Instinct Wants to Run Your Errands. Can It Earn Your Trust?

Jason Yeh asked Instinct to find open dinner reservations. According to his account reported by Forbes, the AI assistant booked a table carrying a $200 cancellation fee.

The problem was the distance between two seemingly similar instructions: find me an option, and commit me to it. Permission to investigate is different from permission to commit. For an AI assistant capable of acting on your behalf, understanding that distinction is part of the job.

It is also what makes Instinct such an interesting product. You can text or call it, connect the accounts it needs and delegate everyday chores. The company promises an assistant that follows up on neglected conversations, arranges rides and books services, with enough personal context to anticipate what needs doing.

The appeal is easy to understand. Getting a useful answer is satisfying. Getting a task off your list can change your day.

Behind the product is Noah Shinn, whom The Wall Street Journal described as a 23-year-old who left Northeastern University in 2023. A former Sierra research scientist, he had also worked on machine learning and programming-language research at Northeastern and MIT.

Investors have moved quickly: Forbes reported a valuation increase from $500 million to $2.5 billion in August; a September 17 New York Post opinion column described subsequent fundraising talks at $10 billion. The fivefold and fourfold headlines refer to successive jumps, with the latter still a proposed deal.

For users, the attraction starts with a familiar message thread. Instinct works through channels including iMessage and WhatsApp, while drawing on connected email, calendars and other information. Its pitch is that you can hand over the errand without learning another elaborate application.

That pitch is becoming more tangible. Stripe's Link integration supports purchases, while Shinn has announced an email address for the assistant itself and a trusted network that lets users' assistants coordinate plans. Its new Instinct Concierge feature is rolling out to selected users for calls, including restaurant bookings and cable-bill inquiries.

Together, those tools move the product closer to the ordinary work of a human assistant: finding information, communicating with someone else and completing the transaction. The promise is continuity. You should not have to start from scratch every time an errand moves from an email to a website to a phone call.

The strongest praise centers on precisely that kind of practical help. Investor Sarah Guo described successful restaurant bookings, marathon entries, campsite reservations and purchases. Her firm, Conviction, is an early backer.

Betaworks' John Borthwick, whom the Post identified as having no stake in Instinct, said he would recommend it to nontechnical family members. The same column described business owner Matthew Ammirati using it to help manage more than 500 incoming emails a day.

These are recognizable problems. Dinner plans, an overflowing inbox and household shopping compete for attention with everything else a person is trying to accomplish. They also recur, giving an assistant repeated opportunities to become useful.

There is some evidence that users are moving beyond experimentation. In an August 28 announcement, Shinn said purchasing users spent more than $1,300 a month on average through Instinct, citing groceries and household essentials among the uses.

That is meaningful spending, but the denominator matters. The average covers people who buy through Instinct. It does not tell us what proportion of all users make purchases, and a few expensive transactions could pull it upward.

Shinn also said 32% of users shared credentials through its Vault, in his September 4 announcement of a planned 1Password integration. That measures willingness to enable account access. It does not establish how many people regularly delegate shopping, travel or email.

The figures suggest early engagement. Public reporting still lacks the task-by-task adoption and retention data needed to establish which habits are sticking.

Instinct also faces capable competition. ChatGPT Work can act across connected tools, files and desktop apps, including recurring assignments. Claude Cowork handles work across files and connected services, with research, document creation and scheduled reports among its uses.

Instinct's emphasis is personal coordination through a conversation: the ride, the reservation, the forgotten follow-up. Its opportunity lies in making that experience convenient enough to become a habit. Readers already paying for a broader assistant should consider how much of their actual to-do list it can already handle.

Access and cost remain less settled. Forbes described Instinct as free in August. The public website reviewed for this article offers a waitlist and invitations, without a published subscription price. An invitation is a chance to evaluate an evolving service, with its eventual cost still an open question.

The harder question is how much access to grant it.

Yeh's reservation complaint is one example. In another, Claire Vo reported that Instinct retained email copies after she disconnected Google access. Forbes said the company did not respond to its comment request. TechCrunch also documented an unauthorized email and a user's account of tricking Instinct with a malicious message.

On the broader security concerns, Shinn told The Wall Street Journal that the team was taking them seriously and had introduced product changes to prevent certain unwanted actions.

There is a wider security problem here. Developer and AI security commentator Simon Willison calls the combination of private data, untrusted content and the ability to communicate externally the "lethal trifecta". An attacker can place instructions inside an email or webpage that an assistant reads, trying to persuade it to disclose information or act against its owner's wishes.

Willison's warning concerns AI agents generally. It explains why connecting an inbox is such a consequential step: the assistant gains access to your information and to material supplied by other people. It needs to know whose instructions count.

Instinct's published policies now provide several specific controls. Its privacy policy excludes Google Workspace API data from model training, and its terms separately exclude Vault material. Other information may be used for training, subject to an opt-out with stated exceptions. Those distinctions deserve more attention than a blanket promise that data are safe.

That gives the decision to try Instinct a practical starting point: choose a chore, define what completion means, and be explicit about which commitments require your approval. Then judge the result, including how much checking and correction it required.

The attraction is real. So is the difference between an assistant that finds a table and one that decides you are going.

An assistant earns its place when the time spent supervising it is consistently smaller than the time it gives back.