Workers do not hand over what agents do best on benchmarks: data analysis (7.1%) and document reading (5.7%) lead full delegation, while only 25% of record-keepers use AI at all.
Back to the trend of the day: OpenAI's new AI does office work on your computer, at half price
Poly Rating · from the trend of the day
What Can AI Agents Do? 10 Office Tasks Ranked (2026)
One in five US workers say AI has fully taken over at least one of their tasks, yet only 25% of those who keep records use AI for it at all.
- No. 1: File sorting and renaming
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
File sorting and renaming
Every AI put it first: a wrong folder is visible at a glance and undone by dragging it back. No benchmark publishes a success rate for this category.
Included it: Claude, Meta, Grok, Gemini, Nova, Kimi, DeepSeek, Qwen.In their lists: Claude: 1 · Meta: 1 · Grok: 1 · Gemini: 1 · Nova · Kimi: 1 · DeepSeek: 1 · Qwen: 1
What each AI noted 8
- Claudeplace 1
A mistake here is usually reversible and visible at a glance. That makes it safe even at a modest success rate, because checking takes seconds.
Source ↗ - Metaplace 1
Zero cost of mistake — agent can be 60% accurate and still save time vs $41,340 data entry median
Source ↗ - Grokplace 1
File-management tasks are among the highest-success categories on OSWorld; low mistake cost and easy visual check make them safe to hand over even at modest rates.
Source ↗ - Geminiplace 1
A 92% benchmark success rate paired with a zero-cost undo command saves 85% of human clerical time.
Source ↗ - Nova
Named differently in its own list: open the full list.
- Kimiplace 1
The only task class where a mistake costs seconds to fix — misnamed files and wrong folders are visible, reversible, and agents already clear OSWorld's file tasks at among the benchmark's highest rates.
Source ↗ - DeepSeekplace 1
The one task where a 34% agent can still save time: a wrong folder is a drag-and-drop undo, not a restatement.
- Qwenplace 1
High success rate and zero financial risk; a mistake just means moving a file back to its original folder.
Source ↗
8 AIs put it here - 2
Spreadsheet cleanup and data analysis
Data analysis is the task most often fully handed over (7.1%). Benchmarks disagree: up to 97.87% on simple Calc tasks, 16% on full MBABench-2.0 builds.
Included it: Claude, Meta, Grok, Gemini, Nova, Kimi, DeepSeek, Qwen.In their lists: Claude: 4 · Meta: 4 · Grok: 5 · Gemini: 5 · Nova · Kimi: 4 · DeepSeek: 6 · Qwen: 2
What each AI noted 8
- Claudeplace 4
Data analysis is the single most delegated task at 7.1% of workers, yet an older OfficeWorld test put baseline models near 21-23% on Word, Excel and PowerPoint tasks.
Source ↗ - Metaplace 4
30 tool calls per task in OSWorld 1.0 vs 318 in 2.0 — short spreadsheet edits now near human 72.36% baseline
Source ↗ - Grokplace 5
LibreOffice Calc tasks score high on OSWorld-Verified for top agents; still needs spot-checks when formulas or multi-sheet logic are involved.
Source ↗ - Geminiplace 5
Automates string trimming and header matching across 81% of sheets, but syntax anomalies still need validation.
Source ↗ - Nova
Named differently in its own list: open the full list.
- Kimiplace 4
Single-app spreadsheet work is where benchmark numbers are strongest — WorkArena-L1's top agent clears 90.3% of atomic enterprise tasks.
Source ↗ - DeepSeekplace 6
The top spreadsheet agent passes 16% of MBABench-2.0 tasks; debugging accuracy is 12%, so every cell needs a second pair of eyes.
Source ↗ - Qwenplace 2
Agents handle structural cleanup well, and Excel's undo button makes human checking painless.
Source ↗
8 AIs put it here - 3
PDF and document data extraction
5.7% of workers already hand over reading work documents in full, and the output can be spot-checked against the source page instead of re-typed.
Included it: Claude, Meta, Grok, Gemini, Kimi, DeepSeek, Qwen. Did not include: Nova.In their lists: Claude: 2 · Meta: 3 · Grok: 6 · Gemini: 2 · Kimi: 3 · DeepSeek: 2 · Qwen: see list
What each AI noted 7
- Claudeplace 2
Reading work documents is already among the most-delegated tasks, at 5.7% of workers. It is output you can spot-check against the source.
Source ↗ - Metaplace 3
Structured extraction with code execution — low reversal cost, high time saving vs manual copy-paste
Source ↗ - Grokplace 6
Document tasks are common in OSWorld; extraction is high-success when the target schema is clear and the output is reviewed before downstream use.
- Geminiplace 2
Vision agents parse messy multi-page receipts with 89% accuracy, requiring only spot checks on totals.
Source ↗ - Kimiplace 3
The safest extraction job because every cell lands somewhere a human can spot-check — and GPT-6.1 Terra's launch pitch was built exactly on this invoice-and-form grind.
Source ↗ - DeepSeekplace 2
LlamaParse Agentic scores 84.88 on ParseBench; the human check is a two-minute spot-check, not a line-by-line audit.
Source ↗ - Qwen
Named differently in its own list: open the full list.
see list
7 AIs put it here - 4
Web research and browsing
The agent only reads and changes nothing. But scores swing from 78.0% on one browsing benchmark to 23.5% on everyday tasks, so sources still need opening.
Included it: Claude, Meta, Grok, Gemini, Kimi, DeepSeek, Qwen. Did not include: Nova.In their lists: Claude: 3 · Meta: 2 · Grok: 2 · Gemini: 3 · Kimi: 5 · DeepSeek: 3 · Qwen: 7
What each AI noted 7
- Claudeplace 3
The agent only gathers information and changes nothing. The OSWorld 2.0 write-up says Claude does better on dynamic environments and cross-source reasoning.
Source ↗ - Metaplace 2
Agents now read pricing pages and compare vendors unsupervised — 42.7% to 55% success on WorkArena web tasks
- Grokplace 2
Top workplace AI use in Microsoft data is decision support and information retrieval (27.5% and 13%); workers already hand these over freely.
Source ↗ - Geminiplace 3
Browsing agents compile multi-site market catalogs reliably with zero risk of altering underlying live data.
- Kimiplace 5
Agents now gather across tabs faster than people — but the best error rate on record is still 4.3%, which is a hallucination budget, not zero.
Source ↗ - DeepSeekplace 3
Browser Use Cloud hits 78.0% on browsing, but ClawBench's everyday tasks drop Claude Sonnet 4.6 to 23.5% — verification is the job.
Source ↗ - Qwenplace 7
Agents can gather data across tabs, but hallucinated sources require line-by-line fact-checking by a person.
Source ↗
7 AIs put it here - 5
CRM record updates
Only 25% of record-keepers use AI for it, the lowest of ten tasks surveyed. Tool-based agents reach 64–84%, while screen-driven open agents score under 5% on Salesforce.
Included it: Claude, Meta, Grok, Gemini, Kimi, DeepSeek, Qwen. Did not include: Nova.In their lists: Claude: 6 · Meta: 7 · Grok: 7 · Gemini: 7 · Kimi: 7 · DeepSeek: 7 · Qwen: 3
What each AI noted 7
- Claudeplace 6
Only 25% of workers who maintain records use AI for it, the lowest of all measured tasks. They are wary of silent errors in records.
Source ↗ - Metaplace 7
WorkArena CRM tasks only 42.7% success — needs human check because customer data hard to undo
- Grokplace 7
Multi-app and enterprise UI workflows remain harder (WorkArena list filters near 0%); safe only for low-stakes fields with quick human review.
Source ↗ - Geminiplace 7
74% benchmark accuracy leaves 1 in 4 contact fields corrupted if unreviewed, risking client outreach hygiene.
Source ↗ - Kimiplace 7
Atomic CRM edits pass at 90.3%, but the compositional versions of the same ServiceNow work collapse to 2.1% — that 88-point gap is exactly where a human check still lives.
Source ↗ - DeepSeekplace 7
CRM-Agent-Bench strict success spans 64.32–83.60%, but SCUBA's Salesforce screen test puts open-source computer-use agents under 5% — the API and the screen are different worlds.
Source ↗ - Qwenplace 3
Agents navigate multi-tab forms easily, though sales reps still verify key client updates before saving.
Source ↗
7 AIs put it here - 6
Expense entries and invoice posting
Fable 5.1 enters invoices correctly 72% of the time against 51% for the best human tester, but the 28% it gets wrong touches money and audit trails.
Included it: Claude, Meta, Grok, Gemini, Kimi, DeepSeek, Qwen. Did not include: Nova.In their lists: Claude: 9 · Meta: 9 · Grok: see list · Gemini: 9 · Kimi: 9 · DeepSeek: 8 · Qwen: 5
What each AI noted 7
- Claudeplace 9
It is a small data-entry task that touches reimbursements and audit trails. The data-entry job category is projected to shrink 25.9% from 2023 to 2033.
Source ↗ - Metaplace 9
Touches money — 90% accurate agent that needs line-by-line check saves no time at all
Source ↗ - Grok
Named differently in its own list: open the full list.
see list - Geminiplace 9
Mistaken expense entries trigger audit penalties; 68% of finance workers insist on full receipt verification.
Source ↗ - Kimiplace 9
At clerical wages of roughly $21-22 an hour, a line-by-line check of expense entries costs about as much as typing them — the savings evaporate exactly here.
Source ↗ - DeepSeekplace 8
Fable 5.1 passes 72% of invoice-entry tasks, beating the best human at 51%, but the 28% that fail can move money.
Source ↗ - Qwenplace 5
Agents easily pull dates and totals, but humans must approve payouts before accounting cuts the check.
Source ↗
7 AIs put it here - 7
Payments, wire transfers and payroll
At 98% accuracy per step, a 20-step task succeeds end to end only about 67% of the time. OpenAI, Anthropic and Google all advise human approval on money.
Included it: Claude, Meta, Grok, Gemini, Kimi, DeepSeek, Qwen. Did not include: Nova.In their lists: Claude: 10 · Meta: see list · Grok: see list · Gemini: 10 · Kimi: 10 · DeepSeek: 10 · Qwen: 10
What each AI noted 7
- Claudeplace 10
Even at 98% per step over 20 steps, end-to-end success is only about 67%. For irreversible payments that is a long way from safe.
- Meta
Named differently in its own list: open the full list.
see list - Grok
Named differently in its own list: open the full list.
see list - Geminiplace 10
Even at 88% precision, the irreversible cost of a mistaken bank routing entry forces line-by-line human sign-off.
Source ↗ - Kimiplace 10
The task where 90% accuracy is worthless — the 10% that lands in a bank transfer is the only part anyone remembers.
Source ↗ - DeepSeekplace 10
No benchmark needed: OpenAI, Anthropic and Google all tell you to keep a human approval on money, terms and credentials.
Source ↗ - Qwenplace 10
The ultimate human-in-the-loop task, where an agent's misclick could drain company funds instantly.
Source ↗
7 AIs put it here - 8
Form filling (non-payment forms)
It is the headline use of the new model, yet the end-to-end form benchmark scores 43.31 on multi-file forms and 15.5 on instruction-only forms. Every field still needs a look.
Included it: Claude, Meta, Gemini, Nova, Kimi, DeepSeek. Did not include: Grok, Qwen.In their lists: Claude: 8 · Meta: 8 · Gemini: 4 · Nova · Kimi: 8 · DeepSeek: 5
What each AI noted 6
- Claudeplace 8
A person with a 90%-accurate agent still has to check each field. At the data-entry median of about $19.88 an hour, that check can cancel out the saving.
Source ↗ - Metaplace 8
GPT-6.1 Terra built for filling forms — 15 points higher than predecessor but guessing still 38%
Source ↗ - Geminiplace 4
Achieves 84.8% completion on ServiceNow WorkArena, reducing repetitive manual typing by 70%.
Source ↗ - Nova
Named differently in its own list: open the full list.
- Kimiplace 8
Forms are Terra's whole sales pitch — 'filling out forms' is OpenAI's own example — yet end-to-end enterprise form workflows remain the single weakest benchmark category.
Source ↗ - DeepSeekplace 5
CRAFT reaches 43.31 E2E on multi-file forms but 15.5 on instruction-only forms — the form is easy, the instructions are not.
Source ↗
6 AIs put it here - 9
IT service portal tickets and requests
Single-step ServiceNow tasks reach 79.1–90.3% on aggregator boards (the snapshots differ), and service-catalog forms reached 77.8% even with GPT-4o.
Included it: Claude, Grok, Gemini, Qwen. Did not include: Meta, Nova, Kimi, DeepSeek.In their lists: Claude: 5 · Grok: 4 · Gemini: 6 · Qwen: 6
What each AI noted 4
- Claudeplace 5
On WorkArena-L1 an aggregator shows a top agent near 79.1%, and the same snapshot's best L2 is near 69.4%. All older models scored 0% on list filtering.
Source ↗ - Grokplace 4
WorkArena service-catalog tasks reached ~78% with GPT-4o-class models; UI is conventional and mistakes are easy to spot.
Source ↗ - Geminiplace 6
WorkArena scores show 79.5% accuracy in routing internal incidents without exposing client-facing communications.
Source ↗ - Qwenplace 6
End-to-end success on WorkArena is solid for password resets, but hardware requests still need human approval.
4 AIs put it here - 10
Multi-app workflows
The hardest case: the best strict full completion on OSWorld 2.0 is 48.7%, self-reported by Anthropic, while the paper's best official run finished 20.6%.
Included it: Claude, Meta, DeepSeek, Qwen. Did not include: Grok, Gemini, Nova, Kimi.In their lists: Claude: 7 · Meta: 10 · DeepSeek: 9 · Qwen: see list
What each AI noted 4
- Claudeplace 7
The OSWorld 2.0 write-up calls multi-app work the hardest case, with up to five apps in one task. Best full completion is 48.7% (Anthropic-reported).
Source ↗ - Metaplace 10
OSWorld 2.0 median 1.6 hours, 318 tool calls — 20-24% binary completion on short tasks, collapses on long
Source ↗ - DeepSeekplace 9
OSWorld 2.0's best agent fully completes 20.6% of 108 multi-app workflows that average 318 tool calls — four out of five still need a human.
Source ↗ - Qwen
Named differently in its own list: open the full list.
see list
4 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
Single-step and multi-step results come from different agents and years: about 79–90% on WorkArena-L1 boards against 2.1% for GPT-4o on the 2024 WorkArena++ tasks, which humans finish 93.9% of the time.
Trust lags use: only 27% of employees strongly agree they are comfortable delegating tasks to AI agents.
Surveys disagree on how much AI work needs fixing: Epoch/Ipsos finds 66% of output used unchanged or lightly edited, while Asana finds 54% of workers redo or correct it.
Waiting is part of the cost: in one study an agent took 12 minutes to change line spacing, a job a person does in a couple of minutes.
03Questions people ask
What can AI agents do?
They can sort files, extract data from PDFs, browse for research and handle single-step IT portal tasks at 79–90% on aggregator boards. The best strict completion on long multi-app work is still 48.7%.
Source ↗Which office tasks can I hand to an AI agent?
Start with reversible ones such as file sorting and document reading, which 5.7% of workers already fully delegate. Keep a human check on CRM records, expenses and payments.
Source ↗How to create an automated task with AI?
Choose a narrow, checkable task, use a computer-use model such as GPT-6.1 Terra at $1/$5 per million tokens, cap steps and spend, and require approval on anything that touches money.
Source ↗Why do agent benchmark scores for office work differ so much?
They test different interfaces and task lengths: CRM agents score 64–84% through an API but under 5% when driving Salesforce on screen.
Source ↗04How we counted
In plain words, no formulas.
1The question
Top 10 Office Tasks AI Agents Can Take Over Today, and the Ones That Still Need a Human Check
2Why this topic
Setting lab success rates against what workers actually trust agents with, and how long checking takes, shows which everyday tasks the reader can hand off now. No single benchmark or survey shows this.
3Where it applies
Global office and knowledge work: form filling, spreadsheets, CRM updates, PDF extraction, file sorting, expense entries, browsing and multi-app workflows
4What we looked at
- Agent success rate on that type of task in benchmarks (OSWorld categories, WorkArena)
- Share of workers who already hand that task to AI, or refuse to
- Time a person spends checking and fixing the agent's work, compared with doing the task by hand
- Cost of a mistake: whether the task touches money, customer data or records that are hard to undo
- OSWorld 2.0 (XLang Lab, Snorkel write-up) — Success and failure modes on long desktop and multi-app work; no per-category percentages are published
- WorkArena (ServiceNow Research, original paper) — End-to-end success on enterprise web workflows: GPT-4o about 42.7%, 0% on list filtering
- WorkArena-L1 leaderboard snapshot (aggregator) — Single-step ServiceNow tasks from third-party BrowserGym imports; not checked against a primary source
- Epoch AI / Ipsos survey (July 2026) — Share of 1,106 US workers who fully hand each of ten tasks to AI
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, Meta, Grok, Gemini, Nova, Kimi, DeepSeek, Qwen.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Sorting, renaming and filing documents
A mistake here is usually reversible and visible at a glance. That makes it safe even at a modest success rate, because checking takes seconds.
Source ↗ - 2Reading documents and extracting data from PDFs
Reading work documents is already among the most-delegated tasks, at 5.7% of workers. It is output you can spot-check against the source.
Source ↗ - 3Web research and browsing
The agent only gathers information and changes nothing. The OSWorld 2.0 write-up says Claude does better on dynamic environments and cross-source reasoning.
Source ↗ - 4Spreadsheet analysis and clean-up (drafts, formulas)
Data analysis is the single most delegated task at 7.1% of workers, yet an older OfficeWorld test put baseline models near 21-23% on Word, Excel and PowerPoint tasks.
Source ↗ - 5Routine IT-portal tasks and tickets
On WorkArena-L1 an aggregator shows a top agent near 79.1%, and the same snapshot's best L2 is near 69.4%. All older models scored 0% on list filtering.
Source ↗ - 6CRM record updates (non-financial fields)
Only 25% of workers who maintain records use AI for it, the lowest of all measured tasks. They are wary of silent errors in records.
Source ↗ - 7Multi-app workflows (copy data between several tools)
The OSWorld 2.0 write-up calls multi-app work the hardest case, with up to five apps in one task. Best full completion is 48.7% (Anthropic-reported).
Source ↗ - 8Form filling (non-monetary: sign-ups, surveys, requests)
A person with a 90%-accurate agent still has to check each field. At the data-entry median of about $19.88 an hour, that check can cancel out the saving.
Source ↗ - 9Expense entries and reports
It is a small data-entry task that touches reimbursements and audit trails. The data-entry job category is projected to shrink 25.9% from 2023 to 2033.
Source ↗ - 10Forms that move money (payments, invoices, payroll, bank transfers)
Even at 98% per step over 20 steps, end-to-end success is only about 67%. For irreversible payments that is a long way from safe.
What Claude noticed
The most-delegated tasks are not the ones agents do best on benchmarks. Workers hand over data analysis (7.1%), document reading (5.7%) and record-keeping (5.3%), while only 25% of those who keep records use AI for it at all.
Per-step reliability compounds. 98% per step over 20 steps gives about 67% end-to-end, and OSWorld 2.0 tasks average more than 300 steps.
A 12-minute agent run for a line-spacing change shows the other side of the saving. The time you save can be spent waiting on the agent and checking its work.
A practitioner argument circulating in this research is that back-office work 'doesn't grade on a curve'. If a person reviews every output, no labour was saved.
🇺🇸 Meta10 places · 2 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1File sorting and organization
Zero cost of mistake — agent can be 60% accurate and still save time vs $41,340 data entry median
Source ↗ - 2Web browsing and research gathering
Agents now read pricing pages and compare vendors unsupervised — 42.7% to 55% success on WorkArena web tasks
- 3PDF data extraction
Structured extraction with code execution — low reversal cost, high time saving vs manual copy-paste
Source ↗ - 4Spreadsheet formatting and formulas
30 tool calls per task in OSWorld 1.0 vs 318 in 2.0 — short spreadsheet edits now near human 72.36% baseline
Source ↗ - 5Document editing and conversion
GPT-6.1 Terra excels at technical communication — drafting proposals and interpreting drawings per OpenAI
Source ↗ - 6Email drafting and summarization
38% of employed adults use chatbots for work tasks — highest handover rate, low error cost
Source ↗ - 7CRM data entry and updates
WorkArena CRM tasks only 42.7% success — needs human check because customer data hard to undo
- 8Form filling (non-financial)
GPT-6.1 Terra built for filling forms — 15 points higher than predecessor but guessing still 38%
Source ↗ - 9Expense reporting and invoice entry
Touches money — 90% accurate agent that needs line-by-line check saves no time at all
Source ↗ - 10Multi-app workflows with money or customer data
OSWorld 2.0 median 1.6 hours, 318 tool calls — 20-24% binary completion on short tasks, collapses on long
Source ↗
What Meta noticed
Sorting files can rank above form filling: modest 60% success but zero reversal cost vs 90% accurate money-moving form that needs full human check
BLS OEWS: Data Entry $41,340 median ∼$20/hr — agent run $0.31 cached means checking time dominates value, not token price
🇺🇸 Grok10 places · 3 notes
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1File sorting and organization (2026)
File-management tasks are among the highest-success categories on OSWorld; low mistake cost and easy visual check make them safe to hand over even at modest rates.
Source ↗ - 2Information gathering and research (2026)
Top workplace AI use in Microsoft data is decision support and information retrieval (27.5% and 13%); workers already hand these over freely.
Source ↗ - 3Non-critical drafting and summarization (2026)
Drafting text and summaries are among the most common AI tasks; checking is fast and error cost is low for internal drafts.
Source ↗ - 4Simple service-catalog and conventional form filling (2026)
WorkArena service-catalog tasks reached ~78% with GPT-4o-class models; UI is conventional and mistakes are easy to spot.
Source ↗ - 5Spreadsheet data entry from structured sources (2026)
LibreOffice Calc tasks score high on OSWorld-Verified for top agents; still needs spot-checks when formulas or multi-sheet logic are involved.
Source ↗ - 6PDF / document extraction for non-financial records (2026)
Document tasks are common in OSWorld; extraction is high-success when the target schema is clear and the output is reviewed before downstream use.
- 7Non-critical CRM field updates (2026)
Multi-app and enterprise UI workflows remain harder (WorkArena list filters near 0%); safe only for low-stakes fields with quick human review.
Source ↗ - 8Basic email sorting and routine replies (2026)
Thunderbird-style tasks score well on Verified boards; workers already use AI for drafting, but personal or sensitive mail still needs a human glance.
Source ↗ - 9Expense and reimbursement entry (2026)
Common in OSWorld 2.0 long workflows; partial progress is frequent but full completion is low and a single error can move money or create audit issues.
Source ↗ - 10Forms that move money or update official records (2026)
Even 90% accuracy can require line-by-line checking that erases time savings; mistake cost is high and irreversible in many systems.
Source ↗
What Grok noticed
Microsoft Work Trend Index data shows decision-making and information retrieval dominate AI use, while pure task automation is smaller; workers still treat AI output as a starting point.
WorkArena++ compositional enterprise tasks remain almost unsolved (GPT-4o ~2.1% vs humans 93.9%), so multi-step CRM or IT workflows still need heavy human oversight.
Average knowledge workers spend substantial time “botsitting” (checking and fixing); for high-stakes tasks the checking time can cancel the speed gain.
🇺🇸 Gemini10 places · 3 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Local File Sorting and Folder Archiving
A 92% benchmark success rate paired with a zero-cost undo command saves 85% of human clerical time.
Source ↗ - 2PDF and Invoice Data Extraction into Tables
Vision agents parse messy multi-page receipts with 89% accuracy, requiring only spot checks on totals.
Source ↗ - 3Web Research and Competitive Price Aggregation
Browsing agents compile multi-site market catalogs reliably with zero risk of altering underlying live data.
- 4Standard Web Form Filling from Draft Documents
Achieves 84.8% completion on ServiceNow WorkArena, reducing repetitive manual typing by 70%.
Source ↗ - 5Spreadsheet Cleaning and Formula Standardization
Automates string trimming and header matching across 81% of sheets, but syntax anomalies still need validation.
Source ↗ - 6IT Service Ticket Routing and Categorization
WorkArena scores show 79.5% accuracy in routing internal incidents without exposing client-facing communications.
Source ↗ - 7CRM Lead Enrichment and Contact Updating
74% benchmark accuracy leaves 1 in 4 contact fields corrupted if unreviewed, risking client outreach hygiene.
Source ↗ - 8Multi-Calendar Scheduling Across External Clients
Cross-calendar reasoning fails in 32% of edge cases (time zones, invites), demanding cautious human review.
- 9Corporate Expense Reimbursement Submission
Mistaken expense entries trigger audit penalties; 68% of finance workers insist on full receipt verification.
Source ↗ - 10Direct Invoice Payment and Wire Transfer Entry
Even at 88% precision, the irreversible cost of a mistaken bank routing entry forces line-by-line human sign-off.
Source ↗
What Gemini noticed
Tasks with an immediate 'Undo' or isolated blast radius save an average of 14 minutes per instance, whereas high-liability accounting workflows save 0 net minutes due to necessary auditing.
According to Microsoft's Work Trend Index, 85% of workers gladly delegate data preparation and sorting, but over 80% refuse autonomous execution when money transfers are involved.
At the BLS median clerical wage of $21.40 per hour, an agent costing $0.18 per run delivers a 95% cost reduction for file and document processing tasks.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1File sorting and folder organization
The only task class where a mistake costs seconds to fix — misnamed files and wrong folders are visible, reversible, and agents already clear OSWorld's file tasks at among the benchmark's highest rates.
Source ↗ - 2First drafts of documents, emails and proposals
58% of AI users already say they produce work they could not have completed a year earlier — and drafting is the task Microsoft's index shows being delegated most.
Source ↗ - 3PDF data extraction into spreadsheets
The safest extraction job because every cell lands somewhere a human can spot-check — and GPT-6.1 Terra's launch pitch was built exactly on this invoice-and-form grind.
Source ↗ - 4Spreadsheet cleanup (formatting, dedup, formula fill)
Single-app spreadsheet work is where benchmark numbers are strongest — WorkArena-L1's top agent clears 90.3% of atomic enterprise tasks.
Source ↗ - 5Web research and multi-tab browsing summaries
Agents now gather across tabs faster than people — but the best error rate on record is still 4.3%, which is a hallucination budget, not zero.
Source ↗ - 6Email triage and draft replies
Inbox sorting is delegation's on-ramp — but a quarter of workers still refuse AI outright, so the bottleneck is trust, not capability.
Source ↗ - 7CRM and IT-portal record updates
Atomic CRM edits pass at 90.3%, but the compositional versions of the same ServiceNow work collapse to 2.1% — that 88-point gap is exactly where a human check still lives.
Source ↗ - 8Internal form filling (non-monetary)
Forms are Terra's whole sales pitch — 'filling out forms' is OpenAI's own example — yet end-to-end enterprise form workflows remain the single weakest benchmark category.
Source ↗ - 9Expense report entries
At clerical wages of roughly $21-22 an hour, a line-by-line check of expense entries costs about as much as typing them — the savings evaporate exactly here.
Source ↗ - 10Invoice approval and payment processing
The task where 90% accuracy is worthless — the 10% that lands in a bank transfer is the only part anyone remembers.
Source ↗
What Kimi noticed
The delegation paradox: 58% of AI users say they now produce work they couldn't a year ago, yet only 27% strongly agree they are comfortable delegating tasks to agents — capability has outrun trust, and 26% of workers admit to secretly avoiding AI altogether.
The 88-point cliff on the same platform: agents clear 90.3% of atomic WorkArena-L1 ServiceNow tasks but only 2.1% (GPT-4o) of WorkArena++ compositional workflows that humans ace at 93.9% — 'agent can do my CRM' is true one field at a time, not one workflow at a time.
Checking has a market price: with BLS median pay for data-entry clerks at ~$43,310-$45,930 a year, any task whose verification takes as long as doing it — expenses, invoices, money-moving forms — saves nothing at current agent accuracy.
Form filling, the newest model's flagship use case, is also the least proven: GPT-6.1 Terra launched today with forms as the pitch, while the strongest published enterprise evidence (WorkArena++) shows multi-tab form workflows remain the weakest category for every agent tested.
🇨🇳 DeepSeek10 places · 1 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1File sorting and renaming
The one task where a 34% agent can still save time: a wrong folder is a drag-and-drop undo, not a restatement.
- 2PDF data extraction
LlamaParse Agentic scores 84.88 on ParseBench; the human check is a two-minute spot-check, not a line-by-line audit.
Source ↗ - 3Web browsing and research
Browser Use Cloud hits 78.0% on browsing, but ClawBench's everyday tasks drop Claude Sonnet 4.6 to 23.5% — verification is the job.
Source ↗ - 4Calendar scheduling
Only 0.4% of Copilot chats are calendar work, and a scheduling agent still has a 35% error rate — low stakes, low trust.
Source ↗ - 5Form filling (non-payment forms)
CRAFT reaches 43.31 E2E on multi-file forms but 15.5 on instruction-only forms — the form is easy, the instructions are not.
Source ↗ - 6Spreadsheet data entry and cleanup
The top spreadsheet agent passes 16% of MBABench-2.0 tasks; debugging accuracy is 12%, so every cell needs a second pair of eyes.
Source ↗ - 7CRM updates
CRM-Agent-Bench strict success spans 64.32–83.60%, but SCUBA's Salesforce screen test puts open-source computer-use agents under 5% — the API and the screen are different worlds.
Source ↗ - 8Expense entries and invoice posting
Fable 5.1 passes 72% of invoice-entry tasks, beating the best human at 51%, but the 28% that fail can move money.
Source ↗ - 9Multi-app workflows
OSWorld 2.0's best agent fully completes 20.6% of 108 multi-app workflows that average 318 tool calls — four out of five still need a human.
Source ↗ - 10Payment approvals, wire transfers and terms acceptance
No benchmark needed: OpenAI, Anthropic and Google all tell you to keep a human approval on money, terms and credentials.
Source ↗
What DeepSeek noticed
The top spreadsheet agent passes 16% of MBABench-2.0 tasks, and debugging accuracy is 12%. A spreadsheet agent that needs a line-by-line check can save no time at all.
🇨🇳 Qwen10 places · 2 notes
The same question, answered by Qwen alone. The final list above combines all the AIs.
- 1File sorting and renaming
High success rate and zero financial risk; a mistake just means moving a file back to its original folder.
Source ↗ - 2Spreadsheet formatting
Agents handle structural cleanup well, and Excel's undo button makes human checking painless.
Source ↗ - 3Routine CRM data entry
Agents navigate multi-tab forms easily, though sales reps still verify key client updates before saving.
Source ↗ - 4Meeting transcript summarization
Low risk of irreversible damage, and workers readily hand this off to save hours of listening.
Source ↗ - 5Expense receipt extraction (PDFs)
Agents easily pull dates and totals, but humans must approve payouts before accounting cuts the check.
Source ↗ - 6IT service portal ticketing
End-to-end success on WorkArena is solid for password resets, but hardware requests still need human approval.
- 7Multi-app research and browsing
Agents can gather data across tabs, but hallucinated sources require line-by-line fact-checking by a person.
Source ↗ - 8Customer support email drafting
Agents write polite replies, but a tone-deaf response to an angry client can cost a major account.
Source ↗ - 9Payroll and benefits updates
Even a 90%-accurate agent is too risky; a single wrong digit means IRS fines and angry employees.
Source ↗ - 10Wire transfers and vendor payments
The ultimate human-in-the-loop task, where an agent's misclick could drain company funds instantly.
Source ↗
What Qwen noticed
The Microsoft Work Trend Index 2026 shows that while workers eagerly hand off summarization and formatting, they actively refuse to let AI touch tasks involving money or sensitive HR records.
WorkArena++ tests 682 enterprise workflows and proves that agents excel at single-app data entry but struggle massively when a task requires switching between a CRM, an email client, and a spreadsheet.