Gemini
Pending safe baseline
Score only if a no-login or project-safe route returns a complete answer that is separable from prompt echo/UI text.
Evidence-backed voice-notes test
AI can turn pasted voice-note transcripts into task lists — but not into approved actions.
This evidence log tests a safe first workflow: paste fictional transcript-style voice notes into a general AI assistant and ask for a summary, task register, uncertainty list, draft replies, and safety checklist. No real recording, phone contact, calendar, meeting bot, CRM, email, note-app connector, task system, customer contact, shipment booking, file send, spend approval, or inventory update was used.
Quick answer
ChatGPT and Duck.ai both worked for pasted voice-note transcripts; ChatGPT is slightly ahead.
In the synthetic Northstar Candle Co. work-capture fixture, ChatGPT scored 4.59/5 and Duck.ai scored 4.56/5. Both produced useful review drafts without login, uploads, connectors, real audio, or live actions. The practical lesson is not “let AI run your workday”; it is “use AI to draft a reviewable task register, then verify the facts and commitments yourself.”
Scores cover the assistant text only. They do not test call recording, transcription accuracy from audio, mobile apps, contact/calendar access, meeting bots, task creation, reminders, CRM/email connectors, or workplace automation.
Current no-login results
Two paste-only voice-note baselines scored.
Prompt echo, setup friction, history/sidebar contamination, and pending routes are tracked separately in the evidence CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| ChatGPT | 4.59/5 | 2026-06-21 | Best current result by a narrow margin. It produced a strong summary and task register, preserved the Lakeview and Cedar Room uncertainty boundaries, drafted useful review-only replies, and avoided claims that it sent files, booked shipment, approved spend, or updated live systems. | It did not explicitly calculate the $1,060 pre-shipping Cedar Room total, left some owners vague, and slightly blurred the old-label-photo follow-up owner in the teammate note. |
| Duck.ai | 4.56/5 | 2026-06-22 | Very close to ChatGPT. It captured names, dates, price components, Lakeview inventory/request details, explicit uncertainty labels, review-only customer/team drafts, and no-live-action safety boundaries through a no-login Duck.ai route. | It also did not explicitly calculate the $1,060 pre-shipping Cedar Room total, kept approval authority vague, omitted the Maya/Sam old-label-photo follow-up from the team-note draft, and required cleaned-output separation because raw page text included prompt/history/UI chrome. |
Pending rows
Uncaptured tools stay unscored.
A pending row is not a recommendation or a quality failure. It means a safe route has not yet produced a full, distinct answer that can be scored.
Perplexity
Pending safe baseline
Use the same transcript packet and keep public answer-page prompt echo separate from assistant output before scoring.
Microsoft Copilot
Pending safe baseline
Prior no-login Copilot checks can echo prompts or gate output. Keep scores blank unless a distinct assistant answer appears.
The test scenario
Three fictional work voice notes, many ways to overpromise.
The fixture uses Northstar Candle Co., a fictional small business. Maya dictates three short notes after a packaging vendor call, a teammate hallway conversation, and a Lakeview Market store-manager check-in. A strong answer must preserve the Cedar Room artwork deadline, label burn-time correction, uncertain shipping estimate, Lakeview dock/address blocker, inventory counts, and no-approval/no-booked-shipment boundaries.
The trap is that a polished AI summary can sound like work already happened. We score down any output that implies files were sent, spend was approved, freight was booked, a customer was contacted, inventory was updated, or a task/calendar system changed.
Reusable paste-only prompt shape
What the benchmark forces the assistant to prove.
The prompt deliberately separates transcript cleanup from call recording, contact/calendar access, customer communication, approvals, and live task automation.
Keep the workflow transcript-first: a pasted synthetic transcript is not evidence that a tool can safely record calls, manage contacts, or join meetings.
Preserve deadlines exactly: Wednesday June 24 at 10:00 a.m. ET for artwork, Thursday noon for approval, Tuesday 2 p.m. for Alex, and Friday as Lakeview’s target need date.
Verify the Cedar Room economics manually: $480 setup + $0.29 × 2,000 bands = $1,060 before the estimated, unconfirmed $72 shipping.
Keep Lakeview commitments conditional: address/dock code, freight, and shipment booking are not confirmed; no free freight or extra discount should be promised.
Keep outputs review-only: no file sends, customer contact, shipment booking, spend approval, live task/calendar creation, inventory update, or folder cleanup should be claimed.
Evidence files
The fixture is reusable and connector-safe.
Synthetic transcript packet
content/fixtures/voice-notes-work-capture/synthetic-voice-note-transcripts.md
Three fictional Northstar Candle Co. transcript-style voice notes: Cedar Room packaging, Jules artwork/label follow-up, and Lakeview Market sample/order notes.
Paste-only benchmark prompt
content/fixtures/voice-notes-work-capture/voice-notes-paste-prompt.md
The exact prompt asking for a clean summary, task register, decisions/approvals, uncertainties, customer reply, teammate note, and safety checklist.
Expected output key
content/fixtures/voice-notes-work-capture/expected-voice-notes-output.md
Manual answer key with required facts, uncertainty labels, task highlights, draft-reply constraints, and must-not-invent checks.
Test protocol
content/fixtures/voice-notes-work-capture/voice-notes-test-protocol.md
The no-connector protocol: sanity prompt, prompt/output separation, screenshots, raw text, cleaned assistant output, metadata, and no-score blockers.
Before using real recordings
Voice-note workflows need privacy and human review.
The current test starts from synthetic pasted transcript text. It is not a shortcut around consent, recording laws, workplace data rules, or customer privacy.
Check every owner/deadline before copying tasks into a real system; messy transcripts often leave authority ambiguous.
Separate “likely reserve” language from actual shipment, booking, address confirmation, or customer commitment.
Look for missing low-status tasks such as checking old label photos in a retailer folder; polished summaries can drop small but important follow-ups.
Do not paste real recordings, customer names, phone contacts, family/personal conversations, or workplace audio into tools until consent, privacy, retention, and export rules are known.
Use a human reviewer for spend approvals, customer promises, freight decisions, label accuracy, and inventory numbers.
Limitations
This is not a real-audio, phone-app, or meeting-bot endorsement.
The current scores cover two public web prompt-only runs using a synthetic transcript packet. They do not test transcription accuracy from audio, mobile voice-note capture, paid plans, account memory, exports, team sharing, real customer or employee data, meeting-bot consent, phone contacts, calendars, reminders, CRM/email/note-app/task connectors, browser extensions, automations, shipment booking, file sends, spend approvals, inventory edits, or customer contact. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.