Instinct Raised $1 Billion, Then Started Pushing Stuff Nobody Asked For

Days after confirming a $1 billion Series C at a $10 billion valuation, the AI agent startup Instinct rolled out a feature that turned its assistant into a shop counter. Users did not take it well.

The feature is called Instinct Selections. It pushes personalized recommendations in dining, travel, and shopping directly to users, without waiting for a request. Founder Noah Shinn announced it on X, framing it as an attempt to bring human taste to the platform by partnering with local chefs, interior designers, architects, and outdoor guides. A restaurant search could draw from lists picked by chefs who know the area. A trail suggestion could come from a regional guide. Home decor could come from independent designers, matched to a user's preferences and budget.

The concept is coherent. If you are going to build an agent that recommends things, curating inputs rather than scraping the web is a defensible way to differentiate from a generic chatbot.

What the rollout actually did

The idea was fine. The rollout was the problem.

Recommendations began appearing overnight, unprompted. Shruti Gandhi, a general partner at Array VC, said the assistant pushed suggestions for brimmed hats, sunglasses, and carry-on luggage she had not asked about. She described the shift as turning the assistant into a commerce engine rather than a helpful personal assistant, and asked the company to at least build a personalized shopping list of things people actually need.

Chat Joglekar, founder of a business-buying marketplace, described his first ewww moment with an AI assistant. Instinct suggested a portable coffee flask after reading a Blue Bottle coffee subscription in his email, and travel accessories tied to upcoming trips pulled from the same inbox. Entrepreneur Andrew Yeung and investor Ash Thaker reported similar unsolicited recommendations. Thaker's suggested fix was straightforward: let users opt in and choose their own preference sources.

That last point is the crux. The complaints are about timing: recommendations arrived outside the context of an active request. Selections interrupted users rather than responding to them.

The timing is the awkward part

Shinn had laid out a clear philosophy a day before the launch. On an episode of the Invest Like the Best podcast, he said Instinct should not influence a user's behavior in a way that is not aligned with what the user wants. He went further, criticizing Google, TikTok, Instagram, and Snapchat for treating users as products by pushing ads designed to make people buy things they may not want. He said it would be a dangerous world if Instinct used its intelligence to convince users to purchase or subscribe to something they did not want.

That message sits uncomfortably next to a feature that early users experienced as unsolicited advertising. The disconnect is less hypocrisy than the gap between a principle and a product decision, the same gap every assistant that wants to monetize will eventually have to cross.

Why proactive agents are hard to monetize without breaking trust

The design problem here is specific and worth understanding, because every agent company will hit it.

A proactive assistant needs context to be useful. Instinct connects to email, messaging, screens, audio, and location, and can arrange transportation, contact businesses, and make bookings. The more context it has, the more proactive and useful it can be. Shinn has shared data showing how strong that flywheel is: if a user shares a credit card number within their first three weeks, retention climbs to roughly 80 percent, and there is about a 40 percent chance a user shares a card within that window. He described it as a chicken and egg problem, where more data makes the agent more sympathetic to a user's situation, and it takes several weeks to build trust.

That same context is exactly what makes an unsolicited product recommendation feel intrusive. When the agent reads your inbox and notices a coffee subscription, then pushes a travel flask, it is demonstrating both capability and a willingness to convert your private signals into a sales prompt. The user sees the second part first.

The line between assistance and advertising is drawn by consent and timing rather than by whether money changes hands. A recommendation that arrives when the user is looking for something is service. The same recommendation that arrives because the system noticed a signal is an interruption, even if the underlying suggestion is good. Instinct has not said whether Selections generates revenue through affiliate commissions, commerce partnerships, or another arrangement, and has not detailed what controls users will get over it.

The contrast with bigger platforms is telling

OpenAI and Meta have not pushed unsolicited commerce suggestions into their agent interfaces, which gives users fewer commercial interruptions even as their agents get more capable. That is a competitive position as much as a design choice. The largest assistant platforms are still optimizing for trust and habit, and they have enough other ways to monetize that they can afford to.

A startup with a $10 billion valuation does not have that cushion. Instinct is approaching $1 billion in annual transactions through the platform, with more than half of that in travel bookings, according to Shinn. That transaction base looks like an obvious place to attach commercial revenue, and the geography of the recommendation data, mostly flights, gear, and destinations, lines up exactly with where the suggestions appeared.

What the episode teaches every agent company

Two lessons stand out. The first is that proactivity is a permission problem before it is a relevance problem. Even a perfectly relevant suggestion can fail if it arrives without the user's consent to be interrupted. Opt-in controls and transparent preference sources are conditions for proactive recommendations working at all, not features a company adds after pushback.

The second is that trust has to survive the monetization step. Instinct's own data suggests retention depends on users sharing sensitive information early, which means the relationship is fragile at the exact moment a company is under pressure to show revenue. A feature that reads private email to suggest purchases spends trust faster than a subscription fee does.

The monetization pressure behind the mistake

It helps to see Selections as a symptom of a structural problem in the agent business, not a one-off misstep.

An autonomous assistant is expensive to run. It consumes model inference, connects to many services, and holds long sessions. Revenue has to come from somewhere, and the obvious candidates are subscriptions, transaction fees, or some form of commerce. Instinct already handles close to $1 billion in annual transactions, more than half of it travel, which makes commerce the path of least resistance for a company that just raised at a $10 billion valuation and needs to show a revenue model.

The tension is that the same data that makes the agent useful is what makes commerce feel like surveillance. A travel agent that knows your itinerary is helpful. A travel agent that uses your itinerary to push accessories is selling to you. The line between the two is drawn by consent and context, and the initial Selections rollout crossed it in the most visible possible way: a push notification, unasked, based on a private inbox.

Every assistant with this much context will face the same fork. The companies that handle it well will treat recommendations as an opt-in mode with clear boundaries, and they will accept slower monetization in exchange for keeping the data pipeline open. The ones that do not will keep discovering that the fastest route to revenue is also the fastest route to losing the users who made the product worth building.

What to watch

Whether Instinct backtracks matters less than whether Selections gets redesigned into something users choose to receive. If the company adds granular toggles, discloses its commercial relationships, and lets users pick their own sources, the feature becomes a template for agent commerce done carefully. If it stays as-is, it becomes the cautionary example that other agent companies point to when their boards ask why they are not monetizing proactivity yet.