AI documentation for medical practices: 14 vendors compared, in their own words

How this table was compiled

On 3 September 2026 we read the vendors' product, price and data protection pages, re-checked them on 4 October 2026, and entered into each cell only what is stated there. There was no test, no weighting and no ranking; the order is not a judgement. The pages read are listed below the table.

Two caveats: the entries for the Doctolib assistant come from its public information page for patients, which states no price. Nabla is not listed because we found no German-language offering.

The overview

The table scrolls sideways. One row per vendor, one column per question a practice should answer before a pilot.

How to read the columns

Audio recording and text are separated because they can be processed separately. A vendor may transcribe the recording at the practice and send only the text to a cloud, or transmit both. A statement like "servers in Germany" usually concerns the text and says nothing about the recording; that is why this column so often reads "not stated".

Platform and practice software determine the workflow: a tool built into the practice's own software needs no transfer step, while a standalone tool is copied into the practice software via an interface or by hand. Both work; the question is which practice software the practice has, and whether it would switch.

The price column names only what is public. A missing price is a vendor decision, not a statement about the amount.

What the table does not say

It says nothing about the quality of the drafts, the waiting time after the consultation, how interruptions are handled, or what ends up in the output. Those are the questions a pilot in your own practice answers. The guide with twelve questions for AI documentation software is the template for that, and the guide on cloud versus local processing explains why the first two columns carry so much weight.

Before a pilot with real consultations begins, settle these points in writing with the vendor you have shortlisted.

1. 1

   Sign a data processing agreementIf the vendor processes patient data, a contract under Art. 28(3) GDPR is required. Among other things it sets out the subject matter, duration, nature and purpose of the processing and the type of data.
2. 2

   Bind the vendor to confidentialityUnder § 203(4) sentence 2 no. 1 of the German Criminal Code (StGB), the practice must ensure that persons assisting it are bound to confidentiality. Check that the contract says so expressly.
3. 3

   Ask for evidence for cloud servicesHave the vendor confirm the place of processing, an establishment in Germany and the current C5 attestation, as § 393 SGB V requires.
4. 4

   Check whether an impact assessment is neededArt. 35 GDPR requires one where a high risk is likely. The German data protection conference writes that this will often be the case with AI applications. The vendor supplies the information on how the system works.
5. 5

   Prepare consent to the recordingRecording the privately spoken word without authorisation is a criminal offence (§ 201(1) no. 1 StGB). Before the first consultation, decide how the practice obtains and records consent.

Where KiPT Voice stands

KiPT Voice processes the recording, its transcription and the speaker attribution on the practice computer; the recording is not transmitted. The transcript text goes to the provider configured in settings, unless the practice sets a local model. The reviewed text is copied into the practice software, the application runs on Windows 11 and on Macs with an Apple chip, and the price after the pilot is in the table. The application is available; the first 14 days are free.

Read on