Order in SAP knows the experience. You need to know the transaction code, the document type, the business partner number, the material number, the sales
organisation. Miss one field and the system rejects your input with an unhelpful error. Customer service teams spend hours on tasks that should take minutes. Supervisors monitor dashboards
that require expert interpretation. And all of this friction exists not because the underlying data is wrong — but because the interface was never designed for conversation.
A Different Approach: The AI Agent
At the heart of our solution is an AI agent powered by a LLM such as Claude (Anthropic). Rather than replacing SAP, the agent wraps it: it sits between the user and the ERP, understanding intent in plain language and
translating it into precisely the right system operations.
Claude doesn’t connect directly to S/4HANA. Instead, it works through a set of well-defined tools — structured operations like create_sales_order, update_sales_order, get_order_status, cancel_sales_order, and add_payment_card. Each tool maps to a secure, validated integration flow that sits between the agent and the ERP.
“Claude handles intent, confirmation, and natural-language response. It never touches the ERP directly — every action goes through a validated integration layer that acts as the guardrail.”
Results in Production
The results have been exactly what we hoped for:
✓ Order creation time reduced from minutes of screen navigation to a 30-second conversation.
✓ Support ticket volume for order modifications dropped significantly — customers now self-serve what previously required a human agent.
✓ Order error rates fell because the agent validates every field before submission, resolving common mistakes automatically.
✓ Supervisor oversight improved through the approval workflow — high-value orders get a human review without slowing down routine ones.
Why should I not just use SAP's own AI offering?
While SAP Build and Joule Premium has made breathtaking steps forward, and it's possible to create custom agents by this year, there are downsides:
Complexity to integrate with 3rd party applications
Dependency on SAP's AI Unit pricing
Limited customisation options
Agents can only run in SAP BTP
Dependency on AI Tools Joule supports
No control over which master data is exposed to SAP AI
When to build a custom AI integration scenario?
The advantage of using your own agent implementation is that you can leverage your own AI tool investments and reuse enterprise subscriptions that you might already have. Also you can change between AI tools e.g. Claude, Gemini and ChatGPT as you desire. While the backend calls still need to be forwarded via the Integration Suite, you're free to deploy the agent application itself to SAP BTP or any other IaaS or PaaS provider. And you can modify the custom agent application and MCP server completely to your needs without any constraints.
Built for the Real World
The architecture is built to production standards throughout:
✓ The agent is made to run as a containerised service on AWS ECS or SAP's own Kyma environment — scalable, isolated, and independently deployable.
✓ Integration flows handle authentication, error recovery, and message logging with configurable alerting when failure rates spike.
✓ An approval workflow layer in SAP Integration Suite lets supervisors review and sign off on high-value or unusual orders before they are submitted to the ERP.
✓ The entire solution deploys on top of existing SAP infrastructure — no S/4HANA customisation required, no BTP AI Core dependency, no proprietary AI lock-in.
The Intelligence Behind the Agent
One of the most important design decisions was ensuring the agent always works with accurate, current data — not stale snapshots. This required solving a real data synchronisation challenge.
Always-fresh master data
Before Claude can create a valid Sales Order, it needs to know which customers exist, which materials are available for each sales organisation, current prices, and live stock levels. We replicate all of this from S/4HANA into the agent’s own data store in real time.
The replication is event-driven: whenever a customer record, material, pricing condition, or stock level changes in S/4HANA, an event is published to a message bus. An integration flow picks this up immediately, fetches the complete updated record, and pushes it into the agent’s context. The agent is never working from yesterday’s data.
Secure data
At the same time the master data is not fed into an AI tool directly. It's stored in memory as part of the agent application which the enterprise controls. Master data is fed by the agent into the chat context selectively only to perform necessary validations which are part of the business logic.
Validation before every action
The integration layer between the agent and S/4HANA acts as a guardrail. Every inbound request is validated: required fields are checked, idempotency is enforced (so duplicate requests never createduplicate orders), and missing fields like delivering plant or distribution channel are automatically resolved from the master data the agent already holds.
What the Agent Can Do
The agent supports the full lifecycle of a Sales Order. A customer can walk through any of the following
simply by describing what they want:
• Create a new order — the agent extracts customer, materials, quantities, and delivery date from
the conversation, confirms the details, and creates the order in S/4HANA.
• Modify existing orders — change quantities, add new line items, update delivery dates, or
remove positions from an open order.
• Cancel an order — with a built-in confirmation gate: the agent requires explicit confirmation
before any cancellation, and warns that the action is irreversible.
• Adjust individual item quantities — with an automatic warning if the requested change exceeds
50% of the original quantity.
• Add a payment card — securely, using tokenisation. The agent never requests or stores full card
numbers.
• Check order status — retrieving real-time delivery and processing status directly from the ERP.
Every write operation follows the same discipline: the agent always calls get_order_status first to verify
the order is open and editable, confirms all details with the user before acting, and surfaces any errors
clearly rather than swallowing them silently.
Like to know more or see the demo without any commitment? Please contact us at ingo.taraske@yunit.tech.
Page updated
Report abuse