# How to add AI agents to SAP Business ByDesign

> Source: https://netray.co/integrations/sap-bydesign-ai-agents
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Integrations

## Short answer

Build a small tool layer over ByDesign's communication-arrangement web services and custom OData services, let an LLM plan multi-step work against those tools, and route every write through a human approval. Because ByDesign is a cloud-only product, the agent runs beside it, not inside it.

## System

SAP Business ByDesign

## Capability

ai-agents

## Intro

SAP Business ByDesign is the cloud ERP many mid-sized manufacturers, distributors and project businesses chose when they wanted a single suite without running servers. The people who run it day to day, usually a small finance and supply chain team, spend a lot of time on repetitive exception work: purchase requests that stall in approval, customer orders blocked on credit, supplier invoices that fail matching, and stock that sits below the reorder point while nobody notices.

An AI agent is a worker that handles one of those loops end to end: it reads the exception, gathers context from several ByDesign views, drafts the next action, and waits for a person to approve it. That is different from a chatbot that answers questions. It needs a defined set of tools, a plan-and-act loop, and a strict boundary on what it may change.

ByDesign is a tidy target for this because its integration surface is explicit. Nothing is reachable until you create a communication system, a communication arrangement and a technical user, which also means you control exactly what an agent can see.

## Use Cases

### Purchase request chaser

**Description:** The agent lists open purchase requests older than an agreed number of days, finds the approver and the supplier lead time, and drafts a nudge or a reassignment for a buyer to confirm.

### Sales order blocks triage

**Description:** For orders held for credit or availability, the agent pulls the customer's open items and the ATP position, proposes release, partial delivery or a call to the customer, and logs its reasoning.

### Supplier invoice exception worker

**Description:** When an invoice does not match the purchase order or goods receipt within tolerance, the agent compares quantities, prices and receipt dates and drafts a clarification to the supplier.

### Reorder watch

**Description:** The agent reads stock and planning data for items near their reorder level, checks open supply, and prepares purchase requests for a planner to release in one pass.

### Month-end readiness checker

**Description:** Before period close it sweeps for unposted documents, open goods receipts without invoices and unreconciled items, and produces a short list per owner.

## System Surface

SOAP web services published through the Service Explorer and activated per communication scenario in the Application and User Management work center. The scenario you pick decides which operations the agent can call.

Communication systems and communication arrangements: the system record holds the technical user and authentication method (basic authentication or a client certificate), and the arrangement binds it to a scenario. Exact scenario names vary by release.

Custom OData services created from analytics data sources or built with the Cloud Applications Studio. These give read access to business objects shaped the way an agent needs them, with filters and paging.

Cloud Applications Studio (SDK) for custom business objects, actions and ABSL scripts when a write needs validation logic that lives inside the tenant.

SAP Integration Suite or SAP Cloud Integration as an optional mediation layer for monitoring, retries and message mapping.

## How It Works

### Pick three or four narrow jobs

**Detail:** Choose exception loops with a clear owner and a clear definition of done, for example stalled purchase requests. Avoid open-ended goals in the first release.

### Create a dedicated communication system and technical user

**Detail:** Do not reuse an employee login. Give the technical user a business role limited to the work centers the agent needs, and activate only the communication scenarios for those jobs.

### Expose read tools first

**Detail:** Wrap each web service or custom OData query as a tool with a typed input and a trimmed output: get_open_purchase_requests, get_supplier_invoice_status, get_stock_position. Return identifiers the agent can cite.

### Add the planning loop

**Detail:** Run an open-weight or hosted model with tool calling. Give it the exception, the tool list and a rule that it must state which records it used before proposing an action.

### Add write tools behind approval

**Detail:** Create tools such as create_purchase_request_draft that only produce a proposal. A person approves in a queue, and a separate service performs the actual web service call with its own narrower credentials.

### Log every step

**Detail:** Store the input, tool calls, model output, approver and resulting document ID so an auditor can replay any decision.

### Run in shadow mode, then widen

**Detail:** For a few weeks the agent only proposes and the team compares it with what they did anyway. Widen scope only where proposals were accepted without edits.

## Example Flows

### Stalled purchase request

**Trigger:** A daily schedule finds a purchase request waiting for approval longer than the agreed threshold.

**Steps:**

- Agent reads the request lines, requested date and approver through the purchase request service.
- It looks up the supplier's recent delivery performance and any existing open order for the same material.
- It decides whether to remind the approver, merge into an existing order or flag a risk.
- It drafts the message and queues it for the buyer.

**Result:** The buyer approves a ready-made action instead of hunting through lists, and the decision trail is stored.

### Blocked customer order

**Trigger:** A sales order enters a blocked status for credit or availability.

**Steps:**

- Agent fetches the order, the customer's open receivables and the current availability by line.
- It classifies the block and proposes release, split delivery or hold.
- It writes a short rationale and attaches the numbers it used.
- A sales administrator approves or overrides in the queue.

**Result:** Orders are cleared the same day with the reasoning visible to the account manager.

## On Prem

ByDesign itself is SAP-hosted and cannot be installed on your premises, so on-prem here means where the agent and the model run. Keep the agent service, the vector store for any document retrieval and an open-weight model inside your own network or private cloud, and let it call ByDesign outbound over HTTPS using the technical user. Nothing needs to be opened inbound to your site. If policy forbids sending ERP records to a hosted model, the private model removes that concern while ByDesign remains the system of record. Allow-list the agent's egress addresses on the communication system where your tenant supports it.

## Security

- Use a dedicated technical user with a minimal business role; never a named employee account.
- Separate read credentials from write credentials, and keep write execution outside the model process.
- Require human approval for every create, change or release, and log approver identity.
- Mask personal data in prompts when the task does not need it, and keep prompts and outputs in your own storage.
- Rotate the communication system credentials or certificate on a schedule and alert on unusual call volume.

## Options

### Use SAP's own embedded AI features

**Fit:** Good when your need is a standard assistant task already covered in your release. Check what your tenant version includes before building anything.

**Effort:** Low, but limited to what SAP ships.

### Build in-house on web services and custom OData

**Fit:** Fits teams with a developer comfortable with SOAP or OData and an LLM tool-calling framework.

**Effort:** Several weeks for a first read-only agent, longer for governed write paths.

### Work with a partner such as Netray

**Fit:** Fits teams that want private models, approval queues and audit built once and reused across agents.

**Effort:** A scoped pilot of a few weeks, then incremental agents.

## Pitfalls

- Giving the agent a broad business role because it is faster; the role then defines your blast radius.
- Returning entire documents to the model, which wastes tokens and invites mistakes. Trim to the fields the decision needs.
- Assuming a scenario name or operation exists in your release without checking the Service Explorer in your own tenant.
- Letting quarterly upgrades surprise you. Pin tests that call each tool after every release.
- Treating the agent's explanation as proof. Always show the underlying records next to the rationale.

## Faq

### Can an AI agent write directly into SAP Business ByDesign?

**Answer:** Technically yes, through web services that support create and update. In practice you should have the agent propose and a person approve, with a separate service performing the write. This keeps the model away from posting rights and gives you a clean audit trail.

### Do I need Cloud Applications Studio to build agents?

**Answer:** Not for read access. Standard communication scenarios and custom OData services are often enough. You need the SDK when you want custom objects, validations or actions that run inside the tenant, which is useful for safe write paths.

### Can the model run on our own servers if ByDesign is in the cloud?

**Answer:** Yes. The agent and model can sit in your network and call ByDesign outbound over HTTPS. Only the records the tools return are sent to the model, and they stay inside your environment.

### How is an agent different from the SAP assistant features?

**Answer:** SAP's embedded features are built for generic tasks inside its UI. A custom agent can span ByDesign plus other systems, follow your own rules and run on a model you choose, at the cost of building and maintaining it.

### What should the first agent do?

**Answer:** Pick a read-heavy exception loop with a human owner, such as stalled purchase requests or unmatched supplier invoices. It delivers value with proposals only and teaches you how the model behaves on your data.

## Search terms

sap business bydesign ai agents, ai agents for sap bydesign, sap bydesign automation ai, add ai to sap bydesign, sap bydesign web services ai, sap bydesign odata ai agent, sap bydesign llm integration, sap bydesign purchase request automation, sap bydesign agentic ai, sap bydesign private ai

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json