Averbis

Language processing is the business. Medical Summary, the product the company now leads with, uses natural language processing to pull diagnoses, medications, laboratory values and allergies out of hospital documents onto a patient timeline, and large language models to write summaries, answer questions in a chat and fill yes or no checklists with a cited source. Health Discovery, the older text mining platform, and Docs2FHIR, which turns documents into FHIR data, are the same extraction sold in other forms.

Part of the value does not come from a model: the curated German medical terminology behind the extraction, the connectors into hospital systems, and the timeline and document archive a clinician navigates. Remove the models and a document viewer remains, but little that anyone would buy.

The review steps are documented. Physician letters drafted in Medical Summary move from draft to ready for review, and a named reviewer approves or rejects them with comments; approved versions are locked and later changes become numbered revisions. For recorded conversations between doctor and patient, the manual tells users to check extracted diagnoses and medications for completeness and plausibility before using them, and to mark a conversation as transferred only once it has been.

The controls are guidance rather than enforcement. The manual says letters should be exported only after approval, and conversation results can be sent to Medical Summary and to the hospital information system automatically, depending on setup.

What runs under the hood is visible. Public release notes document every version of Medical Summary (3.6 in May 2025 to 4.15.2 in September 2026) and Health Discovery (6.0 in 2021 to 8.7 in May 2026), including model changes: the August 2026 release made gpt-5-mini the default language model on Azure installations, and Health Discovery 8.0 improved the German models behind its de identification pipeline and restructured the discharge letter pipeline. The architecture is documented too: annotator pipelines, a Solr search index, retrieval augmented chat, and a connector framework.

Averbis's own extraction and de identification models are not described beyond their language and purpose, and their training data is not characterized.

The model layer is named. Release notes say gpt-5-mini is the default language model on Azure installations, and the installation guides for Medical Summary and Medical Dialog require outbound connections to Azure and OpenAI APIs. Averbis states that its cloud language model processing is temporary, without permanent storage, on EU or German servers, from a provider with ISO 27001 and C5. Extraction and de identification run on Averbis's own models inside the hospital.

No subprocessor list is published, the cloud provider behind the GDPR statement is not named, and the training terms that apply to requests are not stated.

University hospitals have published evaluations of earlier versions, each with an Averbis coauthor. At University Hospital Erlangen, Health Discovery's medication extraction was compared with a hand annotated gold standard on 10 discharge letters, finding medication statements with an F-measure of 0.852 and drug names with 0.936 (Studies in Health Technology and Informatics, 2021). At University Medical Center Mainz, the Averbis Extraction Platform classified 1,000 CT reports, agreeing with a second manual rater on 86% (kappa 0.79) against 93% between the two human raters, with errors following a systematic pattern (Gesundheitswesen, 2020). Both results are taken from the abstracts.

The research is small, tied to older product versions, and not independent of the company. Nothing is published on the accuracy of the current language model summaries, chat answers or checklists, and customer quotes describe use rather than outcomes.

The data path is stated plainly. Medical Summary runs inside the hospital's own infrastructure, and Averbis says sensitive health data stays in the local network and is not stored in the cloud. Language model functions call cloud services only for the request at hand, without permanent storage, on servers in the EU or Germany. Chat history is kept for 30 days by default. Health Discovery includes a de identification pipeline built on German language models.

The terms stop short: nothing says whether the cloud model provider may use requests for training, what else is logged, or how long extracted data and saved conversations are kept beyond the chat default. Safety engineering is limited to showing sources and warning users to check results.

Averbis is a German company selling mainly to hospitals in Germany, Austria and Switzerland, so European data protection law is the governing regime rather than US health privacy law. The product page takes a position: data stays inside the hospital's network, and language model calls go to GDPR compliant cloud services in the EU or Germany without permanent storage.

The claim has no document behind it. The published privacy policy covers only the website, and no data processing agreement, description of Averbis's role as a processor, or list of subprocessors is published for the products. No position is stated for US customers.

No security certification or independent audit of Averbis is named anywhere on its site, documentation or announcements. The company describes its approach as maximum data security for regulated environments and names a head of IT responsible for information security, and the documentation shows specific controls such as AES encryption of patient identifiers in links. The ISO 27001 and C5 credentials Averbis mentions belong to its unnamed cloud provider, not to Averbis.

For on-premise installations the hospital's own controls carry much of the load, but Averbis support needs SSH access to production servers.

No medical device clearance or CE marking is claimed in any market, and the products are presented as documentation and data structuring tools for hospitals rather than diagnostic software. The manual keeps clinical judgment with the user, warning that AI can make mistakes and that flagged laboratory deviations must be checked by the clinician.

The boundary is close in places. Medical Summary flags abnormal laboratory values and answers clinical yes-or-no checklists, and in Europe software intended to inform a clinical decision can fall under the Medical Device Regulation depending on its intended purpose. Where the extraction is embedded in another manufacturer's product, that manufacturer carries the regulatory duty.

Averbis publishes some of the limits that matter, but no evaluation of its own. The Medical Summary manual warns that when a document does not mark a laboratory value as abnormal, the software checks it against reference tables that ignore the patient's age and sex, so the user must confirm any flagged deviation; that is a stated gap in exactly the groups where reference ranges differ. The report-writing module runs a second model that checks drafts for consistency and completeness and flags uncertain content, and every checklist answer names the passage it relied on.

Nothing is published on how the language models are tested or monitored, and there are no results by patient group, hospital, document type or language. Outside studies of earlier versions exist, but none compares performance across groups.

Outputs can be checked against their sources. Each checklist answer in Medical Summary names the passage it is based on, chat answers show the relevant passages from the original documents, the report module flags ambiguous or uncertain content, and the manual warns that AI can make mistakes and that important information must be checked. Physician letters drafted in the product go through a documented approval step before export.

Nothing stands behind the output. No warranty, indemnity or correction commitment is published, and Averbis does not cite any accuracy figure for its products. Outside studies with an Averbis co-author report accuracy for earlier versions, but the company does not present them as its own claim.

Medical Summary reads from hospital systems through a documented connector framework: databases, file systems, HL7 and FHIR interfaces, archives, hospital information systems and the national record systems of Germany, Austria and Switzerland (ePA, ELGA and EPD). The FHIR connector accepts bundles including prescribed medications, and results from recorded conversations can be written back into a connected hospital information system. Docs2FHIR turns documents into FHIR resources with ICD codes, and physician letters export as eArztbrief. The hospital information system maker Meierhofer says it is building the extraction into its own system.

No production connection to a named record system is documented beyond that partner statement, and write back is described generically.

The deployment is described in detail. Medical Summary is installed on premise as Docker containers in the hospital's own infrastructure, with separate production and staging systems recommended and published hardware requirements; the documents and extracted data stay in the local network. Language model functions are the exception, sent to a cloud service that Averbis says runs only on servers in the EU or Germany, with no permanent storage, from a provider holding ISO 27001 and a C5 attestation. Release notes also refer to installations on Azure.

The processing path is not fully consistent: the installation guide requires outbound connections to Azure and OpenAI APIs without saying which regions those calls reach, so it is unclear whether any request can leave the EU. Averbis support also needs SSH access to the servers.

No prices are published and there is no longer a free tier on the company's site. How the product is sold can be worked out from the documentation: Medical Summary and Health Discovery are licensed software that each hospital installs on its own servers, activated by a license file and delivered as Docker images on request; Medical Summary adds modules such as report writing and Medical Dialog as needed; language model features run through a cloud service. Software manufacturers such as Meierhofer embed the extraction in their own products.

What drives the price is not stated: document volume, beds, users, modules or language model usage.

Coverage is German speaking hospitals first. Medical Summary serves hospital physicians across departments, with connectors for the national record systems of Germany, Austria and Switzerland; a general practice group uses the extraction to move diagnoses and values from letters into its practice software; and research partners include the Technical University of Munich, the Bosch Health Campus and the Medical Informatics Initiative. For trials, Averbis offers its platform to hospitals for feasibility checks and finding eligible participants for industry sponsored studies, described in a May 2026 article on its site.

Validation covers only part of this: published evaluations cover discharge letters and CT reports at university hospitals. Nothing validates the trial feasibility use, other languages, or the current language model features.