AI agents now answer, book and buy for you. Here's what can go wrong

• What an AI agent is
• The agents you can use now
• Six jobs agents already do well
• What can go wrong
• Six rules before an agent acts for you
• How to try one safely
• How we set up an agent

A chatbot answers questions. An agent works in your calendar, inbox, files or booking system, step by step, toward a goal you set.

At 21:30 a customer calls a hairdresser to book a cut. Nobody answers, so they book with the next salon on the list. An agent can take that call, book the slot and send a confirmation by text message.

OpenAI says Dots works toward your goals "in the background with minimal oversight". That’s why the rules further down matter. Our comparison of ChatGPT, Claude and Gemini shows how the models behind them differ.

Sources: TechCrunch and Tek.no on Dots, Google, TechCrunch and Reworked on Gemini, Anthropic on Cowork (read 13 October 2026).

If you run a business, these are ready for an agent today:

Say you’re a plumber. On Sunday evening a customer asks about a new water heater. The agent asks for photos and the address and drafts a quote from your price list, which you check and send on Monday. If water is leaking right now, it calls you.

In July 2026, by OpenAI’s own account, agents it was testing for hacking skills escaped from a test that was meant to run offline and reached the servers of the AI company Hugging Face. OpenAI’s monitoring would have caught it more than a day earlier, VG reported, but it was switched off. Nobody independent has checked OpenAI’s account, Lotte Skolem of the think tank Langsikt told NRK.

An agent that answers your phone can reach far less: your calendar, your opening hours, your price list. Its risks are more ordinary:

• It gives a wrong answer and sounds sure about it.
• It shares information with the wrong person.
• It does more than you meant, like sending an email or buying something.
• Nobody notices, because nobody reads what it does.

Sources: NRK, 22 July 2026, VG, 27 August 2026.

1. The agent drafts, a person approves. Payments, refunds, quotes and anything that makes a promise go to a person first.
2. Only the access it needs. An agent that books appointments gets your calendar, not your bank or your whole inbox.
3. A record of everything, with alerts on, so you can see every answer and every action afterwards.
4. Data stored in Europe. Choose a service that stores data in the EU or EEA, and sign a data processing agreement: the contract GDPR, the EU privacy law, requires when another company handles personal data for you.
5. Say it’s an AI. Article 50 of the EU AI Act requires it from 2 August 2026, unless it’s obvious.
6. A quick way to a person. Anyone can ask for a human, and the agent hands over the whole conversation.

Sources: EU AI Act Article 50 and Article 113 (dates of application).

1. Count what you miss. In last week’s call log and inbox, count the calls and messages that came after hours, and how many never got an answer.
2. Pick one job, the one that costs you the most time or customers, and write down the 20 questions people ask most, with your answers.
3. Set the limits: what the agent may do alone, and the one thing it must never do, such as give a price. Connect only the tools that job needs.
4. Test it for two weeks while a person reads every answer. Then keep it, change the rules or stop.

When we set up an agent, we start the same way: one job, on real data, for one to two weeks, under GDPR and with the data kept in the EU or EEA where possible. Then there’s a clear yes or no, and a fixed price for the full version.