WIRED Tested Instinct: the AI Agent That Saved $550 and Lost $64

Zoë Schiffer, a WIRED contributing editor, spent several weeks texting Instinct, an invite-only AI agent that communicates through iMessage and WhatsApp and connects to users' email, calendar and messaging apps. In an account published on September 24, she describes an assistant that arranged restaurant reservations and recovered a $550 airline refund — while also costing her $64 and raising questions about how much access an agent should hold.

How Instinct works

Launched in private beta in February, Instinct has no open text box: users text it as they would a person, and the agent itself suggests tasks and then carries them out. Schiffer writes that it also avoids texting too often, unlike the competing agent Lindy, which she deleted after a week because it texted her every morning and joined Zoom meetings uninvited.

Where it earned its keep

On a work trip to Venice, Italy, Schiffer handed over her itinerary and asked Instinct to book tables wherever she would be, including by messaging hole-in-the-wall restaurants on WhatsApp. The reservations came through.

A harder task followed. She had already booked a November trip to New York with her sister when WIRED asked her to attend the WIRED World Fair in Miami the day before. Her saver fare could not be moved, and missing the first leg would cancel the whole booking. The agent noticed that Alaska Airlines had moved her flight 90 minutes earlier — a change she had missed — which qualified her for a full refund. It cancelled the ticket, saving her roughly $550, and rebooked her one-way return to San Francisco.

The security trade-offs

That same access is where the risk comes from. Users who tried to disconnect the agent from their email found it had kept a copy of their inboxes anyway. One venture capitalist said Resy banned him after the bot hit the reservation service's API roughly 200 times an hour. Another tech investor deleted the app after concluding that phishing Instinct would be trivial. The company's terms of service also allow it to use at least some conversations to train its AI models.

Schiffer's own mishap was smaller: told to cancel a delayed DoorDash order only if a refund was possible, the agent cancelled it and she lost $64. It apologised, she writes, without offering to cover the loss. The company did not respond to WIRED's request for comment.

Why it matters

The mixed record lands in a widening split between people who delegate work to AI agents and those who have never tried one. Schiffer notes that agents still require a lot of direction from users. She is keeping the app anyway: it recently warned her not to open an email from a friend that turned out to be a phishing scam. Her conclusion: agents are a security nightmare, but for now she is willing to take the risk.