# SAFFT AI > Private AI for firms that handle confidential data. SAFFT AI installs open-source language models, a searchable knowledge base of the firm's own documents, a task agent and long-term memory on hardware dedicated to that firm: an AI workstation in its office (SAFFT AI Box) or a dedicated GPU server in a UK or EU region (SAFFT AI Cloud). Prompts and files are not sent to public AI services. SAFFT AI is a service of SAF FinTech Advisory Ltd. ## Who it is for Accountancy practices, law firms, wealth managers, fintechs and clinics in the UK: teams that cannot paste client data into public chatbots. ## What is included - The model: open-weight LLMs (for example Qwen, Llama, Gemma, Hermes), chosen by testing on the client's real work. - Your knowledge: retrieval over the firm's document stores (RAG), with source citations and folder permissions. - The agent: Hermes Agent for multi-step tasks, using only tools the client approves. - Memory: per-person and per-team memory that can be viewed and erased. - Controls: role-based access, an audit log of every exchange, and human approval before anything leaves the system. ## Deployment options - SAFFT AI Box: on-premises AI workstation (128 GB unified-memory class), can run fully offline. One-off hardware and setup, then a monthly care plan. - SAFFT AI Cloud: dedicated GPU server, not shared, in a UK or EU region. Monthly subscription including GPU time. - Clients can move between Cloud and Box; documents, memory and settings move with them. ## How to start A 4-week pilot on a dedicated cloud GPU with the firm's own documents (after about one week of scoping). At the end of week 5 the client receives costs for both Box and Cloud and chooses. Ongoing care covers model updates, monitoring, backups and new skills. ## Limits, stated plainly Strong at answering from the firm's documents with sources, first drafts in house style, summaries and extraction. Weaker than the largest public models at the hardest open-ended reasoning. Capacity of a single Box depends on model and usage and is measured during the pilot. SAFFT AI is designed to support data-protection and confidentiality duties; the client's DPO or compliance lead should review the final setup. ## FAQ ### Which AI models do you use? Open-weight models such as Qwen, Llama, Gemma and Hermes. We don't pick one in advance. During the pilot we test several on your work and keep the best balance of quality and speed for your hardware. ### How many people can use one box? That depends on the model, the length of your documents and how many people ask at the same moment. We measure it during the pilot and size the hardware to your team instead of quoting a number up front. ### Is it as good as ChatGPT or Claude? On questions about your own documents it is often more useful, because it has read them and cites them. On the hardest open-ended reasoning the largest public models are still ahead. The pilot shows you where that line falls for your work. ### Can it run without the internet? Yes, a box can run fully offline. Updates are then installed on a schedule you agree, either on site or over a connection you open for that purpose. ### Do you see our data? Only when you give us access for setup or support, and that access is logged. The rest of the time the system runs without us. ### What if we want to stop? You keep the box hardware. With either option you can export your knowledge base, memory and logs, and we delete any copies we held. ## Contact - Website: https://safftai.com/ - Email: info@saffintech.com - Parent company: SAF FinTech Advisory Ltd, https://www.saffintech.com/