Business Insider testers lost 8 pounds and a Notion schedule when an AI agent acted for them

The board-deck summary of these trials is that personal agents are good at finding things and bad at doing them. The record supports half of that. Over two weeks, Instinct booked Will Martin's commute tickets into London and got him a refund on an $80 subscription he had forgotten to cancel ^[4]. Both tasks involved money. The failures were different in a narrower way. In each case the agent made a judgement call on its own, and the result took effect before anyone checked it ^[7][10].

In Oxford, the agent did careful work on the wrong question. Instinct checked Martin's car model against the charging zone's emissions tiers and confirmed it had paid for two days at 4 pounds a day ^[6]. The mistake came earlier, when it chose which charge to pay: it paid a similar but separate one covering a smaller area of the city ^[7]. His partner caught the error, and the payment was nonrefundable ^[7]. At the 5-pound rate she cited, two days in the right zone would have cost 10 pounds ^[9]. "Lesson learned, even if Instinct tried to find a silver lining," Martin wrote ^[8].

Cheryl Teh's loss came from giving the agent wider access. In her first 24 hours with Instinct she connected it to her Notion dashboard and asked for an overhaul. It cleared her goals calendar and schedule and left a blank template in their place ^[10]. The lesson she took was about instructions: "for optimal results, you've got to be precise in your prompts, and lay out exactly and unambiguously what you want the agent to do," she wrote ^[11].

A skeptic would call both episodes user error that sharper prompts would fix. A sharper prompt changes what the agent tries to do. It does not change what the agent can reach, and a one-word request reached an entire dashboard ^[10]. I think the protection that counts is a permission: a confirmation step before any nonrefundable payment, and read-only access to a workspace until the user has seen the plan.

Teh's own setup already works that way. Instinct found a link to signed early-order albums in three minutes. She said it saved her two hours and caught a US-only window she would otherwise have missed ^[12]. The agent asked her twice whether she had set a budget, and she had not ^[13]. "I haven't plucked up the courage to give the agent my credit card details yet, though," she wrote ^[14].

The evidence on Meta's agent is thinner. Editor-in-chief Jamie Heller's test was a research task, working out what an opaquely described card charge was. "Chat couldn't figure it out. Muse did," she wrote ^[2].

Five staffers over roughly a week add up to a set of anecdotes ^[1]. We do not know yet how often either agent picks the wrong zone or the wrong template. Whatever a team grants this quarter, whether a card on file or write access to a shared workspace, sets the size of next quarter's mistakes. Martin's wrong guess cost 8 pounds only because a congestion charge is a small bill ^[7].