Guide · checked October 2026

On-premises vs cloud AI in hospitals

NHS guidance allows health and care data, including confidential patient information, in the public cloud with the right safeguards, so the choice between on-premises and cloud AI is a local risk decision about data flows, model capability, cost and who has to approve it.

What the guidance says

NHS England's guidance on off-shoring and the use of public cloud services sets out the safeguards needed for organisations to put health and care data in the public cloud, including off-shoring. It says providers should use cloud computing services for hosting NHS data in territories deemed adequate, and that the local senior information risk owner, with the data protection officer and Caldicott Guardian, must be satisfied with security. It points to a cloud security good practice guide and a cloud risk framework.

All organisations with access to NHS patient data and systems must use the Data Security and Protection Toolkit, a self-assessment against the National Data Guardian's ten data security standards. The AI and Digital Regulations Service, curated by NICE, MHRA, CQC and HRA, collects regulation and evaluation guidance for developers and adopters.

Weighing the options

Data flow: on-site models keep prompts and records inside the network; cloud models need a data flow assessment and contracts.
Capability: the largest models usually run in the cloud; on-site models need local GPUs and are smaller.
Resilience: an on-site system keeps working when the internet link is down, but your IT team runs it.
Approval: either route needs information governance, clinical safety and procurement sign-off.

Questions to ask a supplier

Where does each piece of data go, and can outbound traffic be blocked?
What happens when the model is unavailable: is there a non-AI fallback?
Are identifiers removed before any outside call, and is every call logged?
Who can switch models, and is that change approved and recorded?

Myths

"The NHS bans cloud for patient data": the guidance allows it with safeguards.
"On-premises means no governance": local systems still need the DSP Toolkit, IG and clinical safety work.
"Cloud AI is always more accurate": it depends on the task and the checks around the model.

Where Oriel fits

Oriel runs on the hospital's own servers or a desktop in the department, with models on site through Ollama or an OpenAI-compatible server on the hospital network. Every AI step has a non-AI fallback, so audits still run without a model. A hosted model can be added only if the site chooses, behind a site policy, with identifiers pseudonymised before any outside call and every call in the action log.

Admin readiness checks include confirming outbound network access is blocked. Oriel is a pilot on synthetic data and is not a medical device.

Questions

Can NHS patient data be stored in the public cloud?

NHS England's guidance allows it with safeguards, in territories deemed adequate, once the local SIRO, DPO and Caldicott Guardian are satisfied.

Does an on-premises system still need the DSP Toolkit?

The toolkit applies to all organisations with access to NHS patient data and systems, wherever the system runs.

Sources

Official pages, checked October 2026. Guidance changes, so check the current version before you rely on it.

General information for healthcare managers and clinical teams, checked October 2026. It is not legal, regulatory or clinical advice; follow the current official guidance and your organisation's own policies. No patient data is used in these examples. Oriel is a working prototype shown on a fictional demo hospital with synthetic data only. It is not a medical device and is not for clinical use. It drafts and flags for people to decide; it does not diagnose or change clinical records. A clinical safety case and information-governance approval come before any use with real patient data. Oriel is a working name.