# How to Automate Tasks with AI: 7 Jobs to Hand Over > How to automate tasks with AI in a small business: seven jobs to hand an AI agent this week, the exact ask and schedule for each, and four to keep. How to automate tasks with AI comes down to three choices per task: what you ask, how often it runs, and how you check the result in under two minutes. Get those right and an AI agent can take over the reading, watching and drafting that eats a small business's week. Get them wrong and you get confident summaries nobody reads, or worse, an email sent in your name. **The short answer:** hand an AI agent the jobs that repeat, read public pages or your own files, and end in a draft or a file you can check. Watching competitor prices, a morning brief on your niche, one-off research, a Friday summary, post drafts, an SEO check of your own site and a prospect list all qualify. Keep anything that sends, pays, publishes or submits in your name. Below is the exact ask and schedule for each. ## The rule: hand over if you can check it in two minutes Hand a task to an AI agent if all four are true: 1. **It repeats** on a cadence you can name (every weekday, every Monday), or it's a one-off question you'd otherwise spend an afternoon researching. 2. **The inputs are public pages or your own files.** Not a stranger's email, not a login you'd be unhappy to share. 3. **The output is a draft, a file or a message to you.** Nothing leaves your business without you. 4. **You can check it in two minutes** by opening one link or one file. Keep it if any one of these is true: it moves money, it sends something to a customer or prospect in your name, it submits a form on someone else's website, or a wrong answer costs more than the time it saves. ## How to automate tasks with AI: the decision table | Task | Good fit? | How to check the result | Cadence | | -------------------------------------------------- | ----------------------- | ---------------------------------------------- | -------------- | | Competitor pricing watch | Yes | Open the one page it says changed | Monday 09:00 | | Morning brief on your niche | Yes | Click the link on the item you'd act on | Weekdays 08:00 | | One-off research (a vendor, tool or supplier) | Yes | Open two of its sources; do the numbers match? | Once | | Friday summary of links and files you give it | Yes | It must say "nothing changed" when nothing did | Friday 16:00 | | Social post drafts | Drafts only | Every claim has a link you opened; you post | Monday 09:00 | | SEO check of your own site | Yes | Open three flagged URLs | Weekly | | Prospect list from public sources | The list only | Open five source URLs at random | Weekly | | Publishing AI-written pages at scale | No | — | — | | Cold email or LinkedIn outreach, sent by the agent | No | — | — | | Replying to customers | Not yet; drafts at most | Read every draft before it goes | — | | Paying invoices, issuing refunds | No | — | — | The first five cadences are the defaults on [Octomind Cloud](/cloud)'s starter jobs. Disclosure: Octomind is ours, and every run quoted below came from Octomind Cloud on a demo account. Move the times to when you actually read things: a brief that lands while you're asleep and gets buried under ten other messages is wasted. ## Write the ask so the agent can't wander Every ask that works on a schedule has five parts. Copy this skeleton and fill the blanks: ```text Check . Compare against what you found last time. Tell me only what changed: , one line each, with the link. If nothing changed, say so in one line. Save the full result to . ``` The fourth line does the most work. Without it, an agent asked to "report changes" finds something to report every time, and within two weeks you stop reading. With it, a quiet week costs you one line. Here is the competitor watch from the table, run for real on October 10 against Zapier's and n8n's pricing pages. The first run had nothing to compare against, so it said so and saved a baseline. The second run, the same day, came back with one line: _"Nothing changed — both zapier.com/pricing and n8n.io/pricing are identical to the baseline captured earlier today (prices, plans, and wording all match)."_ The conversation cost $0.66, as the panel reported it. That's the shape you want: a long answer once, then one line until something moves. **Run every new ask twice by hand before you schedule it.** The first run shows you whether the output is useful. The second shows you whether "compare against last time" works, which is where most scheduled jobs quietly fail: the agent has nowhere to keep the previous result, so every run is a first run. ## The five jobs to hand over first ### 1. Competitor pricing watch: Monday 09:00 ```text Check and . Compare them against what you found last time and tell me only what changed — prices, plans, wording. If nothing changed, say so in one line. ``` **Check:** open the one page it names. **Failure mode:** a pricing page that loads prices with JavaScript or shows a different currency by country. Fix: add "if you can't read a price, say which one and why" so a blank never reads as "unchanged". ### 2. Morning brief on your niche: weekdays 08:00 ```text Find what is new about since yesterday and write me a brief: the three things that actually matter, one line each, with links. Skip anything I would already know. ``` **Check:** click the link on the item you'd act on, every time. In our own October 10 run on AI agent news, two of the three items held up against their sources, including a [report from Anthropic](https://www.anthropic.com/research/investigating-unintended-model-actions) published the day before. The third overstated what its source said, and the reply still ended by calling all three verified. A one-line summary is a pointer, not a fact. **Fix:** ask for the source's own sentence under each item ("quote one line from the source"), so a mismatch is visible without clicking. ### 3. One-off research: once ```text Research online and write a report to report.md: the answer, the key facts and the sources. Then give me the short version here. ``` This is the ask for "which payment provider", "which accounting tool", "what does this regulation require". Our run on [Stripe vs Paddle for a two-person startup](/cloud/results/research-report) came back as a sourced report with seven numbered sources and an access date, for $0.06. **Check:** open two sources and confirm the numbers you'd base the decision on. Fees and terms change, so a report is good for the decision you make this week, not next quarter. ### 4. Friday summary: Friday 16:00 ```text Go through and write me a short report on what changed this week and what matters. If nothing changed, say so in one line. ``` Give it a fixed list, not "the internet". Our run over Hacker News, OpenAI's and Anthropic's news pages cost $0.01 and said plainly where it found nothing: _"No newsroom post dated Oct 3–10 was found."_ That sentence is the check. A summary that never reports an empty source is padding the others. ### 5. Social post drafts: Monday 09:00 ```text Write five post ideas with ready-to-paste captions for , based on what people are talking about this week. Match this tone: . ``` Drafts only. In our [run for a solo founder](/cloud/results/content-drafts) ($0.08), three of five posts cited sources we could check. The other two cited sources we couldn't, so they were cut before publishing. Budget for that ratio. **Fix:** add "one link per factual claim; drop any post you can't source", and you'll get fewer, usable drafts instead of five you have to audit. ## How to automate SEO with AI without getting penalized Automate the checking. Write the pages yourself. Google's spam policies list "using generative AI tools or other similar tools to generate many pages without adding value for users" as an example of [scaled content abuse](https://developers.google.com/search/docs/essentials/spam-policies) (checked October 22, 2026). Publishing a batch of AI-written pages is the one SEO task to keep off the list. What to hand over instead, with the ask: **Weekly site check.** An agent with a machine can write and run a small crawler: ```text Crawl https:// (same domain only, up to 500 pages). List every page that returns an error, has no title, has a title over 60 characters, or has no meta description. Write it to seo-check.csv with the URL and the problem. Compare against last week's file and tell me only the new problems. ``` **Check:** open three flagged URLs. If all three are real problems, trust the rest of the file. **Competitor page watch.** The pricing-watch ask, pointed at the three pages that outrank you: "tell me only new sections, headings or claims." **Content brief for one keyword.** "Search , list the headings of the top five results and the questions none of them answers. Save to brief.md." You write the article from the gaps. ## How to automate lead generation with AI, and where it becomes illegal The agent builds the list and drafts the first line. You send. ```text Find 20 in that . For each: name, website, the page where you saw the signal, and a contact address listed on their own site. Save to leads.csv. Skip any you can't source to a page. Do not contact anyone. ``` Then, as a second ask: "For each row in leads.csv, draft a two-line first message that mentions the signal. Save to drafts.csv." **Check:** open five source URLs at random. If one is wrong, check them all before you write to anyone. Keep sending manual, for two sourced reasons: - **Cold email has legal requirements.** In the US, CAN-SPAM requires an honest subject line and headers, your physical postal address and a working opt-out honored within 10 business days, and it "makes no exception for business-to-business email." Each violating email can cost up to $53,088, per the [FTC's compliance guide](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business) (checked October 22, 2026). An agent that sends on its own makes every one of those your problem, at volume. - **LinkedIn bans automation.** Its [prohibited software page](https://www.linkedin.com/help/linkedin/answer/a1341387) rules out bots and tools that "scrape, modify the appearance of, or automate activity" on the site, and accounts that use them can be restricted or closed. Don't let an agent log in as you there. ## What not to hand over yet | Keep it | Why | | ---------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Anything that moves money | A wrong payment is the one mistake a draft can't catch. | | Messages to customers in your name | One confident wrong answer costs more than the week of replies it saved. Let it draft; you send. | | Submitting forms on other websites | Agents overreach on the live web. On October 9, Anthropic [reported](https://www.anthropic.com/research/investigating-unintended-model-actions) Claude "submitting a sensitive form on a real website when it should not have" during evaluations and internal use, and cut live internet access from all its internal evaluations. | | Acting on what strangers send you | Text in an email or web page can carry instructions. OWASP ranks prompt injection first in its [2025 LLM Top 10](https://genai.owasp.org/llmrisk/llm01-prompt-injection/), including "indirect" injection from websites or files. An agent that reads untrusted input should only be able to write files and drafts. | ## Which tool for which job Not every task above needs an agent. Match the shape of the job: - **"When X happens in app A, do Y in app B"** (new form entry to a spreadsheet, new invoice to Slack): use an app-to-app tool. [Zapier](https://zapier.com/pricing) is free for 100 tasks a month, and Pro starts at $19.99 a month billed yearly for 750; its separate Agents feature is billed in "activities", 400 a month on Free. [n8n](https://github.com/n8n-io/n8n) is fair-code and self-hostable, with cloud plans from €20 a month billed yearly for 2,500 workflow executions ([pricing](https://n8n.io/pricing/)). [Gumloop](https://www.gumloop.com/) calls itself "the multiplayer AI agent builder" and lists Gmail, Slack, HubSpot and Google Sheets among its connections. All prices checked October 22, 2026. - **"Answer this once"**: a chat assistant like ChatGPT or Claude is enough. Paste the research ask above and read the sources. - **"Go read these pages every Monday and tell me what changed"**: an agent that keeps its own files between runs, so "compare against last time" has something to compare against. If the job lives inside your inbox or calendar, start with the app-to-app tools: that's where the Gmail connections are. ## Where Octomind fits Octomind covers the third bullet. Each workstream gets an octo, an agent with its own folder and memory, so the competitor watch can keep last week's snapshot and diff against it. A routine reruns the ask on its schedule, wakes the octo to do it, and sends the result to Telegram, Slack or WhatsApp. Every scheduled run, including one that was skipped, lands in a history row with its reason and cost: ![A routine's run history in the Octomind panel: a completed run with its report, a run with nothing to report, one that didn't run because the machine was in cold storage, and one stopped at its cost cap](/blog/routines-scheduled-agents/run-history.png) The limits, before you sign up: - **Octos sleep when idle** and wake on a message or a routine. - **No Gmail or calendar connector.** It can't read your inbox or calendar, so inbox jobs belong in Zapier, n8n or Gumloop. - **Routines need a paid plan.** The free plan has no cloud machine to run them on. If you'd rather describe the job in a sentence than fill in the skeleton, start from the [no-code agent builder](/use/cloud/agent-builder). For one octo that keeps your own errands and research, see [personal AI agent](/use/cloud/personal-ai-agent). How routines handle skipped runs and cost caps is in [Routines: your agent shows up without you](/blog/routines-scheduled-agents). ## Your first week - **Monday:** write the competitor watch. Run it twice by hand: once for the baseline, once to see the "nothing changed" line. Then schedule it for Monday 09:00. - **Tuesday:** write the morning brief. Run it, click all three links, add the "quote one line from the source" fix if any summary drifted. Schedule weekdays 08:00. - **Wednesday:** one research question you've put off. No schedule. - **Thursday:** the SEO site check on your own domain. Open three flagged URLs. - **Friday:** the Friday summary, then a ten-minute review: which results did you actually read this week? Turn off the ones you skipped. A job you don't read is a cost, not a time saving. Add the prospect list and post drafts in week two, once you trust the first five. ## FAQ ### How do I use AI to automate tasks without coding? Write the ask in plain sentences using the five-part skeleton above (source, compare, what changed, empty case, where to save), run it twice by hand, then put it on a schedule. App-to-app tools like Zapier and n8n also have no-code editors; they fit when the job is moving data between apps rather than reading and judging. ### What should a small business never automate with AI? Payments, messages to customers or prospects that go out without a person reading them, form submissions on other websites, and publishing AI-written pages at scale. Each has either a legal or policy cost (CAN-SPAM, LinkedIn's terms, Google's spam policy) or a mistake a draft can't catch. ### How do I know a scheduled AI task is still working? Make it say "nothing changed" in one line when nothing did, so silence means it didn't run. Then look at its run history once a week, and turn off any job whose result you haven't opened in two weeks.