--- title: "Clinical AI Systems: Ambient Documentation & RAG" url: https://ypai.ai/solutions/healthcare/clinical-ai-systems/ description: "YPAI builds clinical AI into the workflow: ambient documentation with EHR writeback, retrieval assistants over approved records, human-in-the-loop review." source: "src/copy/routes (route /solutions/healthcare/clinical-ai-systems/)" --- # A model is not a system. Your clinicians need the system. > YPAI builds clinical AI into the workflow: ambient documentation with EHR writeback, retrieval assistants over approved records, human-in-the-loop review. Ambient documentation that writes back to the record, assistants that answer only from your approved sources, document flows that stop where judgement belongs. Built into the systems you already run, and evidenced for the review your procurement team will run. ## The pilot impressed everyone. Three questions decide whether it reaches the ward. Clinical AI rarely fails on the model. It fails on the record it never wrote to, the answer nobody could trace, and the paperwork nobody could produce. ### Does it reach the record? 72-year-old with exertional chest pain since Tuesday. written to the record · signed - A note that lives in a demo is not documentation. Production means clinical speech recognition, speaker separation, a structured note, and a write path into the record system your clinicians actually work in. The integration is the product. ### Can we trace what it said? Plan: cardiology referral per the chest pain pathway. cited · chest pain pathway, version 3 - Ungrounded language models are wrong often enough on clinical text that grounding is not a refinement, it is the design. Every answer is retrieved from your approved sources, cited to its document and version, and rejected when nothing supports it. ### Who writes our conformity file? Signed · attending clinician · system 2.4.1 logged with the version that produced it - We do, for the workload we build. Technical documentation, data governance, human oversight design and logging are produced as part of delivery, in the form your GDPR and EU AI Act review asks for. ## Five systems, built into the workflow you already run. Each is scoped to one workflow and one clinical setting. What the system may say, where it may write, and where it must stop are decided with you before the build starts. ### Ambient clinical documentation - The encounter becomes a draft note, in the record, before the clinician stands up. - Clinical speech recognition with medical terminology, speaker separation, structured note generation and a write path into the record system, with the clinician holding the signature. Available in Danish, Norwegian, Swedish and Finnish, where the clinical language cannot be scraped and has to be collected first. - **Input**: Consultation audio · diarized - **Draft**: Structured note · to the agreed template - **Write path**: confirm. Into the record system, on signature Clinician signs · the draft becomes the note ### Retrieval assistants over your records - An answer a clinician can check, or no answer at all. - Assistants grounded in your approved documentation and records: permission-aware retrieval, every claim cited to its document and version, refusal when the sources do not support an answer, and handoff to a named person at the boundary. - **Retrieval**: Approved sources only · permission-aware - **Answer**: confirm. Cited to document and version - **Out of scope**: flag. Refused and routed to a person Grounded or nothing · both are logged ### Document and records flows - Referrals, letters and forms handled as a controlled workflow, not an inbox. - Extraction, classification and routing for clinical documents, with review queues where judgement belongs, exception handling that surfaces the hard cases, and an audit record on every step. - **Extracted**: To the agreed schema - **Routing**: To the right queue and role - **Exception**: flag. Surfaced, never silently resolved Routed · with the trail intact ### Human-in-the-loop review - The place where a person rules on what the model produced. - Review stations for clinical model output: queues, escalation paths, credentialed reviewers where the domain requires them, and a record of who ruled on what, when and on which version of the system. - **Queue**: Routed by risk and role - **Reviewer**: confirm. Credentialed where required - **Ruling**: Recorded with system version Every consequential output has an owner ### Domain adaptation, when grounding is not enough - The honest order of operations. Ground it first, adapt it only if that falls short. - Where retrieval alone cannot carry the reasoning or the clinical register, we adapt a model: dataset curation, the training run, evaluation against held-out benchmarks and a deployment handoff in the EEA by default. Most clinical workloads do not need this, and we will say so. - **First**: confirm. Grounding and retrieval - **Then**: Adaptation, only where it is warranted - **Proof**: Held-out benchmark before handoff Deployed in the EEA · evaluated, not assumed ## The model drafts. The clinician rules. The record keeps both. One synthetic encounter, five systems working on it at once. The consultation is transcribed and drafted into the record, every claim is traced to an approved source, the clinician signs or sends back, the referral waits for a human decision, and the record logs all of it with the version that produced it. Scripted demonstration on a synthetic encounter. Nothing here is a live model. System 2.4.1 · note template v7 · pathway v3 - Chest pain since Tuesday. It comes on when I climb the stairs. - Any shortness of breath with it? - You are 72. Still on the aspirin? - I will write this up and refer you to cardiology. - Hva kan jeg hjelpe deg med i dag? - Brystsmerter siden tirsdag. Det kommer når jeg går i trapper. - Nei, ikke i det hele tatt. - Du er 72. Bruker du fortsatt aspirinen? - Jeg skriver dette opp og henviser deg til kardiologen. - Vad kan jag hjälpa dig med i dag? - Bröstsmärtor sedan i tisdags. Det kommer när jag går i trappor. - Du är 72. Tar du fortfarande aspirinet? - Jag skriver ner det här och remitterar dig till kardiologen. - Hvad kan jeg hjælpe dig med i dag? - Brystsmerter siden tirsdag. Det kommer, når jeg går på trapper. - Du er 72. Tager du stadig aspirinen? - Jeg skriver det ned og henviser dig til kardiologen. - On aspirin 75 mg since March. - 72 år gammel med anstrengelsesutløste brystsmerter siden tirsdag. - Bruker aspirin 75 mg siden mars. - Plan: henvisning til kardiologi etter brystsmerteforløpet. - 72-årig med ansträngningsutlöst bröstsmärta sedan tisdag. - Står på aspirin 75 mg sedan mars. - Plan: remiss till kardiologi enligt bröstsmärteförloppet. - 72-årig med anstrengelsesudløste brystsmerter siden tirsdag. - Får aspirin 75 mg siden marts. - Plan: henvisning til kardiologi efter brystsmerteforløbet. ### Consultation audio - this encounter · 00:00 to 00:31 · two speakers ### Chest pain pathway - Exertional chest pain, age over 65: refer to cardiology and record functional class. - Antiplatelet therapy: continue existing aspirin unless contraindicated and document the dose. Answers come only from the two approved sources. Ask something they do not cover and the assistant says so. ### What does the pathway say about referral? Exertional chest pain at age over 65: refer to cardiology and record functional class. ### Should the aspirin continue? Continue existing aspirin unless contraindicated and document the dose. This encounter records 75 mg since March. ### What did the ECG show? Not in the approved sources. No ECG is recorded for this encounter, and the pathway does not cover it. Waiting for the draft and its sources Sent back with a correction. The corrected sentence is marked; the clinician signs again. Signed by the attending clinician and locked with the system version. Generated from the signed note. One field waits for the clinician. - 72-year-old, exertional chest pain since Tuesday, no shortness of breath. On aspirin 75 mg daily. - Requested: assessment per chest pain pathway, version 3. Set by the clinician, never by the model Draft composed · note template v7 ### The draft is the model's. - Generated from this encounter and the approved template, never from memory of another patient. ### The signature is human. - No draft becomes documentation without a clinician's decision. Sending it back is a first-class outcome, not an error state. ### The record outlives the shift. - Draft, sources, decision and system version are logged together, so the answer can be reconstructed months later. ## What we operate, in the words your reviewers use. These are controls we run, not outcomes we promise. They are written into the engagement so your privacy, procurement and clinical governance reviews have something concrete to inspect. Each one is shown at work on the desk above. - Every claim traced to a retrieved, cited source - note sentence 1 · source 1, consultation audio 00:04 - Retrieval restricted to approved documents and versions - chest pain pathway · approved version 3 - Refusal when the sources do not support an answer - the ECG question · refused, no supporting source - Who and what it may touch. - Prompts, models and knowledge bases versioned and pinned - system 2.4.1 · note template v7 · pathway v3 - Role-based access, with retrieval limited to what the task needs - attending clinician · two sources in scope - Immutable logging of inputs, sources, outputs and rulings - the record · eight events, each with its version - Human review on consequential output, by role and by risk - sent back at 00:41 · signed at 00:45 - Evaluation before release and regression on every change - not on this desk · see clinical model evaluation - Monitoring, change control and rollback after release - the version in every log line is what rolls back ## The system runs in your environment. The evidence comes with it. A Norwegian company under GDPR. Clinical project data is stored in Europe by default and processed in the EEA by default, with the paperwork to prove both. US healthcare work is handled HIPAA-compliant. - **Jurisdiction**: Norway · GDPR-native - **Deployment**: Your environment · EEA by default - **Access**: Role-based · minimum necessary - **Logging**: Inputs, sources, outputs, rulings - **US engagements**: HIPAA-compliant handling - **Contracts**: Standard DPA terms · SCCs available - **Erasure**: 30-day end-of-contract SLA - **Configuration**: Authority stays with you - [The healthcare hub](https://ypai.ai/solutions/healthcare/) - [Clinical data collection](https://ypai.ai/solutions/healthcare/custom-data-collection/) - [Clinical model evaluation](https://ypai.ai/solutions/healthcare/clinical-model-evaluation/) - [Clinical text and coding](https://ypai.ai/solutions/healthcare/clinical-text-and-coding/) - [Medical imaging annotation](https://ypai.ai/solutions/healthcare/medical-imaging-annotation/) - [The AI Implementation service line](https://ypai.ai/ai-implementation/) ## Start with one workflow, one clinic, one signature. We scope against the workflow as it runs today, agree what the system may say and where it may write, and prove it on a bounded pilot before it reaches care. Scope · Design · Pilot · Acceptance · Production A bounded pilot carries its own acceptance criteria and runs in a controlled workspace. You judge the output against your clinic, not against a slide. Briefs are treated as confidential. We are used to systems that cannot leave the EEA and programmes that cannot be named. Tell us the workflow, the record system it runs in, and where it strains today. We reply with a feasibility read. We come back with a feasibility read on the workflow and its constraints. Inquiry details are treated as confidential.