Ask the vendor to trace a fictional client sentence from microphone to the final matter record in the exact signed build and configuration the firm would install. A promise that “speech is on-device” answers one step. The firm still needs to know whether transcript text reaches an optional AI provider, what the Mac retains, what the app sends for licensing or telemetry, what other apps can read through Accessibility or the clipboard, and what happens when the finished text enters a shared or synced matter system.
The ABA's Formal Opinion 512 addresses lawyers' competence, confidentiality, supervision, communication, and output review when generative AI is used. It does not certify a vendor or impose one questionnaire on every jurisdiction. The public summary of Formal Opinion 477R addresses reasonable safeguards for electronic client communication and cases where stronger precautions are needed. Have firm counsel map these questions to the applicable rules, client agreements, and practice area.
Questions to send with a request for evidence
| Ask the vendor | Evidence a reviewer can inspect |
|---|---|
| Which release and device configuration are you describing? | Signed installer, version, supported macOS and Mac chips, update channel, release notes, and a versioned data-flow diagram. |
| What leaves the Mac in the default mode? | Synthetic network capture covering speech, formatting, account/license checks, telemetry, model downloads, and updates; destination and payload schema for each flow. |
| What changes when a lawyer selects a hosted model? | The exact audio or text sent, instructions sent with it, provider and endpoint, subprocessors, account owner, retention, training, logs, abuse-review terms, and deletion path. |
| Can the firm enforce approved settings? | MDM or other documented control for formatter endpoints, cloud mode, history, and updates; demonstrate that an ordinary user cannot silently select an unapproved provider. |
| What remains locally? | Raw, cleaned, and formatted history fields; temporary recordings; clipboard insertion; app clipboard history; diagnostic logs; file locations; backup effects; deletion behavior after relaunch. |
| What can the app read from a client's active document? | macOS permission list, focused-app/window/field context captured, whether any context enters a prompt or telemetry event, and controls to limit it. |
| What happens when transcription or insertion fails? | Recovery path, retained copies, wrong-window behavior, user notification, diagnostic logs, and a way to report a problem with synthetic content. |
| How are provider keys and account identifiers stored? | Keychain or other storage design, encryption and access controls, token lifetime, device identifier use, and offboarding instructions. |
| Who checks accuracy and the final matter record? | User guidance for names, citations, dates, negations, quoted language, recipient checks, and reviewer handoff; the app should not imply its fluent output is a verified legal record. |
SuprFlow illustrates why the questions need release evidence. Version 1.0.0 uses local WhisperKit recognition and local default cleanup, but Cloud API formatting sends transcript text and instructions to a selected provider or custom endpoint. A configurable Ollama endpoint can be remote. The app saves recent raw, cleaned, and formatted text in local history, temporarily uses the Mac clipboard for ordinary insertion, reads some destination context for mode suggestions, and includes release telemetry and license traffic. A law firm should request release-specific disclosures rather than infer a contractual guarantee from a marketing page.
A synthetic pilot that exposes the real failure modes
Create a test matter called Project Juniper, with no real client data. Dictate: “Casey Vale did not approve clause seven. We have not agreed to an extension. Send the marked draft after partner review.” Use the actual managed Mac and approved matter software. Record the installed app and model, formatter mode, network path, local history setting, clipboard manager, and destination field. Ask a reviewer to compare the final text to the script and to inspect which matter record received it.
This trial tests meaning and routing, not just spelling. The words “did not,” “not agreed,” and “after partner review” are material. The firm should also test a wrong-window interruption, an unavailable hosted endpoint, and clearing a synthetic history item followed by app relaunch. Do not use real client facts until the vendor evidence and firm approval cover the configuration.
Make the decision record usable after an update
Keep the approved signed version, vendor answers and evidence, allowed matter classes, formatter and endpoint, retention settings, destination systems, responsible counsel and IT owner, user training, and stop criteria together. Re-review when a release changes model selection, context capture, telemetry, history, provider terms, or insertion. A per-user toggle is insufficient if the firm requires an enforceable ban on hosted formatting. A source repository is useful evidence if available, but the firm must still evaluate the binary it installs and the contract it accepts.
Questions people ask
Does local speech recognition protect attorney-client privilege by itself?
It limits the speech-to-text transfer path. Privilege and confidentiality still depend on the device, optional formatter, local copies, destination, sharing, vendor terms, and applicable law. Firm counsel makes the matter-specific decision.
Can the vendor promise client dictation is never used for AI training?
Only with the scope stated. SuprFlow's local path does not submit dictation to a training service, but a lawyer can opt into a hosted formatter. The chosen provider's account and contract govern text submitted there. Ask about telemetry and error strings separately; “no model training” does not mean “no retention.”
Is a privacy policy enough evidence?
It is necessary public documentation, but the firm also needs a configuration-specific signed-build test and contract review. Where policy, FAQ, onboarding, and observed behavior disagree, resolve the conflict before using client material.
Related guides
- Offline dictation for lawyers
- Dictate a client-call attendance note
- Does SuprFlow use dictation for AI training?
- Where does my dictation go?
- Can dictation appear in Mac diagnostic logs?
- Bank dictation vendor checklist