Service 03 · AI & Automation

AI where your people already work.

Integration of RAG systems and LLM agents into your business applications. Semantic document search and automated data preparation, with full control over write actions.


Live process

Process control with human-in-the-loop.

Example: incoming invoice. The AI reads, checks and prepares. Nothing is posted until a person approves – and every step is logged.

  1. 01
    ReceiptDocument arrives from DMS, email or scanner.
  2. 02
    ExtractionFields are read, each with a confidence score.
  3. 03
    Validation rulesPurchase order match, amounts, duplicates. Discrepancies are flagged.
  4. 04
    ApprovalA person sees the actual system state and decides.
  5. 05
    Posting & logOnly now does the system write – traceably.
agent · incoming invoicesYour environment · logging on
09:41:02ReceiptInvoice_4711.pdf · 2 pages · from DMS
09:41:03ExtractionSupplier: Müller GmbH0.99
09:41:03ExtractionAmount: €12,480.000.98 · Due: 20 Nov0.97
09:41:03ExtractionIBAN: DE89 …30000.81 → human
09:41:04Rule checkPurchase order 4500012 found
09:41:04Rule checkAmount ≠ purchase order (+3.1%) → flagged
Approval required
SupplierMüller GmbHAmount€12,480.00 (purchase order €12,105.00)IBANplease checkPurchase order4500012
ApproveReturn to supplier
Waiting for: Accounts payable · only a human has write access
09:41:21Approvalgranted by M. Berger · comment: back-order delivery, amount OK
09:41:22PostingERP document 1900003 created · log #a81f saved
09:41:22DoneNext document waiting
BEFORE · WITHOUT AGENT
AFTER · WITH AGENT
Invoices are keyed in field by field.
Fields are read, each with a confidence score. Uncertain values go to a person.
Discrepancies with the purchase order surface in the payment run.
Discrepancies are flagged on receipt, before anything is posted.
Approval by email, system state unknown.
The approval card shows the actual system state. One click, logged.
Who decided what and when is buried in email threads.
Every step logged – traceable and audit-proof.

What we build

AI integrations for your system landscape.

AI functions as decoupled microservices, directly within your core processes.

Document search (RAG)

Semantic search across contracts, manuals and case files – with precise source references and a direct jump to the passage.

  • Semantic querying of unstructured data sets and ticket histories.
  • Strict inheritance of existing user permissions from your source system.
  • Operated either locally (on-premises) or on EU-hosted instances.

Structured data extraction

Automated extraction from unstructured documents into defined ERP, CRM or DMS fields.

  • Field-level extraction from invoices, delivery notes and forms.
  • Rule-based confidence checks: uncertain values are flagged for review.

Agents with an approval process

Automated preparation of complex cases, system queries and drafts.

  • Multi-step preparation of requests and system data.
  • No write action in the live system without explicit approval (human-in-the-loop).

Vector data layer & integration

Connecting existing data silos to a central vector database as the basis for search and agents.

  • Integration of ERP, CRM and DMS with an optimised vector database (e.g. PostgreSQL/pgvector).
  • Efficient delta updates instead of compute-intensive full loads.

High-performance AI & multimodal systems

Real-time interaction and visual AI interfaces.

Development of real-time AI systems for complex interaction requirements – from multimodal voice agents to photorealistic video interfaces.

01 · VOICE & VIDEO

Multimodal real-time agents

  • Development of low-latency interaction systems with synchronised voice and video streaming.
  • Direct coupling of generative avatars / lip-sync engines to internal knowledge bases (RAG).
  • Use in kiosk systems, digital citizen services or high-touch service portals.
02 · ON-PREMISE

Custom Model Training & Fine-Tuning

  • Domain-specific fine-tuning of open-source models (Llama, Mistral, vLLM) on your terminology and proprietary data.
  • Runs on your own GPU clusters (on-premises) for full data sovereignty and predictable operating costs.
03 · COMPUTER VISION

Computer vision & multimodal sensor data

  • Real-time processing of video and image streams for quality control, object detection or automated analysis.
  • Seamless transfer of results via structured APIs to your ERP or control system.

PROTOTYPE · ASPIRITS · 2026

Answers only from verified sources

A voice avatar at a city-hall kiosk answers citizens’ questions exclusively from the city’s published pages. 1,131 pages are indexed, and around 6,000 facts were checked one by one against the source text. Off-topic and political questions are declined. Speech output and lip-sync run locally on a single GPU.

Prototype, not a production system.


Data flow & architecture

Decoupled data storage. Hosting of your choice.

Full control over data storage and model hosting. Business logic, vector database and logging remain within your infrastructure.

The model runs on your hardware (Ollama, vLLM). Nothing leaves your network.

Model-agnostic architecture: switch easily between Llama, Qwen, Mistral or proprietary LLMs without code changes.


Governance & Audit

Binding security principles.

Strict system boundaries in software design. Every integration follows four defined principles:

RULE 01

No mutation without approval.

Drafting and preparation are automated. Write actions in the live system always require an approval card.

RULE 02

Mandatory source tracing.

Every system response cites the underlying source document and the exact passage. Without a validated source, there is no output.

RULE 03

Threshold-based validation.

If the statistical recognition certainty (confidence) falls below the defined target, the field is passed to the business unit for manual review.

RULE 04

Strict permission cascading.

The AI system only accesses data that the requesting user is authorised to see in the source system (ERP/CRM/DMS).

EU AI Act: what applies when.

Regulation (EU) 2024/1689, deadlines as amended by the Digital Omnibus (Regulation (EU) 2026/1744). Logging, human oversight and source references are standard for us – not just from the deadline.

2 Aug 2026Transparency obligations (Art. 50): users are informed that they are interacting with AI.Our approach: labelling in the interface and a notice in every agent response.
2 Dec 2026Labelling of AI-generated content also for systems placed on the market before August 2026.Our approach: generated texts and documents are marked in the metadata.
2 Dec 2027High-risk systems under Annex III (e.g. recruitment, credit scoring, critical infrastructure).Our approach: risk documentation, a log of every action, human approval.
2 Aug 2028High-risk AI in regulated products (Annex I: machinery, medical devices, vehicles).Our approach: documented together with the Machinery Regulation and the CRA.

Contact

Have your requirements checked.

Talk to us directly about your project.