Every AI tool you use sends your data somewhere. ChatGPT goes to OpenAI's servers. Copilot goes to Microsoft. Even the privacy-focused alternatives process your prompts in data centers you do not control.
For most people, that tradeoff is fine. For some, it is not. Maybe you handle sensitive client data. Maybe you are in a regulated industry. Maybe you just do not like the idea of your business processes being processed on someone else's hardware.
Here is an alternative: a Raspberry Pi 5, an open-source AI agent, and zero data leaving your local network. Total cost: EUR 85 for the hardware, a few hours of setup time, and zero monthly subscriptions.
What You Are Building
A Raspberry Pi 5 running a local AI agent that can:
- Answer questions from your documents (contracts, procedures, knowledge base)
- Automate simple tasks (file organization, email drafts, data extraction)
- Process text (summarization, translation, analysis)
- Run on your local network, accessible from any device in your home or office
- Operate without an internet connection after initial setup
What you are NOT getting:
- The power of GPT-4 or Claude (local models are smaller and less capable)
- Real-time web browsing
- Complex multi-step agent workflows (hardware limitations)
- Voice processing (CPU is not fast enough for real-time speech-to-text)
Let's be upfront: a Raspberry Pi AI agent is a Honda Civic, not a Ferrari. It gets you where you need to go for common tasks. It will not win any races against cloud-based AI. But it is yours, it is private, and it costs EUR 85.
Hardware Shopping List
Here is exactly what you need:
Raspberry Pi 5 (8GB RAM model) EUR 85
MicroSD card (128GB, A2 speed class) EUR 15
USB-C power supply (27W, official) EUR 12
Case with cooling fan EUR 10
Ethernet cable (optional, WiFi works) EUR 3
----------------------------------------------
Total: EUR 125The 8GB RAM model is mandatory. The 4GB version will not run any useful AI model. If you want better performance, add a USB 3.0 SSD (EUR 30-50) instead of relying on the microSD card. The SSD upgrade makes model loading 3-4 times faster.
Optional but recommended:
- NVMe HAT + 256GB NVMe SSD: EUR 45. Best performance option. Model loading becomes nearly instant and you have room for a larger document collection.
Software Setup
This guide assumes basic comfort with the command line. If "SSH into your Pi" sounds alien, this project might not be for you.
Step 1: Flash 64-bit Raspberry Pi OS Lite onto your microSD card using Raspberry Pi Imager. Enable SSH. Boot.
Step 2: SSH in and update: sudo apt update && sudo apt upgrade -y. Set a static IP.
Step 3: Install Ollama (open-source local AI runtime): curl -fsSL https://ollama.com/install.sh | sh
Step 4: Download a model. For 8GB RAM, the 3B parameter models work well:
ollama pull llama3.2:3b # Recommended (3.8GB, good balance)
ollama pull phi3:mini # Lighter, faster (1.3GB)7B models work but are noticeably slower. Anything above 7B will not fit in memory.
Step 5: Install the agent framework:
pip install openclaw-edge
openclaw-edge init --model llama3.2:3b --port 8080Step 6: Index your documents: openclaw-edge index --source /path/to/your/documents. The agent indexes PDFs, Word files, text files, and spreadsheets into a local search index.
Performance Benchmarks
Tested on Raspberry Pi 5 (8GB RAM, NVMe SSD, Llama 3.2 3B model):
Task Time Quality (vs GPT-4)
Simple Q&A 2-4 seconds 85% as good
Document search (500 docs) 3-6 seconds 90% as good (factual retrieval)
Text summarization (1 page) 8-12 seconds 75% as good
Email draft 5-10 seconds 70% as good
Code generation (simple) 6-15 seconds 60% as good
Translation (EN<->PL) 4-8 seconds 80% as good
Complex reasoning 10-30 seconds 50% as goodThe "Quality vs GPT-4" column is subjective but based on side-by-side comparisons across 50 test queries. The local model is significantly less capable for complex reasoning, creative writing, and nuanced tasks. It is surprisingly competent for factual retrieval, simple text generation, and structured tasks.
Response speed: Expect 5-15 tokens per second with the 3B model. That is noticeably slower than cloud AI (which generates 50-100 tokens per second). For short answers, the delay is acceptable. For long documents, you will be waiting.
5 Practical Use Cases
1. Private document search
You have 200 client contracts, 50 internal procedures, and a bunch of policy documents. Instead of scrolling through folders, you ask: "What is the termination notice period in the Acme Corp contract?" The agent searches your indexed documents and returns the answer with a source reference.
This is the strongest use case for a Pi-based AI agent. Document search is mostly retrieval (finding the right passage), not generation (creating new text). The local model handles retrieval well.
2. Email draft assistant
You describe what you need to communicate, the agent writes a draft. "Write an email to Supplier X asking about delivery delays. Tone: firm but professional. Reference our contract's delivery timeline clause."
The drafts need more editing than GPT-4 output, but they are a solid starting point. Saves 5-10 minutes per email for someone who struggles with writing.
3. Data extraction from invoices
Photograph invoices, send images to the Pi. The agent extracts vendor name, amount, date, VAT, and category. Exports to CSV. Accuracy: 85-90% for clean, printed invoices.
4. Local automation scripts
The agent writes and runs automation scripts on your local network: file backups, disk space monitoring, batch file renaming. Simple tasks that most people never get around to doing manually.
Privacy: What This Actually Means
No data transmission - everything stays on your network. No terms of service, no account, no tracking. Works offline after setup. You can wipe everything by formatting the SD card. For businesses handling sensitive data (legal firms, medical practices, financial advisors), this means AI capabilities without GDPR concerns about third-party data processing.
Limitations - Honest Assessment
Not powerful. A 3B model will not match GPT-4 on complex tasks. Slow. 5-15 tokens per second. Technical setup. Requires command-line comfort and basic Linux knowledge. No multi-modal. Image analysis and voice interaction are beyond practical capabilities.
When to Use This vs. Cloud AI
Use the Pi for private document search, repetitive text tasks where privacy matters, offline environments, and learning. Use cloud AI for complex reasoning, high-quality creative content, and real-time applications. Use a custom agent from Syntalith when you need both quality and privacy - EU-hosted, GDPR-compliant, connected to your CRM and business systems.
Syntalith is an AI software house based in Warsaw. We build custom AI agents from EUR 1,499 - hosted in the EU, no data shared with third parties. The Pi is a great learning tool. For business, you want production-grade infrastructure.
Talk to us about a private AI agent for your business - EU-hosted, GDPR-compliant, demo in 7 days.