Local AI or Cloud for SAP Business One
Local model, hosting in Germany, your own API key or SAP BTP: data flows, hardware, costs and limits of the four operating models compared.


A supplier sends an invoice as a PDF. The AI is supposed to recognize the supplier, amount, tax, cost center and purchase order reference and prepare the document for SAP Business One. This is exactly where the question of local AI or cloud becomes concrete: may the document leave your network? Does it even have to? And which solution really saves work instead of creating a new technical construction site?
The short answer: there is no overall winner, but four operating models with clear strengths. Local gives you the greatest data sovereignty and the most operating effort. A model hosted in Germany is the pragmatic middle ground for many SMEs. Your own API key gives you the strongest models, and SAP’s own AI always runs via SAP BTP. This article compares the four options, states the hardware requirements of local models and shows how to decide.
Four operating models at a glance
| SAP AI via SAP BTP | Hosted in Germany (RC.AI) | Your own API key | Local in your own network | |
|---|---|---|---|---|
| Where the data is processed | SAP AI Core, region selectable during setup | exclusively in Germany | at the model provider, depending on the contract | in your network |
| Data processing agreement | with SAP | with RConsult, included | with the model provider | not required for the model |
| Your own operating effort | set up and monitor the BTP account | low | manage key, costs and settings | hardware, updates, monitoring, backup |
| Running costs | usage in tokens | 15 EUR per user and month | usage with the provider | electricity, maintenance, hardware replacement |
| Invoice PDFs in RC.AI | not relevant | currently no | yes, with Anthropic Claude | currently no |
| Fits if | you use the Web Client and only need questions and queries | you want to start quickly and expect processing in Germany | you already have contracts with a provider | data must not leave the building |
In all variants with RC.AI, the RC.AI backend runs on or next to your SAP server. The only difference is where the language model does its computing. Permissions always come from SAP: every user works under their own SAP login and only sees what they are allowed to see in SAP.
Which data may the AI see?
The right architecture does not start with choosing a model, but with a sober classification of your data. An AI that suggests texts for internal knowledge articles does not need access to financial accounting. An assistant for incoming invoices, on the other hand, needs document data, supplier master data, account assignment rules and purchase orders.
For general texts, translations or drafts without confidential content, a cloud service is often sufficient. For processes involving financial data, payroll, prices or customer agreements, you set up stricter rules. That does not mean everything has to run locally. It means access rights and data paths have to match the risk. Anyone who may only read purchase orders should not be able to trigger payments or change master data through the AI.
Cloud: start quickly, but do not transfer blindly
Cloud AI is attractive because getting started is quick. You need no graphics card, no model installation and no capacity planning. The providers’ large models are usually stronger at open, complex tasks than anything that can be run economically on premises.
The catch is often not the model but the integration. If employees copy documents into a chat window, store results in Excel and transfer values to SAP by hand, all you get is a new isolated solution. Also check storage location, retention periods and logging and whether the provider uses your data for training.
Calculate the running costs beforehand. With few requests, usage-based models are convenient, but with many documents and users they can become hard to predict. A fixed price per user keeps the calculation simple: with RC.AI, 10 users with the model hosted in Germany cost 150 EUR per month, which is 1,800 EUR per year.
Local: what hardware you need
A local model mainly needs graphics memory. As a rule of thumb: the size of the model file plus 10 to 30 percent for context and runtime, more for long documents. The file sizes below come from the Ollama model library and apply to the usual 4-bit quantization.
| Model class | Example | File size | Realistic graphics memory |
|---|---|---|---|
| 8 billion parameters | Llama 3.1 8B, Qwen3 8B | 4.9 to 5.2 GB | 8 to 12 GB |
| around 20 billion | gpt-oss-20b | 14 GB | 16 GB, sufficient according to the manufacturer |
| 24 to 32 billion | Mistral Small 3.2 24B, Qwen3 32B | 15 to 20 GB | 24 GB, more with long context |
| around 70 billion | Llama 3.3 70B | 43 GB | 48 GB, for example two cards with 24 GB each |
| around 120 billion | gpt-oss-120b | 65 GB | one card with 80 GB according to the manufacturer |
Models such as gpt-oss and Qwen3 30B-A3B only compute with part of their parameters per request. They still need memory for all weights, but respond faster than models of the same size that always use all parameters. As a runtime, Ollama is the easiest way to get started, and RC.AI connects local models through it. For many concurrent users, vLLM is the more powerful choice. Check the model’s license before use: gpt-oss is licensed under Apache 2.0, Llama under Meta’s own license.
For clearly defined tasks such as questions about inventory, open items or sales orders, a model in the 20 to 32 billion class is often enough. Bigger is not automatically better. What matters in daily work is whether the answers are correct, traceable and fast enough. So test with real questions from your business before you buy hardware.
Local means operations
With a local model, you take on an operational task. Hardware, updates, access protection, backup and monitoring all have to work. You should be able to trace which model version was active when and who approves changes. The AI system only needs access to the services it requires for its purpose, not to the entire network.
Local does not automatically mean secure. A locally operated model with overly broad SAP authorizations remains a risk. And there is a limit you should know about: RC.AI’s PDF analysis currently requires Anthropic Claude. If you want to process incoming invoices with AI, you therefore cannot run fully locally today.
Graduated instead of black or white
You do not have to choose the same path for all tasks. Sensitive queries about payroll or terms can run locally, drafts and general texts via a cloud service. What matters is that the boundaries are technically enforced and not just written down in a work instruction. Data minimization is part of this from the start: the AI only receives the data it needs for the respective task.
Your own API key is also a deliberate choice. You keep the contract, billing and settings in your own hands. In return, you have to store the key securely, separate access and define how it is rotated or revoked. Your own key does not replace a clear data architecture.
How to make the decision
| Criterion | Speaks for local | Speaks for hosting in Germany | Speaks for your own key |
|---|---|---|---|
| Data must not leave the building | yes | no | no |
| In-house IT available for operations | required | not required | partly |
| Incoming invoices from PDF | currently no | currently no | yes |
| Predictable costs per user | upfront investment | yes | no, usage-based |
| Strength with open, complex questions | limited | good | highest |
Start with a process that visibly costs time today and define which data it needs. The operating model usually follows from that. Which use cases pay off is shown in the article AI in SAP Business One: Features and Use Cases. How to clarify the contract, the AI Act and approvals before the start is covered in AI in ERP: Data Protection, AI Act and Getting Started. You can find the RC.AI options on the page for the RC.AI assistant.
Frequently asked questions
What hardware does a local language model need?
Can a local model process incoming invoices from PDFs?
Is a local AI automatically GDPR-compliant?
What does the model hosted in Germany cost with RC.AI?
Can I run SAP's AI features locally?
Updated on November 6, 2026

