The paid work is in software that has no API
A script can fill in a form when the software gives scripts a way in. Plenty of business software doesn't, and a computer use agent works around that by doing what a person does: it reads the screen, clicks and types. That is the whole commercial case. Anthropic's page on using Claude Code with Chrome lists what people can do with it. They can test a web app, pull data off pages, and automate "repetitive browser tasks like data entry, form filling, or multi-site workflows". Its sample prompt hands Claude a file of contacts and asks it to add each one to a CRM.
The clearest buyers are people stuck with old software. Cyberdesk, a company backed by Y Combinator, sells developers a Windows agent to automate "EHRs, ERPs, and other legacy desktop apps". Its page names no customer, no result and no price. Deck sells credential and session handling for agents, and its own post calls authentication the main blocker: sessions expire without warning, and multi-factor sign-in is hard to automate. Both sell tooling, so the finished work, the part a client pays a person for, is open to anyone who can run an agent and stand behind the result.
We searched for a freelancer, virtual assistant or small agency that publishes what it charges for agent-run work and found none with a page you can open. What came up were course pages and income claims on blogs. The market is new, so there are no posted earnings to copy. What you can build on is what buyers pay today for the same work done by people. Our LLM money guide covers who has posted numbers for other routes.
Seven jobs, and where a person stays in
Each job below has a buyer who already pays a person for it. Start where a wrong cell is cheap: price checks, lead research and QA runs. Invoice entry and anything that files a form for a client can cost real money when the agent misreads a digit, so those wait until you have a month of clean runs.
| Job | Who pays for it now | What the agent does | Where a person stays in |
|---|---|---|---|
| Data entry into an old system | Clinics, shops and firms whose software has no import | Reads a spreadsheet or scanned form and keys each record into the screen | Counts records in and out, and reads a sample against the source |
| Product listings across marketplaces | Online sellers with one catalog and several storefronts | Copies title, price, photos and stock from a master sheet into each seller dashboard | The seller owns the account and presses Publish after checking prices and claims |
| Invoice and receipt entry | Small firms and the bookkeepers who serve them | Reads each document and enters it into the accounting program | A bookkeeper approves amounts, coding and tax before anything posts |
| Lead research | Sales teams and small agencies | Visits company sites and records name, role, contact page and one fact in a sheet | Spot-checks the facts, and skips sites whose terms ban automated access |
| Form filling | Brokers, clinics, schools and anyone who files the same form again and again | Fills portal forms from the client's file | The client signs, and banking and health portals stay out of the agent's reach |
| Web app QA runs | Small software teams with no tester | Follows a written checklist in the browser and reports what broke, with screenshots | A person decides what counts as a bug and reruns what failed |
| Price and stock checks | Shops that watch their rivals | Opens each competitor page on a schedule and records the price and availability | A person reviews the outliers, and sites that ban automated access are left out |
What buyers pay people for the same work
Pay for the people doing this work runs from a data entry keyer's median to a software tester's, and the average bids on Freelancer.com's listing pages sit under the data entry median. Those are bids on live projects, not budgets, and the pages don't say who won.
That's the trap in selling hours. An agent can't win a race against a bid of a few dollars an hour, and a client who has seen those bids won't pay extra for the word agent. The edge is the work a script can't reach: software with no API, and jobs that cross apps.
| Who or what | Price | What the figure is |
|---|---|---|
| Data entry keyer | $19.88 an hour | US median pay, 2025 |
| Bookkeeping clerk | $24.36 an hour | US median pay, 2025 |
| Administrative assistant | $22.86 an hour | US median pay, 2025 |
| Insurance claims and policy processing clerk | $23.67 an hour | US median pay, 2025 |
| Software QA analyst or tester | $50.14 an hour | US median pay, 2025 |
| Freelancer.com data entry listings | $2 to $18 an hour, or $20 for an Excel product entry project | Average bids on live listings |
| Freelancer.com bookkeeping listings | $5 to $18 an hour, or $176 to turn statements, receipts and invoices into a spreadsheet | Average bids on live listings |
| Freelancer.com lead generation listing | $378 for lead generation and CRM setup | Average bid, 51 bids |
| Freelancer.com website testing listing | $11 for a website user test with a bug report | Average bid, 41 bids |
| Time etc, human assistants by the month | $390 for 10 hours, up to $2,160 for 60 | Plan prices |
| Fancy Hands, human assistants by the request | $35 for 3 requests, up to $125 for 15 | Plan prices, each request up to 20 minutes |
| Prisync, price tracking software | $99 a month for 100 products, up to $399 for 5,000 | Plan prices, tracked by link |
| One invoice entered by hand | $13.11 | Ardent Partners' average for 2022, as DocuClipper cites it |
Agents still miss most real freelance projects
Check every output, because even the best agents get a minority of real projects past a human judge. The Remote Labor Index, run by Scale AI and the Center for AI Safety, gives agents 230 private projects taken from freelancers on Upwork. People then judge each result, and the score is the share of projects where the agent's work is judged as good as the human's or better.
The leaderboard's top agent, Claude Opus 5.5, reaches about one project in five. When the index launched, the best agent managed 2.5%. The curve is steep, and about four projects in five still fall short.
The projects include 3D design, video and audio, so routine data entry may score better, and nobody has published a rate for it. The index also leaves out projects that need direct client contact, and your jobs will have some.
A data job gives you a count of records in and a count out, so a gap shows at once. Then read results against the source: every record at first, and a sample once the runs have come back clean for a month.
Source: Scale AI: Remote Labor Index leaderboard (read November 8, 2026)
Charge per batch or per month
Buyers already pay for this kind of help in three shapes. Time etc sells blocks of a person's hours each month. Fancy Hands sells bundles of small requests, each up to 20 minutes of work. Prisync, the price tracking software, charges by how many products you track. Match the shape to the job.
A one-off, such as loading a catalog onto three marketplaces or clearing a backlog of receipts, fits a price per batch. A job that returns every week or month, such as price checks or invoice entry, fits a monthly fee set by the number of items, products or documents. Don't price by the hours the agent runs. The client never sees them, and your real cost is the plan's allowance.
Promise a delivery day, not overnight. Anthropic's page says Claude pauses at a login page or a CAPTCHA and asks you to handle it. A run that starts at midnight can sit at a sign-in screen until you wake up. Our agency guide covers charging for upkeep when a site changes, and its contracts section covers client data.
A flat monthly fee on top of an allowance the vendor can cut is a risk. Our Grok Bot post covers how usage there has run out faster since 1 November.
What the plan allowance does to a job
Anthropic says Cowork consumes limits faster than chat and gives no multiple. OpenAI's Work shares one allowance with Codex, counted in five-hour windows with weekly limits on top. On Plus, its estimates run from 5 to 45 local messages a window on GPT-6 Astra to 350 to 3,000 on GPT-6 Luna. They are estimates, not fixed limits.
Plans also decide who can do the work at all. Claude's computer use is in beta for Pro, Max and Ultra, and Team and Enterprise seats don't have it, so an agency can't put its staff on Team seats and get computer use. ChatGPT's Pro plans have no five-hour limit, and Plus and Pro can buy extra credit when a window runs out. Past a plan, an API key bills by the token, and the computer use tool counts every screenshot as image input, so cost grows with the number of steps. OpenClaw has no plan: the software is free and the model's bill is yours.
No vendor page says how much of a week one screen task uses. Run one batch on a quiet day, note the allowance it spent, and quote from that. For which agent suits which plan, the comparison lines up the whole market.
A site's terms can end the job
eBay's user agreement bars any robot, scraper or "other automated means" from its services for any purpose. It names "buy-for-me agents, LLM-driven bots, or any end-to-end flow that attempts to place orders without human review". The only exception is eBay's prior express permission. The version in force took effect on 28 June 2026. A listing job on eBay needs eBay's permission or its own bulk tools.
Whether a court agrees is still unsettled. Amazon sued Perplexity over Comet, an agent that shops for users, and won an injunction in March. An appeals court lifted it in August. It found that on the current record the user, not Perplexity, was the one accessing Amazon, because Perplexity's assistant was "a tool acting on the user's instructions". The court limited the ruling to the record before it and set no wider rule for agents. The case went back to the district court.
That case was about Amazon's accounts and Amazon's rules. A client's accounts belong to the client, so get the client's written permission before an agent signs in, and leave out any site whose terms ban automated access.
Amazon sues Perplexity
Over Comet, a browser agent that shops on Amazon for users.
Source: eWeek: Amazon wins injunction against Perplexity (read November 8, 2026)
A judge blocks Comet on Amazon
Strong evidence that it entered users' accounts with their permission but without Amazon's authorization.
Source: eWeek: Amazon wins injunction against Perplexity (read November 8, 2026)
An appeals court lifts the block
On the current record, the user and not Perplexity accessed Amazon.
Source: Retail Insight Network: US court overturns Amazon injunction (read November 8, 2026)
Client data, approvals and logins
A screen agent sees whatever is on the screen. Anthropic's help page says Claude can see "any information visible on your screen or those apps". It advises against giving computer use banking, healthcare or government apps, and it strongly advises against using it to manage or act on sensitive information. On a personal plan, that points away from some of the best-paid jobs on the list, such as claims forms and health portals.
Health data is where the paperwork gets heavy. The Claude Code page says a business associate agreement, the HIPAA contract, doesn't cover data sent to third-party sites through Claude in Chrome. The extension is off by default in Enterprise organizations that turned HIPAA on. Cyberdesk's page shows HIPAA and SOC 2 badges for an agent built for clinic software, which is a different product from a personal plan.
Anthropic's privacy page says consumer chats and coding sessions are used for training if you allow model improvement in your privacy settings, or if a conversation is flagged for safety review. It doesn't say whether a computer use session counts as a chat. Check that switch before a client's screen goes through a personal account.
Both vendors tell you to stay present. Anthropic's safety page says not to schedule tasks that access sensitive files, send messages for you or make purchases. OpenAI's Computer Use docs say to avoid tasks that need secrets "unless you're present and can approve each step". An agent run is a job you supervise, not one you start and leave. Quote your hours of watching and checking, because the agent's own time is the cheap part.
Pages and files carry risk too. Anthropic calls web content "a primary vector for prompt injection attacks" and says safeguards leave the chance "non-zero". A client's supplier site or a scanned PDF can hold text written for the agent. Give it a login with only the rights the job needs. If you run OpenClaw instead, its security policy says the sandbox is off by default, so the agent has your computer account's access, and everyone who can message it can steer it. Run one agent per client.
Computer use has no sandbox between Claude and your applications.
A first paid job, step by step
Pick one job from the cheap-mistake end of the table, for one client who has been paying someone to do it. Price checks on a few dozen products make a good first one: the agent reads pages and changes nothing, and a wrong cell is easy to catch. Give the agent a brief with edges, like this one. The brief is ours.
Open the shared sheet "Supplier prices". It has 40 products, each with a link to a competitor's page.
For each product, open the link, then record the price and whether it is in stock.
Don't sign in anywhere, don't add anything to a cart, and stop if a page shows a CAPTCHA.
Write the results to a new tab, with the time of each check.
List every product you could not read, and why.- Get the client's written permission for the accounts the job touches, and ask for a login limited to what the job needs.
- Run ten items first with approvals on, watch the whole run, and note how much of your allowance it used.
- Compare the count of rows in and out, then read every result against its source page.
- Quote the job by the item, with a delivery day that leaves room for sign-in and CAPTCHA stops.
- Hand over the sheet with a log of what was skipped and what you fixed by hand.
If the same workflow fits many clients, selling it as a product instead of a service is covered in our guide to selling AI agents and workflows. This post is a reading of the pages listed under sources, on 8 November 2026. Nothing in it was run.
Frequently asked questions
What do I need to start?
There are three routes. ChatGPT's desktop app has Computer Use on Mac and Windows, and Work comes with every plan, with a small allowance on Free and Go. Claude's route starts at Pro. OpenClaw is free software where you pay only the model's provider. Pick the route you already hold a plan for.
Can the agent keep working with my laptop shut?
Not with Claude's computer use, which needs the desktop app and an active desktop, according to Anthropic's help page. ChatGPT's Computer Use takes over the pointer and keyboard on Windows, so the PC is busy while it works. The exception is ChatGPT Work's cloud computer on the web and phone, whose tasks don't need your laptop.
Do I need to write code to sell this?
No. The sample prompts on the Claude Code and Chrome page are plain sentences that name a file, a site and what to fill in. Code comes in only with an API, where your own program carries out each click.
Is demand for this work growing or shrinking?
For the clerical jobs, shrinking. O*NET lists data entry keyers, bookkeeping clerks, administrative assistants and insurance claims clerks as "Decline (-1% or lower)" for 2025 to 2035, and software testers as "Faster than average (5% to 6%)". Those are Bureau of Labor Statistics projections of jobs held. They don't measure how much of the work agents do, so the data can't say whether it moves to agents, offshore or software.
Will the agent stop before it deletes or submits something?
It asks before some actions. Claude asks before permanently deleting files in any mode and before using each new app, and OpenAI says ChatGPT asks before sensitive actions. Neither page says it will always stop before a form is submitted, so say so in the brief.
Sources
- Claude Code docs: use Claude Code with Chrome https://code.claude.com/docs/en/chrome
- Cyberdesk: the computer use agent for Windows https://www.cyberdesk.io/
- Deck: automate legacy systems with computer use agents https://deck.co/blog/automate-legacy-systems-computer-use-agents
- O*NET: data entry keyers https://www.onetonline.org/link/summary/43-9021.00
- O*NET: bookkeeping, accounting and auditing clerks https://www.onetonline.org/link/summary/43-3031.00
- O*NET: secretaries and administrative assistants https://www.onetonline.org/link/summary/43-6014.00
- O*NET: insurance claims and policy processing clerks https://www.onetonline.org/link/summary/43-9041.00
- O*NET: software quality assurance analysts and testers https://www.onetonline.org/link/summary/15-1253.00
- Freelancer.com: data entry projects https://www.freelancer.com/job-search/data-entry/
- Freelancer.com: bookkeeping projects https://www.freelancer.com/job-search/bookkeeping/
- Freelancer.com: lead generation projects https://www.freelancer.com/job-search/lead-generation/
- Freelancer.com: software testing projects https://www.freelancer.com/job-search/software-testing/
- Time etc: pricing https://www.timeetc.com/pricing
- Fancy Hands: pricing https://www.fancyhands.com/pricing
- Prisync: pricing https://prisync.com/pricing/
- DocuClipper: the cost to process an invoice https://docuclipper.com/blog/cost-to-process-an-invoice
- Scale AI: Remote Labor Index leaderboard https://labs.scale.com/leaderboard/rli
- Remote Labor Index paper (arXiv 2510.26787) https://arxiv.org/abs/2510.26787
- Center for AI Safety: a significant increase in digital labor automation https://safe.ai/blog/significant-increase-in-digital-labor-automation
- Claude Cowork product page https://claude.com/product/cowork
- OpenAI docs: pricing https://learn.chatgpt.com/docs/pricing
- Claude Help Center: let Claude use your computer in Cowork https://support.claude.com/en/articles/14128542-let-claude-use-your-computer-in-cowork
- Claude Help Center: use Claude Cowork safely https://support.claude.com/en/articles/13364135-use-claude-cowork-safely
- OpenAI docs: Computer Use https://learn.chatgpt.com/docs/computer-use
- Anthropic privacy center: is my data used for model training https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training
- Claude API docs: computer use tool https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool
- OpenClaw security policy https://github.com/openclaw/openclaw/blob/main/SECURITY.md
- VentureBeat: OpenAI introduces ChatGPT Work, a cloud-based AI agent https://venturebeat.com/technology/openai-introduces-chatgpt-work-a-cloud-based-ai-agent-that-manages-tasks-across-email-slack-and-calendars
- eBay user agreement https://www.ebay.com/help/policies/member-behaviour-policies/user-agreement?id=4259
- eWeek: Amazon wins injunction against Perplexity over Comet https://www.eweek.com/news/amazon-perplexity-comet-injunction/
- Retail Insight Network: US court overturns Amazon injunction against Perplexity https://www.retail-insight-network.com/newsletters/us-court-overturns-amazon-injunction-perplexity