What I learned from Instinct

Instinct reached a $10 billion valuation about a year after Noah Shinn started building it, while Meta, OpenAI and the company behind Grok were pursuing the same personal-assistant market. That made me curious about what this new company was doing.

Shinn's conversation with Patrick O'Shaughnessy gave me a closer look at its choices: how people start using Instinct, become comfortable delegating and find reasons to keep using it.

Start with something people already do

Instinct takes requests through text and calls, so someone can start with a familiar way of asking for help. A person can have a task in mind without also knowing which application to open or where its controls live.

Lovable's Telegram interface offers a similar starting point for building and managing projects, although some steps still send the user back to the website.

For an established product, I would look inside its existing workflows. If a customer is already reviewing an invoice, that may be the place to help them discover a useful approval feature. Announcing it does not ensure they will remember it when the invoice arrives.

Make the behavior understandable

Noah emphasizes whether users can predict what Instinct will do, including how it communicates.

A person has to understand enough to try something, recognize whether it worked and know how to correct it. With an assistant, that includes knowing when it will act and when it will wait for approval.

I would check whether people can explain what they have delegated, then compare that understanding with what happens. Repeated corrections could reveal a confusing agreement as well as poor execution. Adding another capability would leave that problem in place.

Let trust develop through use

Noah says trust develops over weeks and users should share access at their own pace, with the ability to withdraw it.

This makes me reconsider onboarding that asks for every connection before someone has experienced much value. The user is making a decision with very little evidence.

Notion's August release lets an agent propose edits for individual review. Lovable's Telegram interface asks for approval before changing something through a connected tool, unless the user has already granted permission for that scope. Both let people inspect a proposed action.

More permissions would not, by itself, show that this works. People who already find a product useful may be more willing to connect their accounts. I want to know whether the product can remain useful while that confidence develops.

Give people a reason to involve someone else

Instinct's trusted network lets assistants coordinate scheduling, with different access for different relationships.

Arranging something together gives people a practical reason to involve another person. That gives me more confidence than an invitation reward by itself, although it does not establish a durable network advantage.

Before copying the invitation system, I'd want to see whether bringing another person in removes coordination work. Sending an invitation is a small result if the two people still have to settle everything themselves.

Keep earning the next use

Instinct includes cancelling forgotten subscriptions among its uses. If handing off that task becomes easier, businesses that depend on customers postponing cancellation have more to lose.

But inertia also affects useful products. A customer can intend to use a feature and never get around to it. Instinct describes proactively calling or texting users and following up on unfinished threads. An assistant could use that contact to offer a specific next step.

Suppose someone asks it to watch for a supplier's reply. A scheduled daily check could notice that the reply is still missing and send: "The supplier hasn't replied. I've drafted a follow-up. Want me to send it?"

The person can review and accept, without remembering the task, opening the thread and writing a new request. That offer could also introduce a feature they have never used. This is the kind of engagement I would want to build, with timing and frequency the user can adjust.

Continue the work between visits

Noah describes moving toward ongoing objectives, such as maintaining inventory within a range, rather than waiting for each individual order.

A person could close the interface while the service remains responsible for something they care about. Counting visits would miss some of that use.

Keeping inventory in range also means dealing with cash limits, uncertain demand and supplier failures. The assistant needs to know when it can place an order and when it needs a decision. Sometimes the useful return happens when the product has finished something or needs help resolving an exception.

Plan for the cost of being useful

Noah says invitations help control demand against compute capacity. He also describes making work that can wait cheaper to serve through customized inference.

When Anthropic introduced its batch API, it offered lower prices for work that did not need an immediate response. Quora used it for summarization and extracting highlights. That supports the mechanism, without verifying Instinct's own cost or performance claims.

Preparation for tomorrow's meeting can wait longer than a reply someone needs during a conversation. Treating both as urgent could make background work unnecessarily expensive.

Cancelling a forgotten subscription can save the user money without generating a new purchase. A business funded by purchases would still need to cover that work, along with reminders and checks that never lead to an order.

Before opening access further, Instinct has to arrange capacity for the work new users will delegate, including tasks that continue after those users put down their phones.