Gemini
Pending safe baseline
Run only after a short no-login sanity prompt returns a distinct assistant answer. Keep no-score if the page echoes the prompt or errors.
Evidence-backed timesheet reconstruction test
AI can draft a forgotten-timesheet estimate — but it is not proof of work.
This evidence log tests a safe first workflow: paste synthetic calendar, email, task, note, and activity snippets into a general AI assistant and ask for a reviewable timesheet draft. No real workplace data, calendar, mailbox, task system, time tracker, payroll, legal-billing, invoicing, HR, or monitoring connector was used.
Quick answer
ChatGPT is the safer first draft in this fixture; both tools still need manual correction.
In the synthetic Northstar Candle Co. Friday fixture, ChatGPT scored 4.22/5 and Duck.ai scored 3.94/5. Both produced separable no-login answers, but neither should be copied directly into payroll, billing, invoicing, HR, legal, compliance, or monitoring systems.
The practical workflow is: sanitize the source traces, paste only approved excerpts, ask for a confidence-labeled draft, then have the worker or manager verify every minute, category, and exclusion before any downstream use.
Current no-login results
Two timesheet-reconstruction baselines scored.
Scores cover the separable assistant answer only. Setup friction, prompt echo, UI cleanup, and pending routes are tracked separately in the evidence CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| ChatGPT | 4.22/5 | 2026-06-18 | Best current timesheet draft: clean table, category totals, excluded-time list, uncertainty questions, and a strong human-review-only warning. It excluded break/lunch, counted only 32 minutes of the weak packing-queue block, and did not count the full 16:00-16:30 buffer. | Still needs correction before use. It counted the Birch & Cedar block as 45 minutes instead of the stronger 40-minute evidence, undercounted the vendor-call/admin segment, assigned the HP-721 check to fulfillment, and missed several specific follow-up questions. |
| Duck.ai | 3.94/5 | 2026-06-18 | No-login route worked with a visible GPT-5 mini label. The answer covered all requested sections, excluded explicit break/lunch, flagged packing and buffer uncertainty, and kept payroll/billing/legal use under human review. | It counted an unsupported 11:45-12:00 admin segment, included the low-confidence 16:05-16:12 buffer check in the headline total, and drifted on category totals and vendor-call categorization. |
Pending rows
Uncaptured tools stay unscored.
A pending row is not a quality judgment. It means we have not yet saved a separable full answer through a safe no-login or project-safe route.
Perplexity
Pending safe baseline
Keep as no-score until a public or project-safe route produces a separable answer and not just search UI or setup friction.
Microsoft Copilot
Pending safe baseline
Prior no-login Copilot checks often accepted prompts but showed no visible answer; treat as setup-friction recon unless a distinct answer appears.
The test scenario
Fictional Friday work traces from a small ecommerce shop.
The fixture uses Northstar Candle Co., a fictional ecommerce and workshop business. The packet includes calendar blocks, sent-email snippets, task comments, vendor-call notes, owner check-in notes, browser/activity ranges, break/lunch exclusions, and uncertainty around a packing queue and an end-of-day buffer.
A strong answer should reconstruct supported segments, avoid double-counting calendar/email/task evidence, calculate category totals, label confidence, exclude break/lunch/unconfirmed buffer time, and ask the right questions before anything reaches a real system.
Reusable paste-only prompt shape
What the benchmark asks the assistant to do.
The prompt is intentionally connector-safe: it tells the assistant to use only pasted synthetic traces and never claim access to calendar, email, task, payroll, billing, HR, legal, or monitoring systems.
Use only pasted source records; do not invent exact work, clients, payroll status, billable codes, project codes, or legal/compliance conclusions.
Do not connect calendars, mailboxes, chat, task systems, time trackers, payroll, invoicing, legal-billing, HRIS, or monitoring systems.
Exclude breaks, lunch, commuting, and unconfirmed buffer time from work totals, and do not double-count overlapping records.
Use confidence labels and call out weak evidence such as the packing queue and 16:00-16:30 buffer block.
Return a table, category totals, excluded/not-counted time, follow-up questions, and a checklist before any downstream use.
Evidence files
The fixture is reusable and connector-safe.
Synthetic Friday work traces
content/fixtures/timesheet-reconstruction/synthetic-friday-work-traces.md
Fictional Northstar Candle Co. calendar titles, sent-email snippets, task activity, call notes, owner check-in notes, browser/activity traces, and caveats.
Paste-only benchmark prompt
content/fixtures/timesheet-reconstruction/timesheet-reconstruction-paste-prompt.md
The exact connector-safe prompt asking for a draft timesheet table, category totals, excluded time, uncertainty questions, and human-review checklist.
Expected answer key
content/fixtures/timesheet-reconstruction/expected-timesheet-answer-key.md
Manual scoring key with supported segments, reasonable total range of about 325-333 minutes, category totals, and must-catch caveats.
Test protocol
content/fixtures/timesheet-reconstruction/timesheet-reconstruction-test-protocol.md
The no-connector protocol: sanity prompt first, screenshots, raw page text, submitted prompt, cleaned assistant output, metadata, and no-score blockers.
Before using real work traces
Keep the AI in draft mode.
Timesheet reconstruction can cross privacy, payroll, billing, HR, legal, and monitoring boundaries quickly. Start with sanitized excerpts and keep human review between AI output and every real action.
Use this workflow only on synthetic data or sanitized excerpts approved for AI use; real work traces can expose employees, customers, vendors, private communications, and business-confidential activity.
Treat the result as a memory aid and review draft, not evidence for payroll, legal billing, invoices, compliance, performance management, discipline, or employee monitoring.
Have the worker or manager verify every time range, category, confidence label, excluded block, and open question before copying anything into a timesheet or billing system.
Keep calendar/mailbox/task/time-tracker/payroll/legal-billing/HR connectors off until there is a clear privacy, retention, approval, and action-taking policy.
Manually check the arithmetic. A polished answer can still count unsupported calendar time, misclassify a category, or turn weak activity traces into false precision.
Limitations
This is a paste-only benchmark, not a time-tracking or payroll automation endorsement.
The current scores cover two prompt-only public web runs using a synthetic fixture. They do not test paid plans, native time-tracking tools, Google Calendar, Outlook, Gmail, Slack, Teams, Asana, Jira, ClickUp, Harvest, Toggl, QuickBooks Time, payroll, legal-billing, invoicing, HRIS, browser-history, device-monitoring, memory, or real action-taking. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.