AI tools are moving from answering to acting on your behalf

Taken together, this week's updates show a clear direction. AI assistants are not just meant to answer questions in a chat window. They are meant to carry out tasks, remember who you are and have access to more of your information.

From chat to action

Amazon has opened up shopping by text message with Alexa for Shopping. For now, it is only available to invited customers in the US. Airbnb is planning an AI agent for travel in the first half of 2027. An agent is an AI program that starts working on a task on its own instead of simply giving an answer. Meanwhile, Airbnb's assistant already resolves roughly 40 percent of support cases without a human employee.

Perplexity is heading the same way with Teammates, persistent AI agents that take over ongoing work tasks. They are reserved for those who pay for the expensive Max subscription.

The most personal version comes from Anthropic. With Personal Finance in beta, Claude can read bank accounts and credit cards without being able to move money. For now, it is only available to paying users in the US. In other words, it is a feature that watches but does not act for you.

Memory is the new battleground

For an assistant to be truly helpful, it has to remember. Meta has updated the models behind Muse and Meta AI so that they remember information from earlier conversations and stick with a task even when something goes wrong. The figures, however, come from Meta's own tests. Living Brain from ChatAndBuild tries to solve the problem across tools, so that ChatGPT, Claude, Cursor and Codex can share the same knowledge about your project.

That is convenient, but it also means more of your knowledge is gathered in one place.

Take the numbers with a grain of salt

Many of this week's figures come from the developers themselves. OrcaRouter reports that an uncensored version of the coding model Flint rejects harmful requests in 10.4 percent of cases, compared with 86.3 percent before. Those are the publisher's own numbers, and they are also a reminder that open models can be modified so that their safety guardrails are removed.

Comparisons between assistants should also be read with caution. Android Police writer Parth Shah names Gemini the clear winner over Claude and Grok after a month on Android. That is one person's experience, however, not an objective measurement.

What can you do?

Try new features on tasks where a mistake won't be costly. Check whether a feature is even available in Denmark before you plan anything around it. Several of this week's announcements apply only to the US, to invited users or to those with expensive subscriptions. And pay extra attention to who gets access to your data when an assistant gets better memory or looks into your finances.