This evidence log tests a safe first workflow: paste a fictional spreadsheet export plus source notes into a general AI assistant and ask for a manager-ready operations report. No file upload, Docs/Word/Excel/Sheets/Drive/OneDrive connector, BI system, email, CRM, accounting, customer, supplier, legal, tax, HR, medical, workspace connector, browser extension, or live-system action was used.
In the synthetic Northstar Candle Co. report-generator fixture, ChatGPT scored 4.83/5 and Duck.ai scored 3.00/5. ChatGPT calculated the core spreadsheet totals correctly. Duck.ai produced a plausible-looking report but materially mis-summed orders, revenue, support tickets, shipping issues, and late orders.
The practical workflow is: paste only approved low-risk source material, require formulas and caveats, then verify the numbers yourself before sharing the report.
Early search hints around report-from-Excel, spreadsheet assistants, automated weekly status reports, proposal drafts, CSV reports, and agent workflows are still sparse. The 2026-07-15 monitor shows weekly-status/report-from-Excel as the only click-bearing cluster, so this route map points readers to the nearest evidence-backed workflow instead of adding another thin page.
Use the proposal benchmark when report-from-Excel evidence becomes client-facing scope, price, proof, or follow-up copy. Keep generated reports separate from contracts, signatures, payment links, marketplace submissions, and legal/tax advice.
Use the support bridge when the real job is a recurring weekly/monthly update assembled from Excel rows, project notes, and source labels — especially “ai report generator from Excel”, “CSV report generator”, “automated project status report”, and “how to automate weekly status report in Excel” searches. The 2026-07-15 GSC monitor still makes this the only click-bearing cluster, so route rather than duplicate it.
Use the spreadsheet assistant benchmark before report drafting when the risky part is missing values, formulas, totals, or sheet cleanup.
Use the weekly-status evidence when the source is scattered work-log notes rather than a spreadsheet-plus-source-notes report packet.
Read the agent-mode boundary guide before letting any AI browse, connect files, edit spreadsheets, send messages, or take report-related actions.
Scores cover the separable assistant answer only. Prompt echo, setup friction, UI cleanup, and pending routes are tracked separately in the evidence CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| ChatGPT | 4.83/5 | 2026-06-19 | Best current result. It calculated the required totals correctly, including 705 orders, $54,380 gross revenue, $1,505 refunds, $52,875 net revenue, 58 shipping-issue orders, 44 late orders, 119 support tickets, week-level net revenue, and approximate CSAT of 4.38/5. | The appendix had a small CSAT numerator typo, and the source labels were broad rather than row-by-row. It is still a draft that needs human review before any management, finance, accounting, customer, supplier, legal, tax, HR, or live-workflow use. |
| Duck.ai | 3.00/5 | 2026-06-19 | The answer had a complete manager-report structure, useful caveat handling, visible source-style labels, and no claims of emailing, filing, approval, reconciliation, or live-system action. | It materially miscalculated the spreadsheet-style numbers: 574 orders, $35,955 gross revenue, $1,405 refunds, $34,550 net revenue, 49 shipping-issue orders, 33 late orders, and 178 support tickets instead of the answer-key totals. |
A pending row is not a quality judgment. It means we have not yet saved a full, distinct assistant answer through a safe no-login or project-safe route.
Pending safe baseline
Run only if a no-login or project-safe route produces a distinct full answer. Keep scores blank for prompt echo, login gates, or incomplete output.
Pending safe baseline
Prior Steel navigation to Perplexity has been unreliable; score only if a full separable answer is captured and raw page evidence is saved.
Pending safe baseline
Prior no-login Copilot checks often showed prompt echo instead of a usable answer. Treat as setup friction unless a distinct assistant response appears.
The fixture uses Northstar Candle Co., a fictional small business. The task asks the assistant to turn May source notes and an eight-row CSV export into an executive summary, key numbers table, week-by-week readout, risks and caveats, decisions needed, draft next actions, and an appendix.
A strong answer must calculate the totals exactly, preserve caveats, show source labels, avoid unsupported record/high/margin/forecast claims, and stay in draft-only mode.
The prompt deliberately separates report writing from document editing, finance action, customer/supplier communication, and live workflow updates.
content/fixtures/report-generator-docs-spreadsheets/monthly-ops-source-notes.md
Fictional Northstar Candle Co. manager request, business context, caveats, decisions needed, and preferred source-label style.
content/fixtures/report-generator-docs-spreadsheets/ops-metrics-export.csv
Eight fictional May rows split by week and channel, with orders, revenue, refunds, shipping issues, late orders, support tickets, CSAT averages, and source notes.
content/fixtures/report-generator-docs-spreadsheets/report-generator-paste-prompt.md
The exact no-upload prompt asking for a manager-ready report, calculations, caveats, decisions, draft-only next actions, and assumptions.
content/fixtures/report-generator-docs-spreadsheets/expected-report-output.md
Manual scoring key with answer-key totals, caveats, must-not-invent checks, and weighted rubric.
content/fixtures/report-generator-docs-spreadsheets/report-generator-test-protocol.md
The no-connector protocol: sanity prompt, screenshots, raw page text, submitted prompt, cleaned output, metadata, and no-score blockers.
Report generation touches numbers, business decisions, customers, suppliers, finance, and sometimes legal or HR contexts. A fluent paragraph is not evidence that the underlying data is correct.
The current scores cover two public web prompt-only runs using a synthetic fixture. They do not test paid plans, document uploads, PDF/OCR extraction, spreadsheet formulas inside Excel or Google Sheets, Docs/Word editing, Drive/OneDrive permissions, BI connectors, email/calendar/workspace connectors, CRM/accounting/customer/supplier systems, stored memory, real files, real customer data, legal/tax/HR/medical material, export formatting, or live action-taking. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.