What happens when an AI gets full access to your email, documents, and calendar - and still can't tell you're dating someone?
Wired's hands-on review of Google's new Gemini Spark - an AI agent that plugs directly into your Google Workspace data - put this to the test. The task: plan a birthday party using the agent. Gemini Spark had full access to the tester's Gmail, Google Docs, and Google Calendar. It handled logistics. It found dates. It processed the information presented to it.
It did not recognize the tester's boyfriend as the most important person in her life. Despite appearing throughout her communications and calendar, he was invisible to the agent in terms of social weighting. Wired described it as the agent effectively "friend-zoning" him.
Access Isn't the Same as Understanding
This is a real limitation in how current AI agents handle personal data. Reading an email isn't the same as understanding the social significance of who sent it. Frequency, tone, and relational pattern - the signals humans use instinctively to map their social world - aren't what an AI optimizes for when scanning your data for "relevant information."
Google's pitch for Gemini Spark is that your own data makes it more personal than a generic chatbot. That claim is directionally right. An agent with access to your inbox is better positioned than one without it. But there's a meaningful gap between "has access to your data" and "understands your relationships" - and this review lands squarely in that gap.
Where Agents Actually Perform Well Right Now
For routine productivity tasks - finding a document, summarizing a thread, scheduling around calendar conflicts - Gemini Spark likely performs well. Google Workspace integration is genuinely useful for that class of task. The problem surfaces when the agent is asked to do something requiring knowledge of what matters to you, not just what's in your files.
Microsoft's Copilot and Apple Intelligence are both building toward deeper personal data access and face the same challenge. The pattern holds across all of them: current AI agents are good at data retrieval and poor at relational reasoning. Gemini Spark appears capable on the former and not yet equipped for the latter.