Glossary: the key terms of AI-assisted practice documentation

How to read this glossary

Each of the following sections covers a single term as it appears in descriptions of AI documentation software. The first sentence defines the term on its own, without assuming anything from an earlier section. A second paragraph places it, where useful, in the everyday routine of a German medical practice.

This glossary does not replace legal advice or a product review. It sorts out terms that commonly get mixed up in conversations with vendors and in data protection documents.

Ambient scribe

An ambient scribe is a system that listens to a conversation in the room and automatically produces a structured written version of it, without anyone taking notes during the conversation. The term describes the task, not a specific product.

For a practice this means: the physician conducts the consultation as usual, a running device records it, and afterwards a draft is ready for review. How much of this runs locally on the practice computer and how much at an external vendor differs by system, and is one of the first questions to ask a vendor.

Automatic speech recognition (ASR)

Automatic speech recognition refers to the software-based conversion of spoken language into written text, without a person typing it. The abbreviation ASR is more common in technical descriptions than the spelled-out term.

Speech recognition delivers the wording, not a structure by topic and not an attribution of who spoke. Both are separate, later steps.

Speaker attribution (Diarisierung)

Speaker attribution, also called diarisation in technical documents, refers to the step in which a system determines which part of a conversation came from which person present, for example the physician or the patient. It works on the already recognised text or in parallel with the audio track, not on the content of what was said.

A speaker attribution can be wrong, for instance when two people speak at once or a third person briefly enters the room. A reviewed draft should therefore make clear where an attribution was uncertain.

Transcript

A transcript is the written version of a spoken conversation, kept as close to word-for-word as possible, including filler words, repetitions and asides. It is an intermediate stage, not a finished document for the record.

A transcript contains everything that was said, including what does not later belong in the documentation. Whoever keeps a transcript keeps the same amount of information as in the recording itself, just as text.

Documentation draft

A documentation draft refers to the text produced from the transcript, ordered by topic, that a physician reads, checks and amends where needed before carrying it into the patient record. It is the result of a further processing step, not the transcript itself.

The difference matters in practice: reading a transcript takes as long as the conversation did, while a draft is condensed to the essentials. What should be approved is always the reviewed draft, never the raw text unchecked.

Large language model (LLM)

A large language model, or LLM, refers to a software system that has learned from large amounts of text to understand text and generate new text. In practice documentation, such a model is used to formulate an ordered draft from the transcript text.

Which vendor operates this model, and where the transcript text is processed for it, is a separate question that should be answered in the vendor's own data protection documents.

Hallucination

Hallucination, in this context, refers to a language model formulating a statement in the generated text that was not actually made in the underlying conversation. The term describes an error in text generation, not a statement about the content of the conversation itself.

For a practice this means: a draft is a proposal, not a transcript record. It must be checked against what was actually said before it is carried into the record, and must not be accepted unchecked.

Local processing

Local processing refers to a given processing step taking place on the device at the practice, rather than sending data to an external server for it. The term always refers to a specific step, not automatically to the whole system.

A recording can stay local while the transcript text for the draft goes to an externally operated language model. Anyone who hears "local" as a blanket property of a product should ask exactly which step is meant.

Hybrid processing

Hybrid processing refers to a combination in which part of the processing takes place locally on the practice computer and another part at an external vendor, for example when the recording stays at the practice and only the transcript text derived from it is sent to a language model.

Vendor descriptions rarely name this path as "hybrid", even though it is common. A practice should call it what it is: partly local, partly external, with two separate questions about the data path.

Data processing agreement (Auftragsverarbeitung, AVV)

A data processing agreement refers to the processing of personal data by an external service provider on behalf of the practice, governed by a contract under Art. 28 GDPR, commonly abbreviated AVV in German. The practice remains legally responsible even when another company technically carries out the processing.

As soon as transcript text or a recording goes to an external vendor, the practice needs such a contract with that vendor. Without it, the legal basis for the disclosure is missing, regardless of how reliable the vendor otherwise appears.

Pseudonymisation

Pseudonymisation refers to a method in which identifying details such as a name are replaced with a marker, so that attribution to the person concerned is possible only with additional, separately kept information. It is named in the GDPR as a measure, not as a state of full anonymity.

A pseudonymised transcript therefore remains personal data for as long as the attribution information exists anywhere. This is what distinguishes pseudonymisation from genuine anonymisation, where that reversal is permanently excluded.

Consent to the audio recording (Einwilligung)

Consent to the audio recording refers to the prior, documented agreement of the people present in the room that their spoken words are recorded as audio. Without this agreement, recording non-public spoken words is a criminal offence under § 201 StGB (German Criminal Code).

For a practice this means: consent must be obtained before the recording starts and recorded in a traceable way, not mentioned in passing or assumed after the fact.

Duty to document (Dokumentationspflicht)

Duty to document refers to the legal obligation of a treating person to record the essential steps and results of a treatment in the patient record promptly, set out in § 630f BGB (German Civil Code). It applies regardless of which tool is used to produce the documentation.

A draft produced with a language model can support this duty but does not replace review by the treating person: they remain responsible for the content of the record, not the system that formulated the draft.

Practice management system (PVS)

A practice management system, abbreviated PVS in German, refers to the software in which a practice manages patient records, appointments and day-to-day operations. A documentation draft from a separate system must be carried into this practice management system to become part of the record.

Whether this transfer happens by hand, copied over manually, or through an interface, depends on the tool and the practice management system in use. Direct access to a third-party practice management system requires its own technical interface; the question therefore belongs in every vendor conversation.

Read on