An adviser finishes a client call and wants to dictate a summary before the next meeting. A bank or investment firm may want that convenience without sending client details to an unapproved speech service. Local speech recognition is one useful control, but procurement still has to answer where the resulting text goes, what is retained on the device, and how the final note enters the firm's record and supervision workflow.
In SuprFlow 1.0.0, WhisperKit recognizes speech on an Apple Silicon Mac and the default cleanup runs locally. The user can also select a hosted AI formatter that sends transcript text to the chosen provider. The app retains recent transcriptions locally and release builds include telemetry. Therefore, an institution should approve a specific configuration and signed build, not a broad “offline” label.
Separate four data questions
1. Is audio sent away for speech recognition? SuprFlow's speech recognition uses a downloaded model on the Mac. Test the signed release with synthetic speech and record the model and version.
2. Is transcript text sent away for formatting? The default cleanup is local. Cloud API formatting sends the transcript and formatting prompt to a selected provider or custom endpoint. Ollama is local only when its configured endpoint resolves to an approved local server. Treat any remote endpoint as a separate provider review.
3. What stays on the device? SuprFlow saves up to 100 recent transcription entries in a local Application Support file; clipboard history is another configurable store. The firm should decide whether these files are permitted on managed laptops, how backups and device loss are handled, and whether a user can clear a record that the firm is required to retain elsewhere.
4. What other network traffic exists? Account and license checks contact SuprFlow, and release builds initialize usage telemetry. That traffic is not the same as transmitting the dictated text, but a security review still needs the destination and payload. Model downloads and updates are separate events.
Test the actual client-note workflow
Create a fictional household and synthetic balances. Ask a pilot user to dictate a meeting summary into the exact approved CRM or document system. Include names that resemble real client names, a date, a percentage, a monetary amount, a negation, and an action item. Check how often a person must correct the result and whether formatting changes the meaning. Check the document's access list, sync path, retention label, and review status after insertion. Dictation does not itself create a supervised or complete client record.
For CFA charterholders and candidates, CFA Institute Standard III(E) calls for preserving client confidentiality and following the employer's electronic-information procedures. It does not endorse a specific tool. For US FINRA member firms, FINRA's 2026 GenAI discussion describes supervision, governance, testing, privacy, and recordkeeping concerns when firms deploy GenAI. Bank and investment-firm requirements vary by jurisdiction and business; the relevant compliance owner should determine which rules apply.
What an approval record should contain
| Evidence | Why a firm needs it |
|---|---|
| Signed app build, source/version match, macOS and model versions | A general description of the product is not proof of a particular installed binary. |
| Allowed formatter modes and enforced endpoint policy | A user may otherwise switch from local cleanup to a hosted provider. |
| Provider contract and training/retention terms, if hosted formatting is allowed | A SuprFlow claim cannot govern another provider's handling of submitted text. |
| Telemetry destination and payload review | “No speech upload” does not answer whether operational metadata or errors are sent. |
| Device storage, history and clipboard controls | Local-only data still needs access, backup, loss, and retention rules. |
| Approved destination apps, record capture and human review | The final note must reach the right record, with corrected figures and accountable approval. |
| User training and incident path | Staff need to know how to recognize a wrong field, mistaken amount, or unapproved cloud mode. |
Does on-device recognition mean the app is approved for customer data?
No. It describes the speech-recognition step. The firm must still review optional formatting, local history, telemetry, license traffic, destination systems, and its own policies. Approval is an institutional decision.
Can we promise client notes are never used for AI training?
The local recognition and default cleanup path in SuprFlow 1.0.0 does not submit dictated words to a hosted model. If staff enable hosted formatting, provider-specific terms apply. A firm that requires a no-training commitment should verify the exact provider agreement or keep the text path on an approved local endpoint. Do not infer a universal guarantee from a single product setting.
Does a local transcript count as the firm's required record?
That depends on the firm's recordkeeping rules and approved systems. SuprFlow history is a convenience copy, not a substitute for an appropriately captured and supervised client record. Confirm the destination and retention process with the firm's compliance owner.
Related guides
- Private voice dictation for bank client notes
- Offline voice to text on Mac: compare local options
- Where does my dictation go?
- Does SuprFlow use dictation for AI training?
- Set up SuprFlow on Mac
- CFA client notes and offline dictation
- Bank AI dictation vendor checklist
- Dictate an adviser CRM meeting note