CNET Test Shows AI Email Agents Require Heavy Manual Oversight

A recent test published by CNET demonstrates that current AI agents struggle to automate personal relationship maintenance without substantial human intervention. The experiment involved using ChatGPT, Google’s Gemini, and Anthropic’s Claude to manage birthday reminders and email correspondence, revealing significant limitations in fully autonomous operation.

How do these AI agents handle email automation?

The CNET report indicates that ChatGPT, specifically using its Dots plugin, was the only tool tested capable of automating the entire process from calendar identification to email sending. The author authorized the tool to read Google Calendar and Gmail data to match dates with contacts. However, the process required the user to review and approve every action, undermining the efficiency gains promised by agentic AI.

What limitations exist in other major AI models?

Other leading AI tools showed restricted capabilities during the same evaluation. Anthropic’s Claude could access inboxes and draft emails or create birthday cards, but it could not automate the actual sending of messages. Google’s Gemini functioned primarily as a drafting assistant within a chat interface, lacking the ability to operate directly inside the user’s inbox for most participants.

Why does this matter for operational risk?

The findings suggest that relying on AI agents for communication tasks may introduce more overhead than savings. According to the CNET article, if the time required to review and approve automated emails exceeds the time needed to write them manually, the value proposition of the technology collapses. This inefficiency highlights the gap between marketing claims of automation and the practical reality of current AI performance.

Questions to ask your suppliers

Organizations should inquire whether vendor-provided AI tools require continuous manual supervision for each action. Teams usually need to verify if the software logs every approval step for audit purposes. Check whether the supplier defines clear boundaries for automated communication, especially regarding sensitive personal data. Ensure that internal policies mandate human oversight before any external message is dispatched by an agent.

Background: AI agents

AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. They move beyond generating text to performing tasks like coding, scheduling, or data analysis. You must treat them as active participants in your network, not passive tools.

Read the full guide: AI Agents Examples: How Autonomous Software Changes Security

What to do and how to stay safe: ChatGPT

• Review permissions granted to AI plugins and revoke access to calendars and inboxes unless strictly necessary.
• Require manual approval for all outbound communications generated by automated agents.
• Monitor logs to ensure AI tools are only accessing data explicitly authorized by the user.
• Wait for vendor updates that provide clearer controls over autonomous sending capabilities.

Step-by-step guide: AI Usage Policies at Work: How to Prevent Data Leakage

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Which AI tool could fully automate sending birthday emails in the CNET test?

ChatGPT, using its Dots plugin, was the only tool that could automate the entire process, though it required user approval for actions.

Can Claude send emails automatically?

No, the report states that Claude could draft emails and create cards but could not automate the sending of messages.

How did Gemini perform in the automation test?

Gemini could only work within a chat interface to draft emails and generate cards, lacking direct inbox automation for most users.