THE WORKDAY TOOLKIT · October 25, 2026
The percentages in the table were correct. The sentence underneath was not. In our AI survey-summary run, the share choosing report drafting moved from 50% to 75%. The assistant calculated a 50% relative increase, then described that share as having “roughly doubled”.
It had increased to 1.5 times its previous level. This walkthrough shows the original response, the arithmetic behind the correction, and a short way to check a survey summary before it enters a management report.
Take the useful part first: copy the corrected summary · check the calculations · try the exact prompt
A corrected summary you can adapt
In the current fictional survey, 18 of 24 respondents (75%) selected report drafting, 12 (50%) selected meeting notes, and 6 (25%) selected email replies. Multiple selections were allowed. Report drafting was selected by 10 of 20 respondents (50%) in the previous survey. The observed share rose by 25 percentage points, equivalent to a 50% relative increase, or 1.5 times the previous share. The respondent groups may differ. These surveys measure stated preferences, not productivity gains. Suggestion: Ask which report types respondents want help with before choosing a workflow to test.
This is our corrected editorial version, not a second model response. For your own report, replace the fictional figures and retain the denominators and survey design that apply to your data.
What we actually ran
On October 25, 2026, we made one separate text-generation request through Picsart using its Claude Sonnet 4.6 model label (claude-sonnet-4-6). We supplied only the prompt and fictional survey facts below, without an answer key or this article. We retained the first response and made no retries. No model settings were exposed in the selected form; the backend model configuration was not independently verified.
The response met the requested under-220-word length and included a table, comparison and suggestion. The issue was inside a fluent comparison sentence.
Read the original AI response, unchanged
## AI Task Preferences – Survey Summary ### 1. Current Survey Results (n = 24) | Task | Selections | Share of Respondents | |---|---|---| | Report drafting | 18 | 75% | | Meeting notes | 12 | 50% | | Email replies | 6 | 25% | *Respondents could choose more than one option.* --- ### 2. Report Drafting: Cross-Survey Comparison In the previous survey (n = 20), report drafting received 10 selections (50% of respondents). In the current survey it rose to 75% — a **+25 percentage-point increase**. In relative terms, that share roughly doubled (+50% relative change). These figures should be interpreted cautiously: respondent counts differ, we do not know whether the same individuals participated in both surveys, and no controlled measurement was conducted. --- ### 3. Suggested Next Step **Suggestion:** Before investing in AI tooling, consider a brief follow-up to understand *why* report drafting is the top priority. Useful questions might include which report types are most burdensome, how much time staff estimate they spend on drafts, and what quality or compliance requirements apply. This would give the team more concrete data to evaluate potential solutions without overstating what the current survey alone can tell us.
Check the denominator before adding percentages
Twenty-four people answered the current survey, and each could select several tasks. The three options received 36 selections in total. Those totals answer different questions: people are the denominator for the share of respondents; selections are the denominator for the share of all selections.
| Task | Selections / respondents | Share of respondents |
|---|---|---|
| Report drafting | 18 / 24 | 75% |
| Meeting notes | 12 / 24 | 50% |
| Email replies | 6 / 24 | 25% |
These percentages add to 150%. That is consistent with a multiple-selection question; it does not mean that 150% of people responded. Dividing 18 by 36 instead would produce 50% of selections. That calculation is valid for a different measure, but it would not answer how many respondents selected report drafting.
Separate points, relative change and multiples
| Measure | Calculation | Meaning |
|---|---|---|
| Current share | 18 ÷ 24 × 100 | 75% of current respondents |
| Previous share | 10 ÷ 20 × 100 | 50% of previous respondents |
| Percentage-point change | 75 − 50 | +25 percentage points |
| Relative percentage change | (75 − 50) ÷ 50 × 100 | +50% relative to the earlier share |
| Multiple of earlier share | 75 ÷ 50 | 1.5 times, not twice |
The original response said: “that share roughly doubled (+50% relative change)”. The parenthetical calculation was right; the description contradicted it. Doubling the earlier 50% share would mean reaching 100%, a relative increase of 100%. We removed “roughly doubled” and stated the actual change.
Compare the shares rather than treating the increase from 10 to 18 selections as the change in preference. The respondent counts also changed from 20 to 24. Since we do not know who answered each survey, the numbers do not establish that particular people changed their minds.
Copy the exact prompt used in the run
Summarize this fictional internal survey for a team lead. SOURCE Current survey: 24 people responded. Question: Which work tasks would you like AI help with? Respondents could choose more than one option. Report drafting: 18 selections. Meeting notes: 12 selections. Email replies: 6 selections. Previous survey: 20 people responded. Report drafting: 10 selections. The previous survey also allowed multiple selections. We do not know whether the same people answered both surveys. No productivity or time-saving measurements were collected. OUTPUT 1. A table with Task, Selections, Share of current respondents. 2. A short comparison of report drafting across the two surveys. 3. A recommendation for what to explore next, explicitly labeled as a suggestion. Keep the whole response under 220 words. Use the respondent count as the percentage denominator. Distinguish percentage-point change from relative percentage change. Do not claim statistical significance or measured productivity improvement. Use only the source facts; keep uncertainty explicit.
The instruction already asked the assistant to distinguish percentage points from relative change. That instruction did not prevent the contradictory phrase in this response. A prompt can make your requirements clear; it does not replace checking the output.
Review the takeaway as carefully as the table
The source asks which tasks people want AI help with. It does not measure minutes saved, better report quality or a return on investment. The AI response kept its recommendation labeled as a suggestion and acknowledged uncertainty about the respondent groups. Those useful distinctions can stay.
For a practical review, read every phrase containing “up”, “down”, “doubled”, “half” or “improved” alongside its calculation. Confirm what was measured, who was counted and which denominator belongs in each percentage. Then check whether the conclusion goes beyond the question people actually answered.
Desk check: is “75% want AI, so productivity should rise 75%” supported?
No. The 75% refers to respondents selecting report drafting as a task they want help with. No productivity measurements were collected. A suitable next step is to explore the workflow and measure a pilot, rather than treating preference as an observed performance gain.
Use the checked result: Put the corrected findings into the weekly status report format. If you are choosing a tool for the pilot, use the seven checks before paying to identify the task and evidence you need.
Scope: This is one connector response on a small fictional dataset. We checked its arithmetic with direct calculations and reviewed the wording against the supplied facts. We did not measure statistical significance, task time or general model accuracy.