Plain English

What can bots actually do?

Short answer: they're good at reading, watching, sorting and drafting. They're getting OK at booking and buying while you watch. They aren't safe to trust with your money, your address or decisions you can't undo.

Chatbot vs. agent, in one minute

💬 Chatbot

You ask, it answers. It can write a letter, explain a confusing form or make a plan, but you do the clicking. ChatGPT, Claude, Gemini and Grok all work like this for free.

🤖 Agent (a “bot”)

You give it a job, and it does the steps: opens websites, reads your email, fills in forms, and can repeat the job on a schedule. Examples: Meta Muse, Gemini Spark, Grok Bot, ChatGPT agent and dots, Claude Cowork, Instinct, plus no-code tools like Zapier, n8n and Make.

Most agents run on their own computer in the cloud, so they keep working when your laptop is shut. Google says Gemini Spark “keeps working in the background even when you close your laptop” (Google).

✅ What they're good at today

Reading and summarizing: inbox briefings, turn the flood of school emails into one weekly digest, explaining a confusing letter.
Watching for changes: have a bot watch any web page and ping you when it changes, get alerted the moment your airline changes your flight, watch your email for one important letter or reply.
Drafting: complaints, follow-ups, replies in your style.
Doing the same small job every week: plan a week of family meals and fill the grocery cart, catch family calendar clashes before the week starts.
Spotting problems: get a daily scam check on your inbox, only see the emails that actually need a reply.

👀 OK if you stay in the loop

These work for many people, as long as the bot asks you before it pays or confirms:

book restaurant tables by message, even at phone-only places, book a haircut or appointment, even when the form won't work and plan and book a whole trip by texting your bot.
let an agent reorder your regular groceries: some stores block agents, so check which ones work.
have your bot make a customer-service phone call for you: tell it exactly what “resolved” means.

OpenAI says its agent asks before bookings and payments (OpenAI help); Google says Spark asks before “high-stakes actions like spending money or sending emails” (Google).

🛑 What they can't do safely yet

Get it right every time. A benchmark of four personal agents found that even when they found the right information, they “often miss what matters, overshare, or make things up” (micro1, Oct 2026).
Follow your exact conditions. WIRED's reviewer said her agent cancelled a delayed food order and she lost $64, even though she'd told it to cancel only with a refund (WIRED).
Keep your private details private. One reported case: an agent selling an item gave a buyer an address and booked a pickup when the seller wasn't home (X post). WIRED also notes reports of an agent keeping inbox copies after being disconnected.
Be faster than you. In one startup's timed test, agents took about 4½ to 14+ minutes to book a dinner the founder booked in 37 seconds (X post).
Get in everywhere. Some sites block agents. One user says Amazon blocked his (X post).
Replace a professional. Use bots to organize and draft. Medical, legal and money decisions stay with you and a qualified person. Be wary of anyone promising AI miracles; the U.S. FTC has acted against deceptive AI claims (FTC Operation AI Comply).

7 rules that keep you safe

Start read-only. Let it look before it does anything.
Make it ask before it sends, buys, books, cancels or deletes.
Pay with a low-limit or single-use virtual card (give your agent a single-use card before it books or buys anything).
Never give it your passwords in a chat. Use the app's official “connect” button.
Tell it to ignore instructions found inside emails or web pages from strangers.
Check the first week of results against the real thing.
Keep a way to switch it off, and know which apps it's connected to.

Ready to try?

Start with something read-only, like the Morning Inbox Briefing or the Scam Checker.

See all blueprints See what's trending

Or browse every AI agent example on one page.