Can AI summarise patient records without them leaving the hospital? Technically yes. Clinical safety and the legal duties stay with the hospital.
Short answer. Technically yes. An open language model running on a hospital’s own servers can summarise records without sending them to a third party, and published work shows models can write clinical summaries that physicians rate as equivalent or better than medical experts’ in most cases. They also make errors: one 2025 study measured a 1.47 per cent hallucination rate and a 3.45 per cent omission rate in model-written clinical notes. Keeping data local helps privacy, but it does not change the hospital’s clinical responsibility or the safeguards the DPDP Act asks of a data fiduciary, and a clinician must review every summary. This is general information, not medical or legal advice.
01
What the studies show
Study
What was tested
Result
Van Veen et al., Nature Medicine, 2024
Eight language models adapted for four clinical summary tasks: radiology reports, patient questions, progress notes and doctor-patient dialogue. Ten physicians rated the summaries.
In most cases the best-adapted models’ summaries were rated equivalent (45%) or superior (36%) to the medical experts’. The authors also link the errors found to potential medical harm.
Asgari et al., npj Digital Medicine, 2025
12,999 clinician-annotated sentences from model-generated clinical notes, across 18 experimental set-ups.
A 1.47% hallucination rate and a 3.45% omission rate. Refining prompts and workflows reduced major errors below previously reported rates for human note-taking.
02
What private does and does not settle
Running the model on the hospital’s own network keeps the records inside it. The DPDP Act defines personal data as any data about an individual who is identifiable by or in relation to such data, and digital personal data as personal data in digital form, so a patient record held electronically falls within it on its face. The Act also says a data fiduciary must protect personal data in its possession or under its control, including when processing is done by a processor, by taking reasonable security safeguards to prevent a personal data breach, and must erase personal data once its purpose is no longer served unless retention is needed by law. Section 1(2) of the Act lets the government bring different provisions into force on different dates, so check which apply to you as of October 2026.
Local processing is one safeguard. It does not by itself settle consent, purpose, access, retention or breach duties, nor the rules of the medical council and the hospital’s own policies. Take advice on those. This page is general information, not legal advice.
03
What a clinician must still check
Every medicine, dose, allergy and laboratory value against the source record.
Negations, such as no history of or ruled out, which are easy to flip.
Omitted findings. In the 2025 study omissions were more frequent than hallucinations.
Mixed episodes: findings from another visit or another patient pulled into the summary.
Copied-forward text that is out of date.
Whether the summary is a draft or is about to be filed as the record. A draft must stay a draft until a clinician has reviewed and signed it.
04
How a hospital would set this up
Start with a low-risk use, such as drafts of discharge summaries that a clinician edits, rather than anything that goes straight into the record.
Run an open model on the hospital’s own servers, on a network segment with no route to the internet if policy requires.
Limit access to staff who treat the patient, and log who generated each summary and when.
Before go-live, have clinicians rate 100 past cases for omissions and invented statements, and set a bar that must be met.
Mark every AI draft as a draft, and keep the source record one click away.
Decide how long drafts are kept and how they are deleted, and write down who can answer a patient’s question about them.
05
What ANVE does today
ANVE answers questions over the office’s own documents and shows the source passage, and its vision models read scanned documents with OCR, all on a device in the building. ANVE has no clinical product. It holds no certifications and makes no claim of compliance with any health-data rule; the data stays on the box in the office, and the office’s own legal duties remain the office’s. Everything sits in one database on the box. Cloud models are used only if an administrator adds a provider, and a policy that blocks confidential data from cloud models is not on ANVE yet. Full-disk encryption and secure boot are part of the appliance image, which is not on ANVE yet, and roles today are administrator and user, with teams, finer roles and single sign-on not on ANVE yet. Backup design, deletion details and administrator access to other people’s chats have not been published, so get them in writing before a paid pilot.
ANVE is pre-release: the software runs today on a development machine, no sealed unit exists yet, and there are no customers yet.
06
Questions people ask
Does keeping the model local settle the DPDP Act?
No. Local processing is one safeguard. The Act’s duties on a data fiduciary, such as reasonable security safeguards, breach intimation and erasure once the purpose is served, concern how the hospital handles the data wherever the model runs, and different provisions can come into force on different dates. Take legal advice; this is general information.
Can an AI summary go straight into the patient record?
Treat it as a draft until a clinician has reviewed and signed it. The studies above found errors even in well-tuned set-ups.
Can it read scanned or handwritten records?
Scans need OCR or a vision model first, and handwriting is harder still. Check the text it produces against the images before trusting any summary built on it.
How should a hospital test a summarising tool?
Have clinicians review past cases and count omissions and invented statements, as the 2025 study did on 12,999 sentences. Set a bar before go-live and re-test whenever the model or prompt changes.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
