Our Take
Spark delivers the first genuinely useful agentic experience by mining data Google already holds; the feature works precisely because privacy is the price of personalization.
Why it matters
This is what agentic AI looks like when it has full account access: capable enough to feel magical, invasive enough to unsettle. Every AI vendor is racing toward this model because it works.
Do this week
Security teams: audit what Google Workspace permissions your Spark-enabled users have granted, and document the data exposure baseline before wider rollout.
Spark Assembled a Weekend Itinerary by Reading Your Life
Google's Spark agent, rolling out on the $99/month Gemini AI Ultra tier, was tested on a trip-planning task. The user asked for a weekend itinerary in Hershey, Pennsylvania for a family of four plus a dog, intentionally withholding specific details like concert tickets and the dog's name.
Spark returned a multi-thousand-word document in a shared Google Doc that included: hotel options with pet fees, driving directions from the user's home address, dog-friendly activities, and the dog's actual name (Frida). It inferred which child (Lewis, age one) would enter Hershey Park free, noted that the three-year-old (Arthur) needed a ticket, and scheduled a nap at 1:30 PM. It listed the wife's name (Anna), noted her aversion to onions and scallions, and pulled the Thomas Rhett and Niall Horan concert details from a Ticketmaster confirmation email. When asked to share the itinerary with Anna, Spark retrieved her email address and sent the document with a drafted note.
The agent failed only when asked to book an Airbnb directly; Airbnb's authentication blocked the action. Spark offered alternative listings instead.
Data Access Is the Moat
Trip planning has been the canonical AI demo for four years. Most chatbots fail at it because they lack context beyond public information. Spark succeeds because it has unrestricted access to Gmail, Google Calendar, Google Photos, search history, and Google's Personal Intelligence feature, which mines this data into a profile.
This is the explicit trade being pitched: usefulness scales with exposure. OpenAI and Anthropic are racing to accumulate similar data access because they lack Google's structural advantage. Spark demonstrates the outcome: an agent that operates like a human assistant because it knows your family names, addresses, dietary preferences, and financial transactions.
The friction point is psychological, not functional. Users intellectually accept that Google holds this data. Seeing it mined and acted upon in real-time, without explicit per-action consent, triggers a different reaction. As the source notes, the system treats personal data not as something to protect but as something to mine for benefit.
Expect This to Become the Baseline
Spark works because it has permission to read everything. Future agentic interfaces will be shaped by the same constraint: access depth determines capability. Organizations deploying agentic AI internally must decide whether to grant similar permissions and document the decision.
The Airbnb failure also signals a real limitation: sites with strong authentication and anti-bot policies will remain out of reach. Agents today can read your data but cannot easily complete transactions on protected platforms. This boundary will tighten or blur depending on how vendors and platforms negotiate integration terms.