The Information

Marissa Mayer's Dazzle turns your camera roll into a personal assistant

Key points

Key takeaways from The Information interview with CEO & Founder Marissa Mayer on her AI startup Dazzle (September 2026):

Camera-roll context. Dazzle's personal assistant builds its context layer from the user's camera roll rather than email or calendars, learning who matters to you, what you like and how you spend your time — Mayer argues photos are the richest and most accessible source of personal information.

Proactive, photo-driven actions. Snap a picture of something to buy, fix or schedule and Dazzle works out what to do and brings you that action; a "get it done" list handles tasks on your behalf with your permission — the average user holds around 25,000 photos, and one has over a million.

Humans in the loop. Consumer experience specialists step in when agents get stuck or a task is more efficient with a person, so complicated jobs still run all the way to completion.

Agent credentials. AI agents and specialists hold their own accounts — they can buy and transfer tickets without transacting on your credit card, and Dazzle asks for confirmation before any transaction; Mayer thinks handing credentials to AI is coming, but not quite yet.

Privacy by discard. Users choose full, selective or no camera-roll access; under "when in doubt, leave it out", photos that look intimate or carry personally identifiable information are dropped and deleted immediately even on full access.

Lessons from Sunshine. Dazzle's photo-sharing predecessor Sunshine built an AI "shareworthy" model, which showed the real need was turning photos into actions — like resurfacing the football cards her son photographed as a ready-made gift idea — rather than sharing them.

Model-agnostic and monetisation. Dazzle runs primarily on LLMs today but uses a model-agnostic harness, experimenting with single-purpose and decision-making models; likely monetisation is a power-user subscription plus transaction fees on bookings, driven by personalised suggestions from past behaviour.

No partnerships needed. Instead of thousands of service-provider deals, Dazzle uses agentic browsing, virtual machines and human agents to source from local or favourite vendors — from a frayed garage door to roofing estimates.

Read more: The Information

More top stories