# Trust builds gradually, then drives much higher retention

Markdown copy of https://realalpha.app/s/trust-builds-gradually-then-drives-much-higher-retention/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Noah Shinn](https://realalpha.app/builders/noah-shinn/) · Noah Shinn is CEO of Instinct.
- Host: [Patrick O'Shaughnessy](https://realalpha.app/people/patrick-oshaughnessy/) · CEO, Positive Sum · Founded Colossus, the network behind Founders and Business Breakdowns.
- Show: Invest Like the Best
- Episode: [The $10B AI Assistant Challenging Meta's Muse, Grok's Bots, and OpenAI's Dots](https://realalpha.app/v/Am7IWP8IpEc/), Sep 28, 2026
- Moment: 23:10 to 27:22 of the episode (4 min), playing on https://realalpha.app/s/trust-builds-gradually-then-drives-much-higher-retention/; the episode from there: https://realalpha.app/v/Am7IWP8IpEc/?t=1390

## Quote

> “three weeks in, there's a **40% chance** that the user has shared a personal credit card with Instinct.”
> — [Noah Shinn](https://realalpha.app/builders/noah-shinn/) at [23:51](https://realalpha.app/v/Am7IWP8IpEc/?t=1431) · Invest Like the Best

## Takeaways

1. Three weeks in, 40% of users had shared a credit card.
2. Users who connect sensitive information have 80% retention.
3. Separate systems can block content and review actions before execution.

## About this segment

Noah Shinn says users take weeks to share sensitive data, and shares trust and retention figures. He describes separate safeguards that monitor agent inputs and actions.

From “The $10B AI Assistant Challenging Meta's Muse, Grok's Bots, and OpenAI's Dots”, Sep 28, 2026. [Watch the full episode](https://realalpha.app/v/Am7IWP8IpEc/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[23:23](https://realalpha.app/v/Am7IWP8IpEc/?t=1403) Talk about trust, what you've learned so far. There's an interesting data flywheel or chicken and egg problem here, which is the more data that you give to it, the more proactive that it can be, the more sympathetic to your situation that it can be, the more that it has to just act and be more useful to you. What we find in the data is actually that it takes several weeks to build trust. And I actually don't have a problem with that because one of the core principles on our end is the user should always feel in control of their data. They should always be in control of their data. They should share data with Instinct at the rate at which they feel comfortable with.

[23:57](https://realalpha.app/v/Am7IWP8IpEc/?t=1437) And if they want to take it back, they can certainly take it back. And what we find is over several weeks, I was looking at the numbers the other day, three weeks in, there's a 40% chance that the user has shared a personal credit card with Instinct. That's 40% of the user base. that's including some proportion of that user base probably churned in that moment too, or at least churned before that point.

[24:15](https://realalpha.app/v/Am7IWP8IpEc/?t=1455) So 40% of the user base, the reason why I'm sharing this is because time to first credit card or time to first account password or time to first sensitive piece of information, these are proxies for trust. And we actually highly value this and we take it very seriously. So 40% of the user base, three weeks in are sharing a credit card. Well, there's something there, no pun intended, it's like instinctual to go to it with whatever you need.

[24:37](https://realalpha.app/v/Am7IWP8IpEc/?t=1477) And so through that process over the first three weeks, the user learns to build trust. And then from there, it's just a snowballing effect. We actually find that when users connect at least one piece of sensitive information to instinct and really trust instinct, there's an 80% retention rate. 80% if you share one piece of money. It's crazy for consumer technology. How should people out there that want to try this, but have a natural reticence to trust, not instinct in particular, but just anything, any AI agent with all of its information?

[25:07](https://realalpha.app/v/Am7IWP8IpEc/?t=1507) Like, how should they think about the actual risk of doing this? And how would you reassure them that you've built your technology in such a way that the odds of something bad happening, if they do trust you with very sensitive stuff, are really low or close to zero? There's a way to break this problem down into two main parts. One is the storage of sensitive information. There are many other businesses, there are many other products that also deal with this problem of they have access to sensitive information and what are they doing proactively to make sure that that is isolated and locked down.

[25:36](https://realalpha.app/v/Am7IWP8IpEc/?t=1536) And then there's a second area, which is the new problems that we need to solve, the new surface areas or the new capabilities that we need to be aware of. The first case is this is a tractable problem. It's just very hard work and attention and care that you need to put into making sure sensitive data is safe and that the user is in full control over that data. There's that second category, which I'll spend more time talking about, which is this is the first time that an agent has been able to have within three weeks, 40% of the user base is giving instinct access to a credit card autonomously to be able to purchase theoretically anywhere. I don't know the numbers on email and what proportion of users connect in email, but you can imagine full access to an email inbox and to a calendar. So there's a lot of service area

[26:17](https://realalpha.app/v/Am7IWP8IpEc/?t=1577) here. There are systems that we've put in place that are detached from the instinct agent architecture itself that are put in place to be proactive about these things and to decouple the risk. So for example, any piece of content, any piece of text, anything, any form of media that comes in that might be consumed by instinct goes through what we call these firewalls, which can intercept, that can reject, that can block malicious pieces of content coming in, hitting instinct and trying to convince instinct to do something. There is also for every action that instinct might take or for every thought that it might have, that is being actively monitored by a system that is decoupled from instinct itself, which is able to pause it, intercept it, to approve or disapprove of what might happen next before it takes the action. I mean, those are just two pieces put in place,

[27:08](https://realalpha.app/v/Am7IWP8IpEc/?t=1628) but there's just so much more under the hood that is put in place to enable the agent to be as