AI
Advice, rollout, agents - GDPR-compliant, without the hype
Artificial intelligence - used where it helps, introduced in a way that holds up under data protection law. We help when you want to work with AI seriously, not just produce marketing material about it.
Was wir in der Säule AI machen.
1 Topics, jedes einzeln buchbar oder als Gesamt-Setup. Klick auf eine Zeile öffnet die Detail-Page.
| No | Topic | Type |
|---|---|---|
| 01 | Copilot rollout· Microsoft 365 Copilot - introduced in a GDPR-compliant way | Topic |
So bauen wir.
Vier Punkte, die in jedem Projekt der Säule AI unverhandelbar sind.
About this pillar.
What AI means at CAVORT
AI is not an end in itself. We bring it in where it removes real friction - and leave it out where it would only be hype.
Advice
Before we introduce tools, we ask: where are you currently losing time that AI could save? Typical answers from our client conversations:
- Processing receipts and invoices
- Email triage and first replies
- Producing documents with recurring patterns
- Searching internal documents for knowledge
- Client onboarding paperwork
- Translation and localisation
- Writing up minutes from meetings
Those are the areas where AI is mature today and delivers real value. Anything beyond that - fully autonomous customer service agents, AI making HR decisions, generative video for marketing - is possible, but carries risks we will name honestly.
Rollout
If you want to introduce Microsoft 365 Copilot, we do:
- Pre-flight check - SharePoint hygiene, labels, permissions. Without this, Copilot surfaces things it should not.
- Pilot group - five to ten power users, supported intensively for four weeks
- Training - half-day workshops per department
- Wider rollout - with usage monitoring
- Optimisation - review after three months, adjust policies
For Claude or ChatGPT Enterprise the steps are similar, but the technical preparation is smaller.
Local AI
Where cloud AI is not an option, we build on-prem:
- Ollama as the model runtime
- Mistral as the default choice for Europe (Small, Medium, Large)
- Qwen as an alternative
- vLLM for production workloads with several concurrent users
- OpenWebUI or a custom frontend
- RAG pipeline over ChromaDB or Weaviate
- Monitoring via Grafana and Sentry
Typical target environment: a Proxmox host with an NVIDIA GPU, or a Mac Studio M4 Max for small teams.
Agents
An agent is a small tool with a limited scope. Examples we have built:
- Invoice agent - reads incoming invoice emails, extracts metadata, files the PDF into DATEV
- Email triage agent - classifies support tickets and routes them to the right person
- Knowledge agent - answers questions from internal documentation via RAG
- Meeting minutes agent - transcribes Teams meetings, extracts decisions and to-dos
Every agent has clear boundaries: what it may do, what it has to hand on, and where a human should look.
Training
The best tool is useless if nobody uses it - or if people use it wrongly and data ends up in random ChatGPT accounts. We build training that stays concrete:
- How to write a prompt that actually helps
- How to spot when the model is hallucinating
- How to use Copilot for recurring Word documents
- How to work with Claude on an Excel analysis
No abstract AI theory - the actual work your people do.
How we work together.
From the first hello to running operations in five steps.
AI - frequently asked.
FAQ.01Which AI providers do you use?
It depends on the data protection requirements and the budget. Copilot for Microsoft ecosystems. Claude from Anthropic or Mistral for independent workflows. Ollama with Mistral or Qwen for local deployments. We rarely use US hyperscaler products directly - and when we do, via an EU region with zero-data-retention terms.
FAQ.02Is Copilot really GDPR-compliant?
With the right settings, yes - but Microsoft does not make it easy. The key questions: are sensitivity labels set, are SharePoint permissions clean, is the EU Data Boundary active, is cross-tenant data inclusion off. We check all of that before switching Copilot on.
FAQ.03When is local AI better than cloud AI?
When data protection is tight, as with law firms, medical practices or public bodies. When large volumes of tokens make cloud costs explode. When you need fine-tuning on your own data. When an air gap is required. For an everyday writing assistant, cloud is usually simpler.
FAQ.04What does a local AI setup cost?
A GPU workstation for team inference starts at around 5,000 EUR with an RTX 4090 in a consumer case and scales up to an A5000 or A6000 at 10,000 to 20,000 EUR. Installation and integration come on top - budget five to fifteen days once, then ongoing maintenance from around 200 EUR a month.
FAQ.05Do you build your own agents?
Yes. For us an agent is a small tool that handles one bounded task - read an invoice and import it into DATEV, triage email, search internal knowledge with RAG. We build these on our own stack, with n8n for orchestration, Python or Rust for specific logic, and OpenAI-compatible APIs locally or in the EU.
FAQ.06Can you train our staff on AI?
Yes. We run half-day workshops on prompt engineering, Microsoft 365 Copilot, using ChatGPT and Claude, and how to apply them in your specific working day. No introductory theory - concrete use you can apply tomorrow.
FAQ.07Will AI replace people in our company?
Probably not in the next few years for most roles. Very probably AI will change the working day - automating repetitive parts, speeding up conceptual work. We advise pragmatically, not evangelically.
Sprechen wir 30 Minuten über AI.
The intro call is free. No sales pressure. We listen, check whether we fit - and tell you straight.