The Case for On-Device AI in Fitness Apps (GDPR, Privacy, and Why It Matters)
By Ollie
The Case for On-Device AI in Fitness Apps (GDPR, Privacy, and Why It Matters)
Every AI fitness app you've used sends your data to a server. Your injury history, your weight, your goals, your workout logs — all uploaded to someone else's cloud, processed by someone else's model, stored in someone else's database.
Most users don't think about this. But they should.
What "AI-Powered" Usually Means
When a fitness app says it uses AI, here's what's typically happening:
1. You enter your goals, body measurements, injuries, and preferences 2. That data gets sent to a cloud API (usually OpenAI, Anthropic, or a fine-tuned model on AWS) 3. A model running on someone else's servers processes your health information 4. The result comes back to your app
Your data now exists in at least three places: your device, the app's backend, and the AI provider's servers. Each with their own data retention policies, security practices, and potential breach surface.
For European users, this creates a GDPR problem. Health data is a "special category" under GDPR Article 9 — it requires explicit consent and has strict rules about processing, storage, and transfer. Most fitness apps handle this with a blanket consent checkbox that nobody reads.
What On-Device AI Actually Means
PumplAI takes a different approach. We use Qwen3 models running via Ollama — directly on-device. Here's what changes:
- Your fitness goals stay on your phone. The embedding model that understands "I want to rebuild core strength after pregnancy" runs locally. No server round-trip.
- Workout generation happens locally. The LLM that creates your training plan processes your data without it ever leaving your device.
- No third-party AI provider sees your health data. No OpenAI. No Anthropic. No cloud inference API.
- Qwen3-Embedding (0.6B) — converts your goals into vector embeddings for trainer matching
- Qwen3 1.7B — routing model (decides task complexity)
- Qwen3 4B/8B — workout generation (simple vs complex plans)
These models are small enough to run on modern smartphones and laptops. The 0.6B embedding model is 600MB. The 4B workout generator is about 2.5GB. That's less than a typical mobile game.
Why This Matters for EU Users
PumplAI is built from the EU, hosted on EU infrastructure (Neon Serverless Postgres, EU region), and designed with GDPR compliance as an architectural decision — not a policy document bolted on after launch.
Here's what that means in practice:
| Data type | Where it's processed | Where it's stored | |---|---|---| | Fitness goals | On your device | On your device | | Workout plans | On your device | On your device + your trainer (if shared) | | Trainer matching vectors | On your device → EU database | EU Neon Postgres (pgvector) | | Account info (email, name) | EU servers | EU Neon Postgres | | Payment data | Stripe (PCI compliant) | Stripe only |
No health data crosses an ocean. No AI inference happens outside your control.
Why This Matters for Trainers
If you're a trainer using a platform that sends client data to US-based AI providers, you have a data protection responsibility you might not know about.
- Explicit informed consent covering the specific processing
- A valid data transfer mechanism (Standard Contractual Clauses or equivalent)
- A Data Protection Impact Assessment for health data processing
Most trainers using AI platforms have none of these. Most platforms don't tell trainers they need them.
PumplAI removes this problem entirely. The AI runs on the client's device. The trainer sees results, not raw health data processed through third-party servers.
The Competitive Landscape
No major fitness platform is doing on-device AI today:
| Platform | AI processing | Data location | |---|---|---| | Trainerize | No AI generation | US servers | | Everfit | Cloud AI (provider unknown) | US servers | | TrueCoach | Assistive AI only | US servers | | FirstRep | Cloud AI agent | US servers (assumed) | | PumplAI | On-device (Qwen3 via Ollama) | EU + on-device |
This isn't just a privacy feature. It's a structural advantage. When regulators tighten health data rules — and they will — platforms built on cloud AI for health data will need expensive retrofits. PumplAI is already compliant by architecture.
Try It
We're opening early access for trainers and clients who want AI that respects their data.