What Can AI Agents Actually Do? 25 Practical Examples
Not demos, not predictions — everyday jobs that today's AI agents can realistically take on, with the prompts to try and the catches to watch for.
What can AI agents do? The short answer
Today’s AI agents are best at research and admin that spans many websites or documents: comparing products, collecting travel options, summarising and sorting email, pulling deadlines into a list, filling routine forms, and watching pages for price or availability changes. They are weaker at anything requiring judgement, logins behind strict security, or spending money without a human check.
Most lists of “things AI agents can do” either describe a distant future or quietly describe what a chatbot can do. This one tries to stay honest: every example below is something current consumer agents (ChatGPT Work, Gemini Spark, Claude’s task mode, Perplexity Computer, Microsoft’s Browse with Copilot and similar tools) are designed to handle, as of September 2026.
Each example includes a sample request you can adapt and a note on what to watch. A small legend helps:
| Rating | Meaning |
|---|---|
| Low risk | Research or drafting; nothing leaves your control |
| Medium risk | Touches your accounts or sends things, but easy to review or undo |
| High risk | Spends money, deletes data or contacts other people, keep a human checkpoint |
If you are new to the idea, read what AI agents are first; it explains why agents behave differently from chatbots.
Shopping and money
1. Build a comparison table for a purchase. “Compare the five most-recommended cordless vacuums under $300: battery life, weight, bin size, warranty and current price at two major retailers. Put it in a table with links.” Agents are good at this because it’s tedious, multi-site and easy to check. Low risk. Watch for: prices that were correct when the agent looked but have since changed, and review sites that are really affiliate lists.
2. Check return and warranty terms before you buy. “For these three products, find each retailer’s return window, who pays return shipping, and the manufacturer warranty length.” The kind of detail people skip and regret. Low risk.
3. Watch a price. “Tell me if this jacket drops below $120 in size M.” Google’s price-tracking and “buy for me” features and Amazon’s Alexa for Shopping both offer versions of this in the US, and general-purpose agents can run it as a scheduled task. Low risk to watch; high risk if you let it buy automatically. More in can AI agents find better prices online?
4. Audit your subscriptions. “Go through my email receipts from the last three months and list every recurring subscription, the amount and the renewal date.” Needs inbox access. Medium risk. Don’t let it cancel anything without showing you the list first.
5. Gather quotes. “Find three local providers for gutter cleaning, their published prices if listed, and their review ratings.” Agents can collect; they usually can’t negotiate or phone. Low risk.
Travel
6. Shortlist accommodation against specific rules. “Hotels within a 10-minute walk of the venue, under $200 a night including taxes, free cancellation, rated 8+.” Agents apply several filters faster than you can, and they can open each listing to verify the filter was honest. Low risk.
7. Build a draft itinerary. “Plan three relaxed days in Kyoto for two people who like food and gardens, with no more than two sights a day, and note which need advance booking.” Treat the output as a first draft. Low risk. Watch for: opening hours and closures that are out of date.
8. Compare routes and total journey times. “What are the realistic options from Manchester to Edinburgh on Friday afternoon by train, coach and plane, including getting to and from the airport?” Low risk.
9. Collect entry requirements and paperwork. “What documents does an Indian passport holder need for a week in Japan as a tourist? Link to the official source.” Useful for orientation, but always confirm with the official government or embassy site. Low risk if you verify.
10. Book a hotel. Google began rolling out hotel booking inside AI Mode in the US in August 2026, with partners such as Booking.com, Expedia, Marriott and Hilton handling the reservation. High risk only in the sense that money moves, read the rate conditions before confirming. See AI travel agents: can AI really plan an entire trip?
Email and communication
11. Triage a messy inbox. “Summarise unread emails from the last week into: needs a reply, needs action, FYI, and can archive.” Medium risk, let it label, not delete, until you trust it.
12. Draft replies in your voice. “Draft polite replies to these three meeting requests; decline the Thursday one and suggest next week.” Leave the drafts for you to send. Medium risk.
13. Unsubscribe in bulk. “List newsletters I haven’t opened in 90 days and unsubscribe from the ones I tick.” Medium risk. Unsubscribe links in spam can confirm your address is live, so only act on senders you recognise.
14. React to specific emails automatically. OpenAI’s documentation describes event-triggered tasks in ChatGPT Work that can run when a new Gmail message arrives. “When an invoice arrives, add the amount and due date to my bills spreadsheet.” Medium risk. Be wary: an agent that reacts to incoming email is also reacting to emails written by strangers. Our email agents guide goes further.
Calendar and planning
15. Pull deadlines out of scattered messages. “Find every date or deadline mentioned in emails from my kids’ school this term and add them to my calendar as all-day events.” Medium risk, check the dates it extracted.
16. Find a time that suits everyone. “Suggest three 45-minute slots next week when I’m free before 4pm, avoiding Wednesday.” Low risk.
17. Prepare a weekly brief. “Every Monday at 8am, give me my meetings, anything overdue, and emails I promised to answer.” A good first scheduled task. Low risk if read-only.
Research and documents
18. Research a question across many sources. “What do the main consumer groups in the UK and US say about extended warranties on electronics? Summarise with links.” This is the strongest current use case. Low risk. Always click through to check the sources exist and say what the agent claims.
19. Turn documents into something usable. “Read these five PDFs of rental agreements and make a table of rent, deposit, notice period and pet rules.” Low risk, though keep sensitive documents out of tools whose data policies you haven’t read.
20. Monitor a page for changes. “Check this visa appointment page every morning and tell me when new dates appear.” Low risk. Some sites block automated checking; respect that.
21. Organise files. Desktop agents that can access folders you grant, OpenAI’s desktop app and Anthropic’s Claude desktop app both offer this with permission, can rename and sort downloads, invoices or photos. Medium risk. Make a backup first and never grant access to your whole drive.
Home, study and life admin
22. Fill in routine forms. “Fill in this change-of-address form with the details from my notes, but stop before submitting.” The “stop before submitting” part matters. Medium risk.
23. Plan meals and a shopping list. “Plan five weeknight dinners under 40 minutes, vegetarian, and produce a combined grocery list by aisle.” Low risk.
24. Build a study plan (not do the homework). “Here’s my exam syllabus and the dates. Build a revision timetable and a set of practice questions for each topic.” Good for learning; see AI agents for students for where the line sits.
25. Compare utility or phone plans. “Given my usage of about 20GB a month, compare current SIM-only plans from the four biggest networks here.” Low risk to compare; switching is a decision to make yourself.
What these examples have in common
Look back over the list and a pattern appears. The tasks that suit agents best are:
- Tedious rather than difficult. Many steps, each simple.
- Spread across places. Several websites, many emails, a pile of documents.
- Easy to check. You can glance at a table and spot a problem.
- Reversible. A label can be removed; a draft can be edited.
The tasks that don’t suit them yet are the opposite: a single high-stakes decision, a subtle judgement call, anything that spends significant money or speaks for you to someone important.
Before you hand over any task
A short checklist that prevents most problems:
- Give the minimum access. Connect only the apps this task needs.
- Ask for a plan first if the product supports it, and read it.
- Add a stop point: “show me before sending/buying/deleting”.
- Spot-check the result. Click two or three of the sources or links.
- Remove access you no longer need.
For the full reasoning behind those habits, read are AI agents safe to use?.
Key takeaways
- Agents shine at tedious, multi-site research and admin that’s easy to check.
- Start with read-only or draft-only tasks; move to actions once you trust the results.
- Always keep a human checkpoint before money moves, messages go out or data is deleted.
- Verify prices, dates and official requirements at the source, agents can be out of date.
What can AI agents do? Quick answers
Very little should be fully unsupervised. Read-only jobs such as research, summaries and price watching are low risk. Anything that spends money, sends a message or deletes data should stop and wait for your approval.
Yes, many can fill routine forms such as bookings, sign-ups and contact requests. Check every field before submitting, and never let an agent handle passwords, card numbers or one-time codes directly.
Judgement calls, sites with strict login security and anything where an outdated price or rule causes real harm. Verify prices, dates and official requirements at the source before you rely on them.
Sources
- OpenAI Help Center, “ChatGPT Work and Codex” (scheduled and event-triggered tasks; local files in the desktop app)
- Google, Gemini Apps Help, “Use Gemini Spark”
- Google, “Shop with AI Mode” (price tracking and agentic checkout)
- TechCrunch, “Google’s AI Mode can now track flight prices, help book hotels, and more” (27 August 2026)
- Anthropic, Claude Help Center
- Microsoft Support, “Browse with Copilot”
