Public run records
129
Reader-safe run records across meetings, admin, email, slides, support chatbots, and small-business website builders.
Original evidence note
The AI productivity paradox starts before the prompt.
People do not only lose time because an AI answer is wrong. In our current public testing database, the bigger early pattern is setup friction: login gates, OAuth-only paths, security checks, unclear upload routes, prompt echoes, export limits, or tools that cannot yet produce auditable output in a safe test.
Setup-friction records
42 (33%)
Runs where the useful finding was friction before usable generated output: sign-in, signup, upload, prompt echo, security, SSO, file-picker, or import issues.
Scored quality runs
39
Only runs with saved output artifacts and rubric scoring. Blank scores are deliberate; they are not hidden ratings.
Short answer: the productivity tax is real, but our data is narrow.
As of 12 June 2026, 90 of 129 public run records (70%) are not output-quality scores. That does not mean those tools are bad. It means the current safe, repeatable test could not yet reach a clean output artifact, or the work is still at the public-doc/availability stage.
The practical lesson for advanced beginners: before trusting a tool to save time, test whether you can safely get your work into the tool, get the result out, and verify the answer without connecting sensitive accounts too early.
Where time disappears before AI saves any
1. Sign-in before a usable workspace
Some tools show a promising public page or composer, then require login, SSO, phone, CAPTCHA, or human verification before a fixture can be tested. We record this as setup friction, not as output quality.
2. Prompt accepted, but no answer appears
A visible prompt box is not enough. Several no-login tests required a sanity prompt first because prompt echo or search-style UI can look like an answer unless the generated output is separated from the submitted text.
3. Import and export are the hidden workflow
Meeting, presentation, chatbot, and website-builder tools can fail the practical test before the model does: unclear upload paths, public-link defaults, weak export options, file-picker blockers, or live-meeting-only assumptions can erase the promised time savings.
Current scenario mix
| Scenario | Scored outputs | Setup-friction records | What it means for a worker |
|---|---|---|---|
| Weekly admin routine | 5 | 20 | Prompt-only outputs exist for five tools, but many public chat surfaces still hit prompt echo, signup, search-result, or sign-in friction before a clean answer. |
| Paste-only email assistant | 4 | 1 | ChatGPT, Gemini, Duck.ai, and Perplexity produced separable synthetic-inbox answers; Copilot no-login returned prompt echo only in the sanity check. |
| Client-ready slides | 0 | 5 | The strongest no-login path reached a composer but produced no auditable deck; other tools hit signup, security, credit-card, or export friction first. |
| Meeting follow-up notes | 0 | 3 | Current evidence is still about account/import/live-meeting boundaries, not summary quality. |
| Customer-support chatbot | 0 | 5 | Free/private testing is often blocked by signup, verification, disabled-submit, or source-ingestion friction before the FAQ answer key can be scored. |
| Small-business website builder | 1 | 3 | One private draft has a score; several builders still need safer signup, draft, export, or publication-boundary follow-up. |
A five-minute preflight before adopting any AI productivity tool
Use a harmless sample first.
Paste a synthetic or low-sensitivity sample task before connecting Gmail, Outlook, Slack, calendar, CRM, Notion, or a customer database.
Demand an export path.
Before scoring the answer, check whether you can copy, download, export, or privately save the output in the format your real workflow needs.
Separate access friction from answer quality.
A blocked signup does not prove the model is weak. A polished answer does not prove the tool is safe to connect to real accounts. Track both axes.
Methodology and limits
This note uses the public evidence-data summary, not a broad market survey. It currently covers 59 tracked tools, 20 test scenarios, 129 public run records, 42 setup-friction records, 10 availability-research records, 5 public-doc blockers, and 39 scored quality runs.
We avoid exact search-volume claims and do not rank setup-friction-only tools. For the raw reader-safe data, download the run index and score summary.