AI and hospital operations
On-premises AI in hospitals: when patient data can't leave the building
On-premises AI means the models run on the hospital's own servers, so patient data doesn't go to a supplier's cloud. NHS England's guidance names the risks of externally hosted AI (data leakage, less control over data handling, harder privacy assessments) but also says on-premises models carry risks of their own, which need appropriate security and risk controls.
Why hospitals look at running AI on-site
NHS England's ambient scribing guidance, written for CIOs and CCIOs, lists the data risks of AI products: leakage and compromised confidentiality in a supplier's hosted environment, being unable to govern privacy and data protection policies in externally hosted environments, and the difficulty of assessing third-party "black-box" products. Running the model inside the hospital takes the supplier's hosting out of that list. It also keeps working when the internet link doesn't, which matters for anything on the operational critical path.
What doesn't go away on-premises
- Security of the AI itself. NHS England warns that AI systems bring their own security challenges, including indirect prompt injection attacks that can extract a user's personal data from large language models.
- Information governance. A DPIA is still legally required before implementing AI, and NHS England's security expectations still apply: role-based access, audit logs, encryption, limits on exports and a record of all movement and storage of personal data.
- Clinical safety. DCB0160 applies to the deploying organisation wherever the software runs.
- Accuracy and monitoring. You still need agreed accuracy levels, ongoing performance monitoring and human oversight of outputs.
- Operations. Someone has to size the hardware, patch the stack, update models, back up and restore, and test it.
Questions to settle before you start
- Which data classes will the model see: operational only, demographics, or clinical? Can that be enforced by the platform, not just promised?
- Will any hosted model be used at all, and if so, for what data and under what data processing agreement?
- How are models updated in a site without internet access, and how is a change reviewed by the clinical safety officer?
- What hardware does the workload need, and what happens when it fails?
- How will you show the DSPT and DTAC reviewers where data goes?
What it does not mean
- On-premises isn't automatically safe. NHS England says on-site models carry risks too and need proper security and risk controls.
- It doesn't remove the DPIA or the safety case. Both still apply to the hospital.
- It doesn't have to mean no cloud at all. What matters is deciding which data any hosted model may see, and enforcing that in the platform.
Where Thalix fits
Thalix runs on the hospital's own hardware from one Docker Compose stack: the node, Ollama with pinned open models, and optional MQTT and NATS brokers. There's no database to install (every store is SQLite in the node's volume). For hospitals without internet access, an offline bundle carries the images, model weights, Docker Engine for Linux, an installer and checksums. Every model declares the data classes it may see and the kernel enforces it; if you choose a hosted model, it sees operational data only, never patient records. The "Thalix in a box" hardware on our home page is concept design (renders, not photographs, with illustrative specifications).
Thalix is at pilot stage. Demos run on synthetic data, and a hospital deployment needs its own clinical safety case, information governance approval and security review. We don't claim any certification or assessment result the product doesn't have.
Questions people ask
Is on-premises AI safer than cloud AI for the NHS?
Not automatically. NHS England's guidance lists the risks of externally hosted AI models but also says on-premises models carry risks, which need appropriate security and risk controls.
Do you still need a DPIA for on-premises AI?
Yes. NHS England's IG guidance says a DPIA must legally be completed before implementing AI-based technologies, wherever they run.
Can large language models leak patient data?
NHS England's guidance warns that indirect prompt injection attacks can be used to extract a user's personal data from large language models, which is one reason to limit what data a model can see.
Sources
Each source below is an official page, read and checked in October 2026. Where a page and this guide differ, the official page wins.
- Guidance on the use of AI-enabled ambient scribing products in health and care settings NHS England · checked October 2026
- Artificial intelligence guidance for IG professionals NHS England Digital · checked October 2026
- DCB0160: Clinical risk management in the deployment and use of health IT systems NHS England Digital · checked October 2026
- Data Security and Protection Toolkit NHS England, DSPT · checked October 2026
General information only, written for hospital digital, IT, operations and clinical safety teams. It isn't legal, regulatory, information governance or clinical safety advice, and it doesn't replace your own organisation's processes or the official guidance linked above.