AI integration — from pilot into production
RAG, document AI, computer vision for industry and the mid-market. EU-AI-Act-aware, auditable, in your company's stack.
Many AI projects end as PoC. We deliver AI that reaches production — integrated into your legacy systems, with audit trail and traceable decisions.
When do you need this?
You have a successful AI pilot
(document AI, vision, RAG), but nobody manages the integration into ERP/MES/workflow.
You want to use LLMs
but data protection, auditability and the EU AI Act are unresolved.
Generative AI
should accelerate concrete back-office or production processes — not just present a chatbot on the website.
You need a clear path
from model to productive, monitored application.
What we deliver
RAG systems
for your enterprise data — secure, with source attribution and permission model.
Document AI
for classification, extraction and plausibility checks of PDFs, scans, emails.
Computer vision
for visual inspection, anomaly detection in sensor data, industrial QA.
EU-AI-Act conformity
as standard artefact: model card, data lineage, human oversight, audit logs.
Integration into legacy systems
REST integration to SAP, Microsoft Dynamics, ProAlpha, Camunda workflows.
Choice between cloud and self-hosted
OpenAI, Azure OpenAI in EU region, or Llama/Mistral/Phi on-premises.
Methodology
- 01
Analysis call
60 minutes on the use case — we check with you whether AI really helps here and where the value is realistic.
- 02
Next step
if the scope is clear, you receive an offer. Where the data situation is unclear or the case is regulated, we clarify feasibility, accuracy, latency, cost and compliance (EU AI Act risk class, NIS-2 implications, GDPR data flows) first in a paid concept phase.
- 03
Implementation
once commissioned, we integrate the solution into your existing systems.
- 04
Rollout & operations
go-live with monitoring (drift detection), model card and maintenance plan.
From real projects
AI only works when the system around it does.
GewerbeLotse maps everyday language to official Austrian trade wordings. The classification comes from public sources and domain rules. The language model is one part of the search, not the product.
Map the language
A description in everyday language is matched to official GISA wordings, with the trade key and whether the trade is free, regulated or a partial trade.
Official sources
The basis is GISA/BMWET via data.gv.at and the federal legal information system (RIS).
Privacy by design
No user account or tracking is required. Search queries are normalized for processing; the original raw text is not retained permanently.
Fallback without AI ranking
When the AI budget is exhausted, lexical search continues.
Typical use cases
Possible scenarios. Not a description of a delivered project.
Computer-vision QA
in final assembly (industry)
RAG for maintenance know-how
service technicians find answers in seconds instead of hours
Anomaly detection
in sensor data for predictive maintenance
Document AI for incoming invoices
classification, extraction, approval workflow
Case-handler assistant
for insurance and banking — pre-qualified suggestions with source attribution
EU AI Act & compliance
From August 2026 the full EU AI Act applies. High-risk AI in industry (e.g. machine safety components) requires conformity assessment. We deliver as standard:
Model card
with training-data description, known limitations, performance metrics
Data lineage
traceable data origin and processing steps
Audit logs
for every AI decision in production
Human oversight
as mandatory workflow step for high-risk applications
NIS-2 supply-chain documentation
for regulated mid-market companies
Frequent questions
Can we self-host the model (no OpenAI)?
Yes. Llama, Mistral, Phi or fine-tuned models on-premises. Alternatively Azure OpenAI in EU region with data residency.
How do you protect our data?
EU data centre by default, NDA, pseudonymisation where possible. No data goes to US providers without sign-off.
Is AI worth it for our problem at all?
That is exactly what we check in the analysis call. If the answer is "no", we say so — and propose a classical solution.
Which process costs your team time every day?
In a free analysis call we look at it together and clarify whether automation makes sense.
60 minutes · no obligation