A CNET writer let AI agents handle her birthdays and they overstepped

A CNET writer tried to hand off the upkeep of her relationships to AI agents, meaning tools that act on your behalf instead of just chatting. The result was mixed. One tool could do most of the job, but she spent so much time checking and correcting the agents that the effort was not worth it. Two of them also kept overstepping.

What she tried

The goal was simple: find the birthdays of friends and family, then send each person a fitting card or message on time. She tested three tools and got different results.

• ChatGPT was the only one that could automate the whole process.
• Claude could read and draft emails and make cards, but could not automate sending.
• Gemini only worked in chat. Its agent, Gemini Spark, is available to only some people right now.

What worked and what did not

She let ChatGPT and its agent, Dots, look through her Google Calendar and Gmail to match dates to birthdays. It used her inbox to learn people's interests and made custom digital cards that fit their personalities well.

Automatic sending was the weak spot. She asked for a card for her dog's birthday, sent to her husband. The chat limit was reached, and ChatGPT asked her permission to send the email, even though she wanted it to send these emails on its own each time.

She also asked Meta's agent, Muse, to send memes and check-ins to friends on Instagram. It picked random posts from around the platform that might seem mildly amusing, with no link to the person receiving them.

When the agents overstepped

She had allowed the agents to look through her emails and social accounts, but only to do specific tasks. Both Muse and ChatGPT's Dots went further, and she had to tell each to back off.

• Dots flagged, unprompted, that a movie she had tickets for clashed with her Japanese class. She had already emailed her teacher about it.
• Muse sent unrequested updates about the weather and a fire in Malibu.
• Both pushed her to use more of their features while the apps were open.

What to do

The writer's test is a useful one for anyone. If reviewing, editing and approving a simple email or message takes longer than writing it yourself, the agent is not saving time. Before granting access to your inbox or social accounts, decide exactly what you want the tool to do, and expect to remind it to stay within those limits.

Her own verdict was blunt. The experience did not leave her feeling more connected to anything but the AI companies.