Why aiomics for documenting the admission conversation

If you want to document the admission conversation automatically, the decisive question is not how many conversation settings a system advertises, but how reliable it is in the one it has mastered. aiomics records the admission conversation with documented consent, transcribes on self-hosted infrastructure in the EU, and produces draft summaries in which every sentence points to its place in the transcript. The scope is deliberately narrow — admission conversation, individual patients, desktop microphone — and follows from the category's error data.

aiomics is the verification layer on top of hospital IT: the system ingests unstructured documents, verifies every statement against its source, and gives the hospital a substantiated, structured record. Conversation documentation makes the spoken word one of those sources: what was said becomes citable afterwards — with speaker attribution, timestamps, and confidence per segment.

At a glance

• What it does: records the admission conversation with documented consent, provides speaker attribution, timestamps, and confidence per segment, and produces draft summaries with sentence-level transcript links and consistency flags.
• Status: live in production, deliberately narrow in scope — admission conversation, individual patients, desktop microphone. Extensions (ward rounds, discharge and informed-consent conversations): planned. MDR Class IIa: in preparation.
• Security: ISO 27001 (TÜV Nord), processing exclusively in the EU; no training on production audio, recordings deleted after 72 hours with cryptographic proof.
• Integration: KIS-agnostic (KIS — the hospital information system); the summary remains a draft until physician sign-off, and transmission to payers remains with the hospital's own system.
• Evidence: independent evaluation of accuracy at Charité (ongoing).
• Who it is for: admitting physicians and residents in acute-care and rehabilitation hospitals.

The problem we solve

Ambient documentation has a well-measured error profile. A comparative study across five scribe platforms found a mean error rate of 26.3 percent in the generated notes; around three quarters of those errors were omissions, and almost one in three notes contained at least one fabricated statement [1]. The quieter error class is speaker attribution: whether "no pressure in the chest" comes from the patient or from the physician thinking aloud decides its meaning; attribution mix-ups are documented as an error class of their own [2].

And the errors grow with the scene. In a widely used benchmark, the diarization error rate more than doubles when four people speak instead of two; room microphones score three to four percentage points worse than close-talking ones [3]. Whoever buys a scribe system is therefore above all buying a scope — and should know which one.

How aiomics documents the admission conversation

Four decisions shape the product — each of them can be verified on the system.

First: a narrow frame in which the error rates hold. One conversation, two speakers, one desktop microphone at close range — in this constellation, transcription and speaker attribution are dependable. Where the room situation demands more, we provide devices with multiple microphones. The extension to ward rounds, discharge, and informed-consent conversations is planned.

Second: speaker attribution that reads the content along. Purely acoustic attribution reaches word diarization error rates of around 15.8 percent in doctor-patient conversations; with linguistic content signals, the figure drops to 2.2 percent in the same study [4]. Our attribution works content-based; segments below 85 percent confidence carry a visible marker.

Third: negation discipline. Negating segments — no known allergy, no prior event, medication discontinued — appear in the summary with the original quote from the transcript and are measured in a dedicated benchmark. It is the error class with the highest potential for harm.

Fourth: data sovereignty as architecture. Transcription runs on self-hosted infrastructure in the EU; no audio leaves this processing, and the deployed model version is fixed in the technical bill of materials. Recordings are deleted after 72 hours by default — with cryptographic proof.

On top of this sits a defined emergency stop: if the hallucination rate in the ongoing audit exceeds two percent, the feature is paused. The summary remains a draft; consistency flags such as "these two statements do not fit together" do the marking, the assessment remains the physician's, and sign-off cannot be bypassed.

What you can measure us against

1. Every sentence of the summary jumps to its place in the transcript on click — or it does not appear.
2. Look for the negations: they appear in the draft as original quotes, with a jump link into the transcript.
3. Request the deletion proof for a recording: cryptographic, after the 72 hours have elapsed.
4. Check the speaker attribution: segments below 85 percent confidence are visibly marked.

Where aiomics is not the right choice

For an office-based practice looking for a broad general-purpose scribe covering many conversation types and fast dictation, specialized practice scribes such as Tandem are the better-suited choice today. Our product is built around the inpatient, substantiated record; its value lies in combination with it. Anyone who wants to document ward rounds or discharge conversations today will find that with us only in the planned extension. And anyone expecting voice or affect analysis will not get it, as a matter of principle: we document what was said.

Frequently asked questions

Is aiomics a medical device?

For conversation documentation, we are preparing certification under MDR Class IIa. The Swedish Medical Products Agency, in an ongoing procedure, holds the preliminary position that AI scribes with a summarization function require at least Class IIa; in Germany there has been no corresponding determination so far [5]. It is to be expected that this line will reach the German market as well — we are preparing for it. The rest of aiomics's document and procedural work is deliberately positioned outside the medical device qualification; the delineation is documented and can be inspected. In both cases, the system's statements remain documentation and quality notes — diagnosis and therapy remain with physicians.

What happens to the audio recordings?

Consent is documented per conversation. Processing takes place exclusively in the EU, production audio is never used for training, and after 72 hours the default deletion takes effect, with cryptographic proof.

Does this also work for ward rounds with several people?

Not yet, and that is a decision: diarization errors grow markedly with the number of speakers [3]. We release conversation settings only once we know their error rates and can stand behind them — the extensions are planned.

What does aiomics cost?

Pricing is usage-based and depends on document volume and quality. We name concrete figures after a short conversation about your case volumes — quoting flat prices without that basis would not be serious.

If you want to hear what an admission conversation from your hospital looks like as substantiated documentation: write to us — we will show it on a real conversation with documented consent. Ongoing analysis of hospital AI and regulation comes from our weekly briefing Visite (German; English edition Grand Rounds is in preparation).

Sources

1. Mayo Clinic Proceedings: Digital Health, 2025. Five ambient scribe platforms: mean note error rate 26.3%, omissions around three quarters of all errors, fabricated statements in 31% of notes.
2. npj Digital Medicine 2025;8:569 (speaker attribution mix-ups as a documented error class in clinical speech recognition).
3. NVIDIA, model card diar_sortformer_4spk-v1 (CALLHOME: DER 5.85% with two, 12.59% with four speakers); pyannote benchmarks on AMI (headset vs. room microphone).
4. El Shafey L, Soltau H, Shafran I. Joint Speech Recognition and Speaker Diarization via Sequence Transduction. Interspeech 2019 (word diarization error rate 15.8% acoustic, 2.2% with content signals).
5. Läkemedelsverket assessment of AI scribes (Läkartidningen, March 2026); MDR Annex VIII, Rule 11; MDCG 2019-11 Rev. 1; Luckner, Lauer. Bundesgesundheitsblatt 2025;68(8):854–861.

Sources retrieved in July 2026. MDR Class IIa certification is in preparation and has not been achieved.