Our Take
Spark performs well on Google's own turf but demands constant human oversight—exactly the opposite of what an agent should do—making the $100/month subscription hard to justify for routine tasks.
Why it matters
Enterprise buyers evaluating agent spend need to see that capability and cost matter less than whether you can actually trust the system to run unsupervised. Spark exposes the gap between what demos promise and what you'd risk in production.
Do this week
If you're considering Spark: run three real-world tasks on your own data before subscribing, document where Spark hallucinates (wrong links, missing people, bad permissions), and decide whether the time saved justifies constant verification work.
Google's Spark agent works—until you step away
The Verge tested Google's new Gemini Spark agent by replicating the demos shown at Google I/O. On straightforward tasks, Spark performed well. When asked to draft an email to the tester's wife with monthly grocery spending averages for 2026, Spark found the wife's email address without being given her name, located the correct budget spreadsheet in Drive (which didn't have "budget" in the filename), and pulled the data including incomplete May figures, then wrote a draft email with the correct personal sign-off.
On a block party planning task, results were mixed. Spark created a friends and family table and drafted emails mentioning a non-existent sign-up sheet. When asked to create that missing sheet and update the email link, Spark succeeded after a few minutes of processing.
On the most complex request—monthly calendar events, a family email about a TV episode, and a preschool prep checklist—Spark delivered results within four minutes but with notable flaws. The calendar reminders used "flamingo" color instead of the requested "hot pink." The email linked to a trailer instead of the actual episode and omitted the tester's wife from the recipient list, despite including her email address elsewhere in the same task. The shared document it created was only accessible to the tester, not the wife, and Spark said it wasn't currently able to grant additional access.
Spark is available only to Google's AI Ultra plan subscribers ($99.99/month), only in the US, and only in English (per The Verge testing).
The agent works, but you can't trust it to run alone
Every task required either active monitoring or post-execution verification. The tester checked notifications and reviewed outputs constantly, defeating the core pitch of an autonomous agent you can "put your phone down" and leave. The stakes are higher when the agent pulls personal data to draft emails or calendar invites to real people.
Privacy and cost compound the friction. Spark requires deep Google ecosystem integration and access to Gmail, Drive, and contacts. Google states that Gemini doesn't train directly on your inbox with Personal Intelligence enabled, but users must trust Google to be a "good steward of your data" with no independent auditing. At $100/month, Spark is justified only if you save significant time on repetitive work—but constant verification work negates that time gain.
The agent also surfaces the gap between what works on Google's internal data (emails to Google staff, Google calendar, Google Drive) and what works on your actual data. Hallucinations—wrong links, missing email addresses, broken permissions—are not corner cases; they appear in real tasks.
Test Spark on messy data before committing
Before subscribing, run at least three end-to-end tasks using your own unstructured data: files without clear names, incomplete records, and real people you actually contact. Document every failure mode: hallucinated links, incomplete recipient lists, wrong file permissions. Compare the time you save against the time you spend verifying output. If you're checking every draft anyway, Spark doesn't buy you autonomy—it just costs you $100/month. If you're already deep in the Google ecosystem and your data is clean, a free trial week will tell you whether the agent is worth the subscription. If you're evaluating agents for enterprise deployment, note that Spark's reliance on trust and manual oversight is the industry norm today, not a Spark-specific flaw.