Synthetic messy order export
A fictional Northstar Candle Co. order export with mixed date, currency, percent, refund, shipping, channel, duplicate, and note fields.
content/fixtures/spreadsheet-assistant/spreadsheet-task-brief.md
Evidence-backed spreadsheet assistant test
AI spreadsheet assistants can help with messy Excel and Google Sheets exports — but only after you check the math.
This evidence log tests a safe first workflow: paste a synthetic messy order export into a general AI assistant and ask for formulas, cleanup checks, manual totals, and a business readout. No Google Sheets, Excel, Drive, OneDrive, add-on, ecommerce, accounting, payment, or live customer-data connector was used.
Quick answer
ChatGPT is the strongest paste-only spreadsheet baseline so far, but every answer needs manual verification.
In the current Northstar Candle Co. spreadsheet fixture, ChatGPT leads at 4.03/5, followed by Perplexity at 3.09/5, Duck.ai at 2.75/5, and Gemini at 2.58/5. ChatGPT was the only run that combined mostly correct headline totals with a complete cleanup/readout/privacy structure, yet even it had a profit mismatch and inconsistent channel rollups.
Treat these as paste-only prompt baselines, not final Excel/Google Sheets add-on recommendations. A useful reader workflow is: anonymize the export, paste only the needed rows/columns, ask for formulas and calculations, then manually check every formula, total, and source-column reference before changing a real workbook.
Current no-login results
Four spreadsheet assistant baselines scored.
All rows used the same fictional order export. Scores are quality scores for the separable assistant answer only. Setup friction, UI cleanup, prompt echo, and connector boundaries are recorded separately in the evidence notes and CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| ChatGPT | 4.03/5 | 2026-06-13 | Best current paste-only spreadsheet answer: correct gross sales, discounts, refunds, and net revenue; useful cleanup checklist, pivot-table plan, and owner-review bullets. | Estimated gross profit was $2,583.30 vs expected $2,584.30, channel rollups were inconsistent, and formulas still need manual review before copying into a real workbook. |
| Perplexity | 3.09/5 | 2026-06-12 | Correctly calculated gross sales, discounts, refunds, net revenue, and the top Wholesale channel; formula explanations were fairly reviewable. | The visible answer stopped before the requested business-readout and privacy/safety sections, and gross profit omitted shipping-paid revenue. |
| Duck.ai | 2.75/5 | 2026-06-12 | No-login route worked and it caught several cleanup issues, including mixed formats, returns/refunds, duplicate notes, and order/privacy cautions. | Several formulas referenced the wrong source columns and core totals were materially wrong: gross sales $5,807.00 vs expected $4,837.00 and net revenue $5,421.10 vs $4,450.10. |
| Gemini | 2.58/5 | 2026-06-13 | No-login Gemini accepted the synthetic spreadsheet fixture and returned a separable six-section answer with useful cleanup categories and privacy reminders. | Most manual calculations were materially wrong: gross sales $3,232.50 vs expected $4,837.00, net revenue $2,912.25 vs $4,450.10, and Wholesale net revenue $1,831.00 vs $3,351.00. |
Unscored candidate
Microsoft Copilot remains pending.
Microsoft Copilot is still in the spreadsheet fixture as a pending row because we have not yet captured a safe no-login or project-safe paste-only spreadsheet answer with a separable output. It should not appear in score summaries until raw output exists and the same seven rubric criteria are completed.
The test scenario
A fictional ecommerce order export with enough mess to catch false confidence.
The fixture uses Northstar Candle Co., a fictional candle business. The pasted export includes mixed date formats, currency symbols, percent strings, refunds, shipping-paid values, channels, discount codes, product categories, possible duplicate notes, and zero-shipping contexts that should not automatically be treated as errors.
A good answer should produce usable formulas and manual totals while warning that the reader must verify source columns, formulas, and totals. A weak answer often sounds confident while referencing the wrong columns or inventing totals that do not match the pasted data.
Reusable paste-only prompt shape
What the benchmark asks the assistant to produce.
The public prompt is intentionally connector-safe: it tells the assistant to use only the pasted synthetic table and never claim access to Google Sheets, Excel, Drive, OneDrive, ecommerce systems, accounting systems, refunds, invoices, inventory, or live files.
Create helper-column formulas for normalized unit price, discount rate, gross sales, discount amount, net revenue, estimated profit, return flags, and possible duplicates.
Manually calculate key totals from the pasted data and show the arithmetic rather than only suggesting formulas.
Flag cleanup issues such as mixed date formats, currency symbols, percent text, refund rows, returns, possible duplicate orders, and notes that should not be blindly deleted.
Suggest practical pivot-table views for channel, category, product, returns/refunds, and discount-code review.
Write a short owner-review business readout that preserves the fixture facts and does not invent tax, payment, demand, or inventory details.
List privacy and connector cautions before using real customer spreadsheets, Google Sheets, Excel, Drive, OneDrive, add-ons, or OAuth connectors.
Evidence files
The fixture is reusable and connector-safe.
Paste-only benchmark prompt
The exact connector-safe prompt used in no-login assistant tests. It asks for helper-column formulas, manual totals, cleanup flags, pivot suggestions, business readout, and privacy cautions.
content/fixtures/spreadsheet-assistant/spreadsheet-assistant-paste-prompt.md
Expected output / answer key
Ground-truth totals and checks: gross sales $4,837.00, discounts $336.80, refunds $101.00, net revenue $4,450.10, estimated gross profit $2,584.30, and Wholesale as the top channel at $3,351.00.
content/fixtures/spreadsheet-assistant/expected-spreadsheet-output.md
Test protocol
The scoring and safety protocol for paste-only spreadsheet assistant runs, including connector boundaries, privacy rules, scoring criteria, screenshots, raw outputs, and no-score blockers.
content/fixtures/spreadsheet-assistant/spreadsheet-assistant-test-protocol.md
Before using real spreadsheet data
Use AI as a formula reviewer, not a live finance operator.
This test is deliberately paste-only because real spreadsheets can contain customer, payment, margin, payroll, tax, or confidential business data. Start with anonymized excerpts and keep a human review step between AI output and any real workbook change.
Never trust an AI spreadsheet answer until you manually verify formulas, totals, and source-column references.
Check headline totals and secondary rollups separately; a tool can get gross sales right and still break profit or channel subtotals.
Remove or mask customer names, emails, addresses, payment data, order IDs, private notes, and sensitive revenue/margin details before using real data.
Avoid Google Sheets, Excel, Drive, OneDrive, accounting, ecommerce, or add-on connectors until privacy controls and rollback/export behavior are understood.
Use AI output as a review draft, not as an automatic live spreadsheet, accounting, refund, inventory, invoice, or tax update.
Search route map · updated 2026-07-16
If you searched for spreadsheet assistants or Excel report automation, continue here.
Early Search Console hints include spreadsheet assistant, AI data-entry automation, form-filler, and weekly-status/report-from-Excel terms. The safest route is to verify formulas here, then send readers to the nearest evidence-backed boundary guide instead of pretending spreadsheet cleanup is ready for browser agents, form extensions, accounting tools, or live connectors.
AI data-entry and browser automation boundaries
Use this when spreadsheet-cleanup searches drift into “AI data entry automation,” browser agents, extensions, live forms, customer records, or production updates. It keeps the workflow in staging-table and human-stop mode.
AI form filler, PDF, and extension safety guide
Use this when spreadsheet rows become form fields, PDFs, or browser autofill tasks. It routes free/online form-filler intent back to paste-only field drafts before uploads or extensions.
Automate a weekly status report from Excel and notes
A support guide for turning checked spreadsheet rows and project notes into a manager-ready update without Slack, Jira, CRM, finance, email, or live spreadsheet connectors.
Report generator from docs and spreadsheets
Shows why polished narratives still need arithmetic review: one report baseline was excellent, while another looked clean but had materially wrong totals.
Admin-professional starter stack
Use this route map when spreadsheet cleanup is one part of an admin workflow with reports, PDFs, forms, calendars, inboxes, tasks, and workflow automation.
Invoice and receipt register drafting
Use this when spreadsheet rows are really receipt, invoice, reimbursement, credit, or missing-document records that need bookkeeping-style status separation before any accounting connector is trusted.
Disclosure and limits
This is an in-progress evidence log, not a final spreadsheet-tool ranking.
The current scores cover four prompt-only web runs using a synthetic fixture. They do not test paid plans, native spreadsheet add-ons, Google Workspace or Microsoft 365 connectors, file uploads, macros, Apps Script, live accounting, ecommerce integrations, memory, automations, or enterprise data controls. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.
Last updated: 2026-07-16 by AIProductivity.guru Editorial Team.