Gemini’s AI Agent Matches Google’s Demo Performance

Google’s new AI agent, Gemini Spark, is impressively capable of handling tasks for you, but the financial investment and potential privacy concerns make me hesitant about its value.

I got my hands on Spark last week, and it’s marketed as an AI assistant that can manage tasks in the background, even those with multiple steps, letting you step away from your devices. The Spark website emphasizes that it operates under your control, only activates when you want it to, and checks in with you before making significant decisions. Given the growing skepticism around AI, it feels like wearing a “not involved in rogue AI” shirt while people still have questions.

Unsure where to begin, I decided to follow my colleague Antonio’s lead and test Spark with tasks similar to those showcased at a recent event. Would it perform as effectively in my home office as it did on stage?

During the event, a Google VP demonstrated Spark by asking it to draft an email to a team, summarizing recent launches and tailoring the tone to match his style. Since that seemed straightforward, I pushed the boundaries a bit further. I asked Spark to draft an email to my wife detailing our average monthly grocery spending for 2026. This would test whether Spark could identify my wife without me naming her, locate our budget spreadsheet in Drive (which doesn’t have “budget” in the title), and create an email in Gmail.

When Spark delivered the results, I was genuinely impressed. It found my wife’s email, extracted the correct data from our 2026 budget spreadsheet, calculated the monthly grocery totals (including incomplete data from May), and composed a draft email in Gmail. The email addressed her by her first name, even though her email address doesn’t include it, and it featured a personal sign-off we use.

In another example, the VP asked Spark for help organizing a block party. While I’m not planning a block party, I decided to ask Spark the same questions. The results were less impressive. It generated a table of friends and family for food contributions, drafted an email referencing a non-existent sign-up sheet, and created a poorly designed presentation about city permits.

To challenge Spark further, I requested it create the missing sign-up sheet and link it to the drafted email. After a few minutes, it successfully generated a spreadsheet and updated the email with the link.

The VP’s final demo was the most striking. He asked Spark to do several tasks: color-code meetings with the CEO, invite a new neighbor to the block party, and create a to-do document for his kids. I adapted this by asking Spark to set up monthly calendar events leading up to my wife’s birthday, draft an email to my family about the latest season of a show, and compile a document with preschool preparation tips for our toddler.

I initiated this request at 3:35 PM PT on a Friday. During the event, the VP made a point of putting his phone down and checking back later, which I did too. After addressing a minor issue—Spark wanted access to my contacts, which I declined—my tasks were completed in about four minutes.

The results were impressive, albeit with some flaws:
* My Google Calendar now has events set for 9–10 AM on the correct days leading up to my wife’s birthday, though the reminders are in a shade Google calls “flamingo,” which is close to hot pink.
* Spark drafted an email with my immediate family’s addresses, but oddly, it left out my wife’s. The email correctly named the first episode of the latest season but linked to a trailer instead of the episode itself. It also included “loool,” a casual expression I often use.
* Spark created a Google Doc with a preschool checklist, but it’s only accessible to me. When I asked if it could share it with my wife, Spark indicated it couldn’t do that yet.

While Spark shows potential as a powerful tool, there are important caveats. Like any AI, you need to verify its output for accuracy, especially when it involves personal information. Although Google promotes Spark as an autonomous assistant, I found myself constantly monitoring its actions and notifications. What’s the point of having an assistant if you can’t trust it to operate independently? And why should I rely on something I’m unsure about when it consumes resources for relatively minor tasks?

Currently, Spark is only available to subscribers of a premium plan, which starts at $99.99 per month, and is limited to users in the US and English speakers. I received free access for testing, but I don’t think it justifies the high subscription cost, especially since I could accomplish all the tasks myself, albeit with more time investment.

Spark seems to work best for those already embedded in the Google ecosystem with Personal Intelligence enabled. Having a Google account for nearly two decades means Spark has a wealth of data to draw from. However, while Google assures that Gemini doesn’t directly train on your Gmail with Personal Intelligence on, you still have to trust that they will handle your data responsibly. For now, I’m uncertain if the cost and risk are worth it.