Before using voice typing for client notes, test whether the installed configuration preserves figures that matter: dollar amounts, percentages, dates, units, and negative statements. A single mistake from “$12,400” to “$124,000” can change the record more than a dozen spelling errors. Use fictional data, keep the same speech script and destination for every trial, and have a human compare the final inserted text with the script. Do not publish an “accuracy rate” from a few examples.
This is a workflow test, not an investment recommendation. CFA Institute Standard V(A) calls for diligence and a reasonable basis in investment analysis and recommendations, including reasonable inquiries into data accuracy. Standard V(B) addresses client communication and distinguishing fact from opinion. The standards do not set a dictation-app benchmark; they explain why an adviser should verify material wording before relying on or communicating a note.
A fictional script worth testing
Read this unchanged script in a quiet room, using a test client record:
I met the Meridian Sample household on 9 September. Their example balance is $124,000, and the proposed example withdrawal is $12,400, or 10% of that balance. The household asked whether a target allocation of 60% equities and 40% bonds should change. I did not recommend a trade. They have not authorized a withdrawal. I will verify the current statement and follow up after the firm's review.
The figures are deliberately related. A formatter could “helpfully” recalculate, round, or merge them. The result must preserve $124,000, $12,400, 10%, 60%, 40%, the date, and both negations. It must also keep a question separate from a recommendation or instruction. Those are more meaningful checks than whether the paragraph sounds polished.
Keep the trial repeatable
- Record the configuration. Note the signed SuprFlow version, Mac and macOS, microphone, speech model, selected language, formatting mode, endpoint if any, and destination app. SuprFlow uses a local WhisperKit speech model, local default cleanup, and optional hosted or configurable Ollama formatting. The formatter choice can change the text as well as the data path.
- Use a permitted test destination. A blank local note or test CRM account avoids real client information. Confirm the field has focus. Ordinary SuprFlow insertion briefly uses the clipboard, so a clipboard manager can retain even the test text.
- Read the exact script three times. Use your normal speaking pace and accent. Do not correct the script between attempts. If testing different models or modes, use the same script, microphone, room, and destination and record each result separately.
- Compare the final inserted text. The reader sees the pasted text, not merely the raw speech output. For each trial, mark whether each material value and negation is exact, omitted, substituted, or added. Also mark whether “asked whether” becomes “decided” or “did not recommend” becomes an affirmative recommendation.
- Correct by source, not by plausibility. Compare the draft with the meeting record or approved source. A value that looks financially plausible may still be wrong. If you cannot verify it, mark it unresolved instead of letting a formatter invent a number.
- Decide what the pilot permits. Keep the trial outputs and configuration record in an approved test location. If a material error recurs, change the model, microphone, prompt, or workflow one variable at a time. Require human review of figures in all real notes even if the synthetic test performs well.
| Material field | Script value | What to flag in the output |
|---|---|---|
| Balance | $124,000 | Missing zero, extra zero, wrong currency, or value swapped with withdrawal. |
| Withdrawal | $12,400 | $124,000, $1,240, a rounded number, or a completed instruction. |
| Portion | 10% | 10 basis points, 0.10%, or an invented recalculation. |
| Allocation | 60% equities / 40% bonds | Reversed percentages or a recommendation to change them. |
| Consent | Did not recommend; has not authorized | Missing “not,” added approval, or a question rewritten as an instruction. |
The test deliberately combines transcription and formatting. If the final text changes meaning, compare the locally saved raw, cleaned, and formatted versions where available under firm policy to isolate where it changed. Those local copies themselves need a retention decision. The final note belongs in the firm's approved CRM or records system, not in a dictation utility's history. CFA Standard V(C) addresses records supporting analyses and client communications; the firm determines the official system and applicable requirements.
Questions people ask
Is three clean test runs enough to prove dictation is accurate?
No. Three trials can expose a problem, but they cannot establish a general accuracy rate. Repeat with different speakers, microphones, noise levels, accents, models, and representative fictional terms under a documented pilot. Keep human review for real figures and instructions.
Should an AI formatter correct a spoken percentage?
Only if the workflow explicitly authorizes a calculation and a human verifies its inputs and result. For a client note, preserving what was said and what was decided is often more important than producing an apparently improved figure. Do not let a rewrite silently turn a question into advice.
Can we test using actual client meetings?
Start with synthetic examples. Real client content should be used only after the firm approves the signed build, formatter and endpoint, retention, telemetry, destination, and supervision workflow. See the bank vendor checklist.