Why aiomics for discharge letters and physician letters

If you want to create physician and discharge letters with AI, quality is decided not by the draft but by what happens between draft and signature. aiomics generates letters in which every sentence points to its source passage in the record — and is building out the verification stage so that omissions, the most common error in AI-generated letters, become visible before anyone signs. Draft generation is live; the build-out to the full verification pipeline is planned — we name precisely which parts of it exist today.

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. The physician letter is the document where this architecture has to prove itself most often: it leaves the hospital, it is read by the next treating physician, and since the same-day obligation for the preliminary discharge letter it is written under time pressure.

At a glance

• What it does: generates drafts for physician and discharge letters from the substantiated record, every sentence with source attribution; export into the hospital's own Word letter template; a derived patient letter from the signed-off document (planned).
• Status: draft generation and Word export live. Omission detector, dress rehearsal, and a software-enforced sign-off chain with versioning that satisfies § 630f (the German Civil Code's medical record-keeping provision): planned.
• Security: ISO 27001 (TÜV Nord), processing exclusively in the EU; no training on customer data (a contractual artifact).
• Integration: KIS-agnostic (KIS — the hospital information system); sending and ePA filing remain with the hospital's own system.
• Evidence: independent evaluation of accuracy at Charité (ongoing).
• Who it is for: physicians, chief physicians, and medical controlling in acute-care and rehabilitation hospitals.

The problem we solve

The evidence on AI-generated letters shows an uncomfortable pattern: the dominant error is the omission. In the CREOLA analysis of clinical language model errors, 3.45 percent omissions stood against 1.47 percent hallucinations; 44 percent of the hallucinations were clinically relevant [1]. An emergency department study at the University of California San Francisco found missing relevant information in 47 percent of AI-generated discharge documents, hallucinations in 42 percent [2]. Anyone who aims their review process solely at hallucinations is checking the smaller error class.

Generation itself has long worked: at Freiburg University Medical Center, 93.1 percent of AI-generated German physician letters proved usable with minimal adjustment [3]. At the same time, the regulatory cadence is tightening — the framework agreement on discharge management makes the preliminary discharge letter a same-day obligation; filing to the ePA (the German electronic patient record) has been mandatory since October 1, 2025, and subject to sanctions since April 1, 2026 [4]. The bottleneck has moved from phrasing to verifying.

How aiomics creates physician and discharge letters

The foundation is live: from the source documents of a case, a letter draft is created in which every sentence knows its origin — whoever checks a passage jumps to the source location in the original. Text without evidence looks visibly different; attention flows to where evidence is thin. The export writes into the hospital's Word letter template via bookmark injection: the letterhead stays pixel-identical, the file opens without a repair dialog.

On top of this we are building the verified pipeline — planned, specified in three phases. We describe it because vendors can be measured against their blueprints:

• The omission detector compares the draft against the structured data of the record and shows in the "Not carried over" panel what the letter leaves out — omitting becomes a documented decision.
• The dress rehearsal reviews the draft adversarially from four perspectives — MD auditor (Medizinischer Dienst, MD — the German payers' medical review service), expert reviewer, recipient, patient — with a quotation and rule reference per finding. "No findings" is a full-fledged result.
• The sign-off chain enforces the review in software: with open markers, finalization is technically impossible; critical changes — a changed discharge dose, say — require an individual substantive confirmation. Every version is documented in a § 630f-proof way; the physician's review becomes provable.
• A style profile learns the hospital's letter voice from the edit history — as a readable, editable, deletable German-language description, without any clinical content.

At every build-out stage, one rule holds: no document is used without a physician's sign-off. This order — Proposed by AI. Verified by You. — cannot be configured away.

What you can measure us against

1. Point at any sentence in the draft: the source pointer leads to the passage in the original document.
2. Upload your own letter template: the export opens in Word without a repair dialog, with your letterhead.
3. Ask about the status of every pipeline component: you get the answer in writing — live, in development, or planned.
4. Watch your physicians' edit rate: a rate near zero we read as a warning sign — blind trust is a safety problem in this product class.

Where aiomics is not the right choice

If all you need is dictation or ambient conversation capture for a practice, specialized tools are the faster route; our approach pays off where letters are created from a complete, substantiated record. If you want to hear a "hallucination-free" promise: we do not make it, because on today's evidence no one can keep it — we give source pointers and verification paths. And if you want letters sent fully automatically without a physician signing them off, you will not get that from us under any configuration.

Frequently asked questions

Is aiomics a medical device?

For document and procedural work, aiomics is deliberately positioned outside the medical device qualification; the delineation is documented and can be inspected. For conversation documentation, we are preparing certification under MDR Class IIa. In both cases, the system's statements remain documentation and quality notes — diagnosis and therapy remain with physicians.

How much time does generation actually save?

The robust evidence speaks of minutes: at Kaiser Permanente, more than 7,000 physicians ended up with around one minute less after-hours documentation time per appointment — with high satisfaction [5]. We think little of promises measured in hours. The larger value lies in completeness, in falling review and correction costs, and in letters that are less often written after hours.

Does the system learn our letter style?

A style profile is planned: it learns from the hospital's edit history and exists as readable German text that you can inspect, change, and delete. Clinical content is never part of the profile; mandatory markers remain an untouchable baseline.

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 see what your last twenty discharge letters would have looked like with source pointers: write to us, and we will show the generation on real cases from your hospital. Why text modules do not solve this problem is described in our article Text modules or contextualized generation; ongoing analysis comes from our weekly briefing Visite (German; English edition Grand Rounds is in preparation).

Sources

1. Asgari E, Montaña-Brown N, Dubois M, et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine. 2025;8(1):274. doi:10.1038/s41746-025-01670-7 (1.47% hallucinations, 3.45% omissions; 44% of hallucinations clinically relevant).
2. Williams CYK, Bains J, Tang T, et al. Evaluating large language models for drafting emergency department encounter summaries. PLOS Digital Health. 2025;4(6):e0000899. doi:10.1371/journal.pdig.0000899 (42% hallucination, 47% missing relevant information).
3. Heilmeyer F, Böhringer D, Reinhard T, et al. Viability of Open Large Language Models for Clinical Documentation in German Health Care: Real-World Model Evaluation Study. JMIR Medical Informatics. 2024;12:e59617. doi:10.2196/59617 (93.1% usable with minimal adjustment).
4. Framework agreement on discharge management under § 39 (1a) SGB V (same-day preliminary discharge letter); ePA filing mandatory since October 1, 2025, sanctions since April 1, 2026.
5. Tierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes: Learnings After 1 Year and Over 2.5 Million Uses. NEJM Catalyst Innovations in Care Delivery. 2025;6(5):CAT.25.0040. doi:10.1056/CAT.25.0040 (7,260 physicians; about one minute less after-hours documentation time per appointment).

Sources retrieved in July 2026.