How a GP drafts a referral letter with Hush AI
A two-week-wait referral letter needs to be right, and it needs to be fast. The clinical content is in front of you; the bottleneck is turning it into a structured letter that the receiving clinician can act on. This is a walkthrough of how that works with Hush AI, from paste to polished letter, in under a minute.
Every step happens on hardware we own. No patient data is stored on our servers and nothing is used for training. If that claim matters to you, and it should, we explain how to verify it at the end.
The five-step workflow
Go to hush-ai.uk/app in your browser. There is nothing to install, no plugin for EMIS or SystmOne, no IT project. It runs in the browser you already have open.
Switch to EMIS or SystmOne, copy the relevant consultation notes, investigation results and any referral criteria the letter needs to address. Switch back to Hush and paste it in.
You can type a short instruction above the paste, for example: "Draft a two-week-wait referral letter to dermatology for this patient. Include the lesion description, timeline, dermoscopic findings and relevant history."
What happens to the data at this point: the text you paste is sent over TLS to hardware we own and operate in the UK. It is processed in memory to generate the draft, then it is gone. It is not written to disk, not stored in a database, not logged and not queued for training. There is no US cloud in the chain and no third-party AI API.
Hush returns a structured referral letter in seconds. It follows standard NHS referral letter conventions: patient demographics at the top, reason for referral, clinical findings, relevant history, current medications, and the specific clinical question for the receiving team.
You read every word. Hush is a drafting assistant, not a clinical decision-maker. The letter is a draft until you say it is not. Check the clinical content against the source notes, verify that nothing has been hallucinated, and edit anything that needs changing. This is the step that matters most.
Once you are satisfied, copy the letter and paste it into your clinical system. Attach it to the patient record, send it through the usual referral pathway. The workflow is copy-paste in both directions, which is deliberate: it means no integration is needed and nothing needs to be approved by your clinical system supplier.
After the draft is generated, the patient context you pasted is not on our servers. If you want a record of the conversation, it is stored encrypted on your own device and you can delete it in one click. The audit log records that a drafting session took place (when, which tool, how many tokens), but the content of what you drafted is not in it.
What this replaces
Without AI, a referral letter is a manual drafting job. You read through the notes, open a template, type the letter, re-check it against the notes, and send it. For a straightforward two-week-wait, that is five to ten minutes. For a complex referral with years of history, it can be longer. Multiply that by eleven letters at the end of an evening session and the arithmetic is uncomfortable.
With Hush, the drafting step drops to seconds. The review step still takes as long as it should, because you are still the clinician. But the mechanical work of turning clinical notes into a structured letter is handled by the AI, on hardware that keeps none of it.
What this does not replace
Clinical judgement. Hush does not decide whether to refer, which pathway to use, what urgency to assign, or what the clinical findings mean. It drafts a letter from the context you give it, and you own the result. It is not a medical device and it does not generate clinical notes. It sits deliberately on the administrative side of the MHRA medical-device boundary.
If you want an ambient scribe that listens to consultations and writes notes, Hush is not that product, by design. We wrote about why Hush is not an ambient scribe separately.
What about other AI tools?
You can draft a referral letter with ChatGPT, Claude, Gemini or Copilot. The drafting quality may be comparable. The difference is in what happens to the patient data you paste in:
- ChatGPT (free and Plus): your inputs can be used to train models unless you change a setting. Processed in the US, by a US company, subject to US legal process including the CLOUD Act.
- Claude: Anthropic states it does not train on inputs by default on paid plans, but it is a US company, processed on US infrastructure, subject to the CLOUD Act.
- Gemini: Google's terms have varied; consumer Gemini inputs may be used for product improvement. US company, US infrastructure.
- Hush AI: no training, no storage, no US entity anywhere in the chain. Processed on hardware we own in the UK, by a UK company with no US parent. Outside the reach of the CLOUD Act.
The drafting is comparable. The data governance is not. For a two-week-wait referral that contains the patient's name, NHS number, clinical findings and the clinician's identity, that distinction matters. (Our AI training checker lets you look up any tool's terms in detail, no email required.)
How to verify our claims
We have a verify-us page with copy-paste terminal checks you can run yourself: TLS certificate (issued to hush-ai.uk, not a cloud provider), security headers, and the post-quantum handshake that proves the connection goes to hardware we operate. The company is at Companies House (17278687) and the ICO register (ZC126901). We publish what we do not yet hold as well as what we do.
Try it with a synthetic patient
Open Hush, paste some fictional clinical notes, and draft a referral letter. No account needed for the free tier. See whether the output is good enough to review, and verify the privacy claims while you are there.
Open Hush →Want to evaluate Hush for your practice? The Founding Practice pilot is a free two-week trial: no card, no calls, up to five logins, capped because our hardware is finite. Details on the GP page.
About Hush AI: Hush AI (hush-ai.uk) is a private AI assistant built by its founder, Dr W.J Carter. It runs on hardware we own in the UK, with no US parent and no third-party AI API. It is not a medical device and not an ambient scribe: it drafts documents under clinician review. Conversation history is stored encrypted on your own device, never on our servers, and nothing is used for training.