Images and Writing, Not Email: The Top 15 Jobs People Hand to AI, and How to Start Free on the Computer or Phone You Already Have

In 815 English-language posts where people describe the work they give to AI, the most common job is not email. It is generating images (48 mentions), then writing articles (41) and translating or proofreading (40). Email's best showing, writing and sending email, is only 15th (20), and all email jobs together add up to 50, about the same as images alone.

You don't need to spend money on hardware or subscriptions before you start. The AI runs in the cloud, so either the computer or the phone you already have is enough, and each of the 10 jobs below costs $0 in AI fees if you ask a free plan.

This article keeps two ways of using AI apart:

• Asking AI. You ask each time, and the AI does the work itself: writing, summarizing, looking things up. Free plans can do this.
• Running it unattended. The AI starts on its own, at a set time or when an email arrives, and you only check or approve the result. This is mostly where paid plans come in.

Asking AI already hands the work to the AI. You start paying when you want a job to run unattended, or when you ask for more than a free plan allows.

Free plans do have usage limits (checked in each vendor's help pages on October 19, 2026). Claude's free plan has a limit that resets every five hours (Anthropic), and Gemini's free tier has limits that reset every five hours and each week (Google); neither publishes a number. When you hit one, you wait for the reset or upgrade. ChatGPT Free has no cap on everyday text chats; the caps are on features such as image creation, file uploads and deep research (OpenAI). So if you mostly ask in text, you will rarely hit a limit on ChatGPT Free. If you hand AI a lot of work, or many images and files, a free plan may not be enough.

The other big difference is which AI model you get, the AI that actually writes the answer. At all three companies, the smartest models are on paid plans only (checked October 19, 2026):

• ChatGPT: Free and Go use the fastest, lowest-cost model; the more capable models start at Plus (OpenAI).
• Claude: the free plan gets Sonnet and Haiku; the more capable Opus starts at Pro (Anthropic).
• Gemini: on the free tier, most questions are answered by a light model built for speed; choosing the higher models yourself takes a paid plan (Google).

A free plan is often enough for everyday writing, summarizing and looking things up. But for smarter answers and more natural writing, when you want the most out of the AI, you need a paid plan: ChatGPT Plus ($20), Google AI Pro ($19.99) or Claude Pro ($20). ChatGPT Go ($8) uses the same light model as the free plan. For more usage, or to run jobs unattended, one plan including Go (about $8 to $20 a month) covers most jobs.

Part 3 sorted AI tools by what starts the work. This part starts from the job instead: what people hand to AI, what you need, how to start for $0, and when you start paying. It is written for sole proprietors and small businesses, not engineers.

The short version

• People hand AI images and writing far more than email. Images, articles, translation and blog posts are the top four. All email jobs together come to 50 mentions.
• No spending before you start. The AI runs in the cloud, so either your computer (our rule of thumb: a Mac from about 2018 on, or a Windows 11 PC, with 8 GB of memory) or your phone is enough.
• Asking AI costs $0 in AI fees. All 10 jobs cost $0 if you ask a free plan. Software that is not AI, such as accounting software, is a separate bill.
• Free and paid differ in three ways: how much you can use, which models you get, and running unattended. Free plans have usage limits too, so if you hand AI a lot of work, a free plan may not be enough. The smartest models are paid-only, so for smarter answers and more natural writing you need a paid plan. Running at a set time, or when an email arrives, is mostly a paid feature.
• To run a job unattended, add one paid plan. One of ChatGPT Go ($8), Google AI Pro ($19.99), ChatGPT Plus ($20) or Claude Pro ($20) covers most of the jobs you would run unattended; the table below says which plan fits which job.
• An AI-worker system costs more. Software that gives each AI worker a role and has a person approve its work before it acts (Paperclip is one example) is free, but the AI inside it runs on your paid plan. It burns through the plan's usage limit fast, which eventually pushes you to a higher plan or to paying per use. Why, and how to bring the cost down, is part 13 (what one reply costs and how to cut it).

Facts are as of October 19, 2026; prices were checked on October 16 and 17 and change often. USD prices exclude sales tax.

The top 15 jobs people hand to AI

I collected English-language posts from Reddit, Hacker News, Dev.to and Medium in which the writer describes work they give to AI, and pulled out each job. One job in one post counts once. When the same person posted the same text in several communities, I kept one copy.

That gave 815 posts and 1,619 jobs: Reddit 504 posts, Dev.to 173, Hacker News 109, Medium 29. I set aside the 539 software-development jobs, since this series is not about coding, and ranked the remaining 1,080.

The 15 jobs people most often hand to AI in English-language posts, by count (software development set aside).

The other email jobs rank lower: sorting and prioritizing email 29th (13), drafting replies 40th (8). Counting each person once, or putting back the 27 jobs I excluded as not really the writer's own delegation (see the end), leaves the top four the same (with those jobs back, blog posts ties for 4th).

The posts lean toward people who already talk about AI tools online, many of them developers, so treat this as what early users do, not the whole market. Earlier in this series I collected about 1,100 Japanese posts about "AI employees" as well. Adding them puts blog posts first and social media drafts second, but email stays far down: its best job is 23rd.

The next table groups every job by area, so you can find your own work outside the top 15.

By what starts the job, two thirds (720 of 1,080) start when a person asks each time. A schedule starts 197, something arriving (an email, a message, a file) 66, and a message to an always-on AI worker 66; 31 don't say.

What you need: either the computer or the phone you have

ChatGPT, Claude and Gemini run in each company's cloud. Your computer shows the screen and carries out the actions the AI asks for, such as reading and writing files or driving a browser. The heavy computing does not happen on your machine.

Most official requirements are met by computers several years old:

• Claude Code (an app that lets Claude work on your computer): macOS 13 or later, or Windows 10 or later, with 4 GB of memory
• The Claude desktop app: macOS 13 or later, or Windows 10 or later
• The ChatGPT Mac app: macOS 14 or later, on Apple's M1 or newer chips or on older Intel chips
• Gemini: only the Mac app needs an M1 or newer chip and macOS 15; the browser version has no such limit

Our rule of thumb is a Mac from about 2018 on, or a Windows 11 PC, with 8 GB of memory. That is our summary of the vendors' numbers, not a vendor statement. Windows 10 still runs these apps, but Microsoft ended support on October 14, 2025, so we do not recommend it for work data.

What runs where. The thinking happens in the cloud; your phone or computer asks, checks, and hands over local files.

To ask AI, either one is enough. A phone alone covers asking one request at a time, and checking and approving an AI worker's output. You need a computer when the AI works with files and folders on it, and to host an AI-worker system.

Running unattended splits in two. A paid plan's scheduled tasks run in the vendor's cloud, so they keep going with your laptop closed. An AI-worker system running unattended needs the computer it lives on (or a rented server) left on. One exception: scheduled runs that use local files, such as Claude Code's, run on your computer and only while it is awake. In every case, the phone is only for checking and approving; you can do without it.

Start with the computer or phone you have. Consider buying one only if it feels slow, or when you want an AI-worker system to keep running while your laptop is closed. That is when a machine like the often-mentioned Mac mini becomes an option; the numbers are in note B at the end.

10 jobs and how to start each

These are the 10 jobs this series sets up. I picked most of them from the top of the ranking, kept two that rank much higher once the Japanese posts are added (social media posts, 2nd, and meeting notes, 10th), and added the jobs I promised in part 3 (email, bookkeeping, Instagram, and replying to customers on chat). Translation, notes, task sorting and everyday questions are usually one request to a free chat, so they don't get their own row; competitor research goes with in-depth research, and image generation comes with social media posts in part 6. Family, health and learning are left out because this series is about running a business.

"Ask a free plan" is the $0 way to start. "Add one paid plan" is the cheapest plan I found on October 16 and 17, 2026, and what it lets you run unattended. For jobs you ask for each time, running them unattended matters less, so the column says what gets easier (such as asking more often). The last column names the part where I set the job up on a real account (parts 5 to 16 aren't published yet).

Notes on the table:

• Names. Meta Business Suite is Meta's free dashboard for scheduling Instagram and Facebook posts. Deep Research is a feature where the AI reads many pages and writes up what it found. ChatGPT Go is a paid plan cheaper than Plus. Insights is Instagram's screen for your numbers.
• Email. Outside the US, "Help me write" needs Google AI Plus or higher, or Workspace; thread summaries are free on personal accounts everywhere. Gemini's email-triggered replies need Gemini's Activity turned on, which also lets Google train on your chats (see Security, step 1); whether the reply lands in your Gmail drafts is checked in part 8. Claude's Google connectors work on every plan, and by default Claude asks before each send, reply or forward. You press Send yourself; part 1 (connecting ChatGPT dots to Gmail and more) shows why.
• Claude's two kinds of schedule. Scheduled tasks in the Claude app run in the cloud, so they arrive even when your computer is asleep or the app is closed. Scheduled runs in Claude Code run on your computer, so they only happen while it is awake. Reading your Instagram screenshots is the second kind.
• ChatGPT scheduled tasks. OpenAI's help page (updated October 18, 2026) says the free plan can run up to 3 tasks, at most once a day, in rough time windows such as morning or night. Exact times and hourly runs need a paid plan. For a job like collecting the news, where a little drift in delivery time doesn't matter, the free plan may be enough to run it unattended. Part 6 tries it on a real account.
• Bookkeeping. The main cost is the accounting software, not the AI. US list prices: Xero Early $27 a month, QuickBooks Simple Start $38. QuickBooks and Xero also publish connections that let an AI read and write your books, but today they are developer tools you run on your own machine; part 7 covers the simpler routes. Always check the entries the AI drafts.
• Customer chat desk. The WhatsApp Business app's greeting and away messages are free, but they are fixed replies, not AI. Connect AI only through official business routes (the WhatsApp Business Platform); automating a personal account risks getting the number blocked. Through the WhatsApp Business Platform, Meta charges per reply since October 1, 2026: the first 1,000 replies a month per number are free, then about $0.0034 each to US numbers, and without a payment method on file replies stop after 1,000. Part 12 (a customer desk where AI drafts and you approve) builds the AI version.

Five steps from asking a free plan to running unattended

Each step costs a little more and takes a little more work off your hands. Part 5 shows which plans to get, and parts 6 to 12 take one service or tool each.

Step 2 costs more because one run of an AI-worker system reads far more text than a direct request. In my own Paperclip setup, the median of my last 185 runs read about 460,000 tokens (the units of text an AI reads). Those were internal jobs such as blog posts and email, not customer replies. A direct reply to one customer message would read a few thousand tokens, by my estimate. Through an AI-worker system, even one customer message carries the instructions and records the system reads every time; how much that adds per message I have not measured yet, and part 13 does.

On a subscription the bill stays the same, but you hit the plan's usage limit much sooner. To keep running, you move to a higher plan or to paying per use, and that is where the cost goes up. A public post covered in part 2 (what "AI companies" really are), from carlginmunsayac on Facebook, described four Paperclip workers nearly using up a usage limit in one night. Paperclip's budget feature does not catch this: as part 3 showed, it records subscription runs as costing nothing.

A caution: letting a subscription answer your customers sits in a gray area of the terms. Anthropic's legal page says plan limits "assume ordinary, individual usage" and that developers building products or services on Claude should use the pay-per-use access sold to developers (called the API). Whether software you run for your own shop counts is not stated. I think it is reasonable for a small business to start on the plan it already has. Still, so a terms change can't stop a job you rely on, set up pay-per-use access as well and give it a monthly spending limit. Preparing it costs nothing; part 14 (start on a subscription, keep a pay-per-use switch ready) covers when to move.

Three combinations and their monthly cost

"Minimum" assumes you ask a free plan for the AI parts.

• Sole proprietor (blog, social media, morning briefing, bookkeeping): the minimum is accounting software only (from $27). Add Claude Pro at $20 and the briefing arrives every morning, and you can ask for blog and bookkeeping drafts more often. Scheduling posts stays free in Meta Business Suite.
• Small shop (WhatsApp desk, Instagram posts and numbers, bookkeeping): the minimum is accounting software only, with the WhatsApp Business app's free greeting and away messages. For AI drafts you approve, you need an AI-worker system (free), one paid plan such as Claude Pro ($20), a computer left on (or a rented server), and WhatsApp's per-reply fee beyond 1,000 replies a month. Meta's own Business Agent, in the WhatsApp Business app, is free to start where it is available (Meta lists only "select markets"); beyond a free limit, Meta One plans start at $14.99 a month. This route instead uses the AI plan and instructions you already have.
• Office worker (email, meeting notes, research, writing): the minimum is $0, assuming your company pays for the meeting app and you use a free AI. To cover all four with one plan, Google AI Pro at $19.99 gives you the AI inside Gmail, Meet notes and research, but only on a personal Google account. Before putting work email into a personal AI account, check your employer's rules; if your company has a business AI plan (Google Workspace, Microsoft 365, ChatGPT Business), use that, since business plans do not train on your data by default.

One AI plan covers the writing, summarizing and researching. Accounting software, the meeting app and paid chat-app options are separate bills.

Security: five settings before you hand business data to AI

Some personal AI plans train on your conversations unless you turn it off, meaning the company may use what you type to build its next AI models. Before you show customer data to an AI:

1. Turn off "use my data for training."
   • ChatGPT: Settings → Data controls → "Improve the model for everyone" → off (OpenAI). OpenAI says it uses conversations on personal plans for training unless you opt out (OpenAI). With that setting on, content read from connected apps can be used too (OpenAI), though data read directly from Google apps such as Gmail and Calendar is not (OpenAI). Turning the setting off is simpler than remembering which apps count.
   • Claude: Settings → Privacy → "Help Improve our AI models" → off. You chose this when you signed up, so check what you chose (Anthropic).
   • Gemini: gemini.google.com → Settings & help → Activity → Turn off. "Keep Activity" is on by default for users 18 and older (Google). Turning it off also turns off the Gmail and Google Workspace connections and tasks that run when an email arrives, so on a personal Google account you cannot keep your chats out of training and connect Gmail at the same time. For mail with customer data, use a Google Workspace business account, or don't connect Gmail and paste redacted text instead.
2. Replace customer names, phone numbers and addresses before you paste them. Write "Customer A" or "555-XXXX". Google says human reviewers may read Gemini conversations, and conversations are kept for 72 hours even with Activity off. The FTC's advice to businesses puts it simply: "No one can steal what you don't have."
3. Use a temporary chat for one-off questions. ChatGPT's temporary chats are not used for training (they are kept for up to 30 days for safety). Saving one turns it into a normal chat that follows your step-1 setting. Claude and Gemini have the same kind of chat.
4. Don't press thumbs-up or thumbs-down on chats with business data. Claude keeps the data from conversations you rate for 5 years. Gemini collects feedback even with Activity off.
5. When you connect an accounting app or chat app to AI, give only the permissions it needs. Anthropic advises reviewing the permissions a connection asks for, narrowing them where you can and denying any that seem unnecessary. The FTC's small business guidance says to "restrict sensitive information access to only those who need it to do their jobs"; it is written about staff, and the same logic applies to AI. And remove connections you no longer use: in QuickBooks, an admin can go to Integrations → Manage integrations → Disconnect (Intuit).

What's next, and tell me which job you want

The next part (part 5, which AI plans you actually need) turns plan choice into a table you can look up by what you want done. The hands-on parts start in part 6 and go one service at a time: ChatGPT, Claude, then Gemini, each taking one or two jobs from the table above.

If there is another job you want covered, tell me in the comments. I will add an article.

The rest is extra detail for those who want it.

Note A. If you want the AI itself at home or in the office

Everything above assumes the AI runs in the cloud. If you would rather keep data in-house, or cut per-use costs, you can run an AI model on your own machine (a local LLM). It is optional.

• The Mac you have can run one. Ollama and LM Studio are free and need Apple silicon (M1 or later) and macOS 14 or later. LM Studio recommends 16 GB of memory and says 8 GB Macs can work if you "stick to smaller models and modest context sizes."
• Memory mostly decides how big a model you can run. Until February 2026, Ollama's README suggested 8 GB for 7B models (the number is roughly the model's size), 16 GB for 13B and 32 GB for 33B; it has since dropped that line. Today's default downloads: Llama 3.2 3B is 2.0 GB, Qwen3 8B 5.2 GB, Gemma 3 12B 8.2 GB, Gemma 3 27B 17 GB.
• Speed. In user benchmarks collected by the llama.cpp developers (a supporting source; 7B model, 4-bit), an M1 managed about 14 tokens per second and an M2 or M3 about 21. Try your current Mac first, and buy hardware only if the speed bothers you.
• The weak spot of small models. They suit summarizing, rewriting and following instructions. OpenClaw's docs warn that "smaller/heavily quantized models are more vulnerable to prompt injection" (instructions hidden in text that trick the AI) and recommend large models for any agent that uses tools. If you can't run a large model at home, a cloud model is the safer choice for jobs where the AI reads text from strangers and then acts, such as a customer desk.
• If you buy. Mac mini from $899, Mac Studio from $2,499, NVIDIA DGX Spark 64GB from $4,999 (on sale October 23, 2026); NVIDIA aims it at "developers, researchers and AI enthusiasts." New hardware makes sense only for bigger models or when your machine is too slow.
• Watch for "cloud" in model names. Ollama models tagged cloud (such as gemma4:cloud) run on Ollama's servers. To keep data at home, turn cloud off in Ollama (OLLAMA_NO_CLOUD=1).

Note B. A machine that runs while your laptop is closed: buy or rent?

To keep an AI-worker system running while your laptop is closed, you need a machine that stays on day and night. Here I compare buying a Mac mini as that always-on machine with renting a small server, over three years. The AI still runs in the cloud; this is not a comparison of local LLM costs (note A).

The always-on machine only holds the software that sends work to the cloud AI, so a small server is usually enough. OpenClaw's FAQ, for example, lists 1 GB of memory as the minimum and 2 GB or more as recommended. I have not checked how much memory Paperclip and the Claude Code it runs need; if 2 GB is not enough, use the 8 GB case below.

US list prices, checked October 17, 2026, before sales tax. Mac mini (M6, 16 GB memory, 256 GB storage): $899. Apple lists it at 4 watts idle and 70 watts maximum. At 4 watts its electricity is about $0.52 a month, at 17.99 cents per kWh (the US residential average for the 12 months to July 2026, from the EIA): 4 W ÷ 1,000 × 720 hours × $0.1799 ≈ $0.52.

Three-year cost of an always-on machine: a Mac mini against three rented 2GB servers, at US list prices before sales tax (as of October 17, 2026). The lower chart shows when the Mac mini pays for itself.

The servers: Vultr 2GB at $10 a month, DigitalOcean 2GB at $12, and Akamai (Linode) 2GB at $12. Over three years: $360, $432 and $432. The Mac mini: about $918 at 4 watts idle, about $1,225 at 70 watts all the time.

Break-even months = Mac mini price ÷ (server monthly fee − Mac mini monthly electricity). Against the $12 servers: $899 ÷ ($12 − $0.52) ≈ 78 months, about 6.5 years; against Vultr's $10 server, about 95 months. If your software needs 8 GB of memory, you compare against an 8 GB server ($48 a month at DigitalOcean or Akamai, $40 at Vultr), and break-even drops to about 19 to 23 months. The totals leave out power while the AI works, resale value, sales tax, your time looking after the machine, and your home internet.

A rented server can't always replace a Mac. OpenClaw's documentation names the cases: using iMessage or macOS apps, needing a home internet address (some sites reject data-center addresses), and needing a browser that stays logged in with a visible screen. Then a Mac mini makes sense.

For comparison, my own company pays for ChatGPT Pro, Claude Max 20x (the $200-a-month tier of Claude's higher personal plan) and Google AI Pro. Our always-on machines are two CPU servers in the office (ordinary computers without an AI chip), and we run AI models on one dedicated machine (an NVIDIA DGX, a computer built for running AI). I also own a Mac mini but don't use it for AI.

Note C. Points I am not stating as settled

• How ChatGPT's free scheduled tasks work in practice. OpenAI's help page (updated October 18, 2026) says Free can create up to 3 tasks, each running at most once a day in a rough time window, but I have not tried one on a free account. OpenAI also does not say outright that scheduled tasks run with your computer closed; it says only that tasks using local files need the computer on. Part 6 checks.
• Whether ChatGPT's free plan can connect Gmail. OpenAI does not say for Free or Go; sending email from ChatGPT needs Plus or higher. That is why the $0 email route above uses Gmail itself or Claude.
• Whether Zoom's free plan gives meeting summaries. Zoom's pricing page lists "meeting summaries (for 3 hosted meetings per month)" on the free plan; its help pages list only Pro and higher. Part 8 checks.
• Whether QuickBooks' own AI works on its cheapest plan. Not confirmed.

About the data behind the top 15:

• "In use" is the writer's own claim. I did not check that each setup actually runs.
• Most posts come from Reddit, and about 85% of the Reddit material comes from communities that discuss AI apps and services. Many writers are developers. Other people may use AI differently.
• Each job was typed by an AI (Claude) and checked by a second AI arguing against the first. A third AI rechecked 40 randomly chosen jobs and found no clear errors (7 were borderline).
• Jobs that weren't really the writer's own delegation, such as lists of app names or other people's stories (27 in the English data), were not counted.