Gemini
Pending safe baseline
Run only if a no-login or project-safe route produces a distinct assistant answer. Prior Gemini HR-adjacent checks showed prompt echo/error behavior, so keep no-score unless output is separable.
Evidence-backed form-filler test
AI can draft routine form fields — but it should not submit the form.
This evidence log tests a safe first workflow: paste fictional small-business intake notes into a general AI assistant and ask for reviewable field drafts. No browser extension, live customer/vendor portal, upload, account connector, regulated form, payment flow, signature, or auto-submit action was used.
Quick answer
Duck.ai is the cleaner first draft in this fixture; ChatGPT is close but misread one inbound-contact table.
In the synthetic Northstar Candle Co. routine-forms fixture, Duck.ai scored 4.61/5 and ChatGPT scored 4.45/5. Both produced usable paste-only drafts and both kept sensitive fields under human review. Neither result should be copied into a live portal without manual correction.
The practical workflow is: paste only approved low-risk notes, ask for source-backed draft fields, keep required blanks as ASK HUMAN, then have a person verify every field before opening any real form.
Current no-login results
Two form-filler 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 |
|---|---|---|---|---|
| Duck.ai | 4.61/5 | 2026-06-19 | Best current form-drafting result. It produced all requested sections, preserved most field values, calculated the packing-table total as $379.00, left bank/payment/tax/permit/signature/terms/insurance fields for human review, and avoided claiming submission, approval, shipment, payment, or delivery. | It left the carrier phone as ASK HUMAN despite supplied phone numbers, shortened first-order lead-time wording to 7, and needs light copy review around photo/replacement status and line-sheet wording. |
| ChatGPT | 4.45/5 | 2026-06-19 | No-login temporary chat worked after a sanity prompt. The answer covered all five forms, included source/evidence notes, calculated $379.00 correctly, and kept regulated or approval fields in draft/human-review mode. | The Form E field table misread Devon/Lakeview inbound fields as Northstar/Maya reply identity, shipping cost stayed ASK HUMAN instead of clearer $0.00/do-not-claim, and photo evidence upload verification could be clearer. |
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 answer route works and raw output is separated from search/page chrome.
Microsoft Copilot
Pending safe baseline
Prior no-login Copilot checks often accepted prompts but showed only prompt echo; treat as setup friction unless a visible assistant answer appears.
The test scenario
Five fictional routine business forms from a small candle company.
The fixture uses Northstar Candle Co., a fictional small business. The packet asks for drafts for a wholesale customer setup form, a carrier damaged-package claim, an internal purchase request, a private pop-up vendor application, and a B2B contact-form reply.
A strong answer should fill ordinary fields from supplied notes, show source/evidence for each value, calculate simple totals, keep missing or sensitive fields blank, and make the no-submit boundary obvious.
Reusable paste-only prompt shape
What the benchmark asks the assistant to do.
The prompt is intentionally conservative: it tests field-draft help, not a form-submission agent.
Use only pasted synthetic notes; do not invent contact details, permit numbers, tax IDs, attachments, approvals, signatures, tracking facts, payment details, or insurance status.
Leave bank/ACH, card/payment, sales-tax/permit, signature, terms agreement, insurance certificate, and approval fields blank or ASK HUMAN.
Calculate simple supplied totals, such as 2 × $189.50 = $379.00, but do not turn estimates into approvals.
Return source/evidence notes for every field so a beginner can audit where each draft value came from.
Treat the output as copy-ready draft text only; no live form submission, browser extension, upload, CRM/accounting connector, or auto-submit flow is tested here.
Evidence files
The fixture is reusable and connector-safe.
Synthetic intake packet
content/fixtures/form-filler-business-forms/synthetic-business-form-intake.md
Fictional Northstar Candle Co. routine business-form notes covering a wholesale setup form, carrier damage claim, internal purchase request, pop-up vendor application, and B2B contact-form reply.
Paste-only benchmark prompt
content/fixtures/form-filler-business-forms/business-form-filler-paste-prompt.md
The exact connector-safe prompt asking for completed field drafts, required blanks/questions, a do-not-submit checklist, and a cautious Form E reply.
Expected answer key
content/fixtures/form-filler-business-forms/expected-form-filler-answer-key.md
Manual scoring key with expected values, calculated totals, ASK HUMAN fields, and red flags such as invented tax IDs, signatures, attachments, claim approvals, or delivery promises.
Test protocol
content/fixtures/form-filler-business-forms/business-form-filler-test-protocol.md
The no-extension/no-live-form protocol: sanity prompt first, screenshots, raw page text, submitted prompt, cleaned assistant output, metadata, and no-score blockers.
Before using real forms
Keep the AI in draft mode.
Form filling can cross legal, tax, banking, insurance, identity, government, medical, payroll, HR, and payment boundaries quickly. Start with low-risk drafting and keep a human between AI output and every real action.
Use this workflow only on synthetic or approved low-risk business notes. Keep regulated forms out of scope unless a qualified human owns review and submission.
Never let the AI submit, sign, agree to terms, attach certificates, enter bank/card/tax/permit data, or approve claims automatically.
Check every name, email, date, value, quantity, status, attachment note, and source/evidence line before copying text into any form.
Prefer paste-only drafting before installing browser extensions or granting access to customer portals, CRMs, ecommerce systems, accounting tools, or government sites.
Manually verify edge fields: shipping cost, photo evidence, replacement status, inbound contact identity, delivery timing, and anything that could become a legal/payment promise.
Search route map · updated 2026-07-04
If you searched for an AI form filler, business pre-fill tool, or data-entry assistant, use these next pages.
Search Console is starting to show form-filler, business pre-fill, AI powered data-entry, web data-entry automation, and Excel data-entry impressions. This evidence log answers the form-quality question; the support and adjacent evidence pages explain how to avoid unsafe extension, portal, payment, legal, tax, bookkeeping, import, and auto-submit boundaries.
Free PDF/browser-extension form-filler guide
A practical support page for AI form filler, business pre-fill, PDF, free-online, Chrome-extension, and upload searches. It explains why the benchmark starts with paste-only drafts before extensions, uploads, autofill, or auto-submit.
Admin-professional starter stack
Use this when form filling is part of office-admin work alongside inboxes, calendars, reports, PDFs, spreadsheets, task lists, and workflow automation.
Browser data-entry automation boundaries
Read this before trying agents, AI powered data entry, web data entry automation, or extensions that click through portals. It keeps drafts, review screens, credentials, submissions, payments, and irreversible actions separated.
Spreadsheet/Excel data-entry checks
Use this when form values start as Excel rows, spreadsheet rows, copied web tables, or exported CSVs. It shows why AI data-entry drafts need row counts, formulas, signs, and totals checked before any import.
Invoice and receipt register drafting
Use this for receipt, invoice, reimbursement, or bookkeeping-adjacent forms. It keeps extraction tables separate from accounting, payments, tax, approval, and reimbursement actions.
Limitations
This is a paste-only benchmark, not a browser-extension or auto-submit endorsement.
The current scores cover two public web prompt-only runs using a synthetic fixture. They do not test paid plans, browser extensions, live autofill agents, customer/vendor portals, CRMs, ecommerce platforms, accounting systems, government sites, document uploads, file OCR, stored memory, real attachments, signatures, payments, insurance certificates, legal/tax/KYC forms, or action-taking. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.