# How our AI works

Our agents draft and do routine work, and a person approves anything that matters.

- **Humans approve what matters:** You decide what an agent can do alone and what needs a tap from a person. Payments, deletions and first messages to a new customer can always wait for approval.
- **Your data stays yours:** Agents run on accounts you own. You can see what they did, export it, and switch them off at any time.
- **No training on your data:** We use providers' business and API plans, which don't train on your inputs by default, and we never train models on client data. [CONFIRM per provider]
- **The right model for the job:** We mostly use Claude (Anthropic) and OpenAI models. When data must stay private or on your own hardware, we use open-source or locally hosted models.
- **Security in plain English:** Each agent gets only the access it needs. Passwords and keys live in a secrets manager, never in code or chat. Every action is logged.
- **When AI isn't sure, it asks:** Agents are tested on real examples before launch. Anything they're unsure about goes to a person instead of a guess.

Book a free 20-minute AI audit: https://modulixsolutions.com/contact/