AI in your own software. You decide where it runs.

What AI can be used for, for example

• Document search with RAG

  Semantic search across manuals, contracts or tickets, with an answer and source citation (retrieval-augmented generation).

• AI assistant for internal systems

  Connection to a ticketing system or, via released interfaces, to SAP. SAP & ERP

• Data pipelines

  Extract data from source systems, transform it and load it into target systems (ETL).

• Speech and dialogue applications

  Applications that understand spoken language, respond and are connected to your systems, for example on the phone or at a terminal.

• AI in existing software

  Summarisation, classification or text generation as a feature in your application. Integrating AI agents into production

• Text-to-SQL

  Queries to databases and reports in natural language.

• Inference optimisation

  Optimise inference batching and GPU utilisation, and measure throughput and latency before and after.

Interfaces and modules without AI: Software & Interfaces →

For example: how a document search is structured

1. Your documents
2. Chunking and vector index
3. LLM
4. API
5. Your application

In this example, all stations run in your environment.

Where your data stays

• IN OPERATION

  Depending on the solution, the application, data and models run in your own infrastructure or with a hosting provider of your choice. Whether an external AI service is used is set out upfront in the statement of work.

• DURING DEVELOPMENT

  As a rule, we develop and test with test documents or anonymised data.

More on data handling: How we work →

What AI needs in production

When AI has write access to your systems, these principles apply.

• Rules in code

  The LLM proposes actions. Deterministic application code decides whether they are permitted and executed.

• Approval

  A person approves irreversible actions (human-in-the-loop).

• Permissions

  Identity and permissions come from the authenticated session, not from the model’s response.

• Prompt injection

  External content and tool outputs are treated as untrusted data.

• Result verification

  The target system, not the model’s response, confirms whether an action took effect.

Integrating AI agents into production →Project: 22 external APIs →

Hardware and operation

You receive the hardware requirements before work begins. They depend on model size and number of users. Operation is handled by your team or your hosting provider.

Frequently asked questions

Can an LLM run in our own data centre?

Yes. Open-weight models can run on your own GPU servers. Which one fits depends on the hardware, the licence and the language of your documents.

What is RAG?

Retrieval-augmented generation: the large language model (LLM) answers based on matching passages from your documents and cites the source.

Do documents leave our premises?

Only with your agreement. We clarify beforehand which data is needed for development and operation, where it is processed and whether it needs to be transferred to external services or our partner teams.

How many users can the service handle at once?

That depends on the hardware. The number is agreed as a target value, measured in our test environment and confirmed in your environment through the agreed checks.

Can the service be connected to SAP?

Yes, through a hard-coded connection to published interfaces such as OData services. The SAP API Policy permits AI systems that plan or execute API calls themselves only through SAP-endorsed architectures.

How scope, testing and handover are handled: How we work →