FlowDule uses AI for documentation assistance in the record and for translation. The model is Anthropic Claude via Amazon Bedrock in the EU, and speech, images and support cases are processed with Amazon Transcribe, Amazon Textract and Amazon Translate. AI does not make professional decisions. Most results are drafts that the practitioner must review before they enter the record or are shared. Your clients’ data is not used to train AI models.
This article is for you if you administer a clinic or a chain. The text your clients read is set out in AI information.
Who has access, and how you switch AI off
The AI features in the record are available to the chains that have switched the journal module on, and to the chain’s users according to their role. There is currently no activation step for each individual feature: when the journal module is switched on, the features are available, and note feedback and record cleanup run automatically in the background. Translation of the chain’s own texts is available to administrators whether or not the journal module is switched on.
You cannot switch AI off yourself and keep the journal module. If you want AI switched off, you ask for it at support@flowdule.com, and FlowDule then switches it off for the chain. This covers the AI features in the record, title suggestions, the template tools and translation of free text. It does not cover translation of shared content for the client and translation of the chain’s catalogue, templates, policies and message templates; these can currently only be stopped for the whole platform.
Which features are there
- Formatting and proofreading: the practitioner can have their note corrected and structured according to a template, for example proofreading or language improvement.
- Summary: a summary of a note with keywords and suggested next steps.
- Dictation: audio recorded by the practitioner is converted into text.
- Text from image: text in a photo of a document or handwriting is recognised and inserted into the note.
- Report drafts: a draft report based on a report template and the notes the practitioner selects.
- Professional insights: at the practitioner’s request, a reflective analysis of the practitioner’s own courses of treatment over a selected period.
- Note feedback and learning profile: finalised and corrected notes are scored automatically in the background, so that the practitioner can get feedback on their notes. Only figures and fixed codes are stored, that is measures of how the note is written and codes for the interventions and themes the note deals with, not text. The results are combined into a learning profile for each practitioner, which only the practitioner can see.
- Record cleanup: finalised notes are reviewed automatically in the background, and a cleaned-up version is proposed if the note contains, for example, interpretation presented as fact, judgemental language, identifying information about other people, speculation, internal remarks or irrelevant sensitive information.
- Translation of shared content for the client: a client can have a note or a report that the practitioner has shared translated into their own language. The translation is machine-made and is shown to the client without a practitioner having reviewed it.
- Translation of the chain’s own texts: administrators can have, for example, the catalogue, templates, policies, message templates and booking texts translated.
- Title suggestions and template tools: suggestions for a note title, and tools with which administrators can test and improve a template’s instruction to the AI.
- Translation of support cases and platform announcements: messages in support cases and FlowDule’s announcements on the platform are translated into the recipient’s language.
What AI may not be used for
- independent diagnosis, triage or risk scoring;
- selection, alteration or recommendation of treatment;
- decisions on access to treatment, a service or any other individual right;
- automatic completion or signing of records without active human approval;
- hidden profiling, emotion recognition or biometric categorisation;
- research, marketing, model training or general product development on client or record data;
- reuse of output for other customers or for purposes other than your own instructions.
The human assessment
AI output that is to be used in the record or towards a client is a draft. Before use, a competent user must:
- check that the material relates to the right person and case;
- compare the output with the original text, audio or other source material;
- review facts, negations, figures, dates, names, medical or professional terms and material omissions;
- correct or reject uncertain or misleading content;
- take independent responsibility for the final text and the professional action;
- save, approve or share the result only once it has been reviewed.
This is how review works in FlowDule:
| Feature | Review |
|---|---|
| Formatting and proofreading | The draft is shown in the writing tool and only becomes part of the record when the practitioner saves. The practitioner can view the changes compared with their own notes. |
| Summary | The summary, keywords and next steps are written on the note when the practitioner has asked for them. The practitioner must read them and can correct or remove them. |
| Dictation | The transcript is inserted into the writing tool, and the text of the note only changes when the practitioner saves. If the dictation is made to a note that has already been created, the raw transcript is stored with the note as soon as it is ready. The practitioner reads and corrects the text like any other text. |
| Text from image | The recognised text is inserted into the writing tool and only becomes part of the record when the practitioner saves. The image and the text are not shown side by side, so the practitioner must compare them. |
| Report drafts | The report must be approved before it can be shared with a colleague or shown to the client. |
| Professional insights | The analysis is never written into the record and is shown only to the practitioner. |
| Note feedback and learning profile | No approval step. Scoring takes place automatically in the background when a note is finalised or corrected, and nothing is written into the record. |
| Record cleanup | The proposal does not change the note. Only when the author applies the proposal is a new version of the note saved; the author can also reject it. |
| Translation of shared content for the client | No approval step. The translation is machine-made and is shown to the client without a practitioner having reviewed it. Only the original content has been reviewed by the practitioner. |
| Translation of the chain’s own texts | The administrator sees the translation and chooses whether to save it. |
| Title suggestions and template tools | The user chooses whether to use the suggestion. |
| Translation of support cases and platform announcements | No approval step. The message is translated automatically into the recipient’s language; the original message is kept. |
Roles and data
You are the data controller for your processing of client and record data and determine the basis for processing under Article 6 GDPR and, where relevant, Article 9, as well as any national legal basis. FlowDule processes the data as a processor under the Data Processing Agreement and the chain’s settings in the platform, which together are your instructions to us.
Depending on the feature, the processing may cover record text, free text, document images, audio, transcription, prompt, output and necessary technical metadata. Your users must limit input to what is necessary for the purpose.
FlowDule must not use your client or record data for its own AI purposes, for training general models, for marketing or for analysis across customers.
Suppliers, region and model training
The AI features use AWS services under FlowDule’s agreement with AWS. The model is Anthropic Claude, more precisely Claude Sonnet and Claude Haiku, which AWS makes available through Amazon Bedrock; FlowDule has no direct agreement with Anthropic.
| Feature | Service | Data processed | Region |
|---|---|---|---|
| Formatting and proofreading, summary, report drafts, professional insights, note feedback, record cleanup, translation of shared content and of the chain’s own texts, title suggestions and template tools | Amazon Bedrock with Anthropic Claude | Note text, selected notes, the template’s instruction, words from the chain’s glossary, text to be translated, and output | EU |
| Dictation | Amazon Transcribe | Audio and transcription | EU |
| Text from image | Amazon Textract | Image of a document or handwriting and recognised text | EU |
| Translation of support cases and platform announcements | Amazon Translate | Message text and translation | EU |
The calls to Amazon Bedrock are processed in one of the EU regions the configuration covers. The applicable list of suppliers is available in Sub-processors and suppliers.
Your data is not used to train AI models. FlowDule does not train models on your data, and under AWS’s terms for Amazon Bedrock, prompts and output are not used to train the models.
Prompts and results from formatting, summaries, dictation and text from image are stored encrypted at FlowDule, so that the practitioner can review the result. The personal data in them is removed automatically after 90 days. Text recognised from an image that is not linked to a note is deleted after 7 days. Machine translations of shared content are stored encrypted, so that the same translation does not have to be made again. Note feedback stores only figures and fixed codes. What the practitioner saves in the record follows the chain’s retention rules, see Retention and deletion for the clinic.
Quality, limitations and errors
AI can hallucinate, misunderstand a context, overlook negations, confuse people, distort figures or dates, translate professional terms incorrectly and reproduce bias from the model or the source material. A fluent and convincing answer is not in itself a correct answer.
FlowDule reviews proofreading, language improvement and summaries against a fixed set of fictitious examples. The other features do not have a comparable review, and the practitioner’s own assessment is therefore particularly important. Serious or repeated errors may mean that FlowDule immediately switches the affected feature or model off, for one chain or for all chains.
Errors in AI can be reported to support@flowdule.com. Security incidents and possible exposure of personal data are reported immediately to security@flowdule.com.
What you must tell your clients
You must assess for yourself when a client is to be informed about the use of AI under the GDPR, professional rules and the AI Act. Your local text must at a minimum give an accurate account of:
- the purpose for which AI is used;
- what information is processed;
- which results a practitioner reviews, and that translation of shared content is machine-made and not reviewed by a practitioner;
- relevant suppliers, regions, retention and transfers;
- how the client can contact you about rights or questions.
If a person interacts directly with an AI system, that person must be informed of this, unless it is obvious in the context. FlowDule has no AI chatbot. When a client chooses to see shared content in another language, the client portal states that the translation is made automatically with AI, and links to AI information.
Automated decisions and profiling
FlowDule’s AI features do not make decisions based solely on automated processing that have legal effect or similarly significant effect for a person. You must not circumvent this by making an AI recommendation effectively binding without a genuine, competent and independent human assessment.
Note feedback and the learning profile are profiling of the practitioner within the meaning of Article 4(4) GDPR: finalised and corrected notes are scored automatically in the background without the practitioner starting it, and the results are combined into a profile of the practitioner’s notes. The purpose is to give the practitioner feedback. The profile is shown only to the practitioner, administrators cannot see it, and it is not used for decisions about the practitioner. As controller, you must inform your practitioners about that processing.
If a feature could affect diagnosis, treatment, access to services, insurance, employment or other significant matters, it must be assessed separately under Article 22 GDPR, the AI Act and any rules on medical devices before it is brought into use.
AI skills among your users
Access to the AI features is controlled by the user’s role in the chain. FlowDule does not follow up on whether the individual user has been through training. You must ensure that the users who use the AI features have sufficient AI skills, see Article 4 of the AI Act. This covers at a minimum:
- the feature’s intended and prohibited use;
- hallucination, bias and the assessment of sources;
- data minimisation and the handling of special categories;
- human approval and professional responsibility;
- the reporting of errors, incidents and security breaches.
Questions from clients
If a client is in doubt about their record, about access, rectification, erasure or the basis for processing, the client must ask you. We assist you on your instructions. If the question concerns FlowDule’s own processing, they can write to privacy@flowdule.com. See also the Privacy Policy and Sub-processors and suppliers.
Need help?
Under Support in the app you can open a ticket to the FlowDule team. You can also write to support@flowdule.com.