Trust and safety at scale requires 40% of employees, functioning as the 'police force and legal system' of the platform -- equivalent to policing two New York Cities.
Grant describes the massive investment in trust and safety (40% of employees), comparing it to the police and legal system of a city with tens of millions of users. They use LLMs and rules engines processing billions of data points to detect harassment, late shipping, and high refund rates, automatically taking action based on seller history. ✦ AI generated
Grant LaFontaine · a16z Podcast · 2026-08-19 · original ↗
starts at this moment · 33:26
“What are some of the things that you encounter with that amount of scale that you constantly have to be on top of?”
trust and safety particular on whatnot is about 40% of employees. Yeah. So, it's au it's a huge huge investment from us because it's it's a challenging one and you don't want to put your >> credit card in a place you you don't trust. You certainly don't want to buy fresh fish from a place you don't trust, >> right? >> Um and and so so so those you know I think lot of measurement, lot of user feedback um and then a big just people investment to get the thing done is is sort of what drives it. Um, and then you know the problems are are pretty diverse and you have to make sure you're building the right set of things for it. So as an example, [clears throat] we have a a system in the background um which is called our our rules engine and it um takes in billions of data points at this point in sub 1 second to be able to take down or find any range of bad behavior on the platform. So if you harass someone in a live stream, the LLM will detect it. It will get plugged into the our action rules engine will look at the history of the seller and based upon other signals programmatically take them down, suspend them, ban them, warn them.
verbatim transcript · starts at 33:26
33:06that could break. So I'd love to understand some of those big things. >> Yeah, I mean it's it's it's challenging. Um I mean you know there's so many dimensions to to trust. So there's sort of like is the platform reliable and is it working and is it stable and is it fast and so um you know we measure just everything. >> Yeah. >> Um like I I think the the punch line on
33:28trust is measure everything. Um and uh we have a invest a lot in it. So I think um trust and safety particular on whatnot is about 40% of employees. Yeah. So, it's au it's a huge huge investment from us because it's it's a challenging one and you don't want to put your >> credit card in a place you you don't trust. You certainly don't want to buy
33:50fresh fish from a place you don't trust, >> right? >> Um and and so so so those you know I think lot of measurement, lot of user feedback um and then a big just people investment to get the thing done is is sort of what drives it. Um, and then you know the problems are are pretty diverse and you have to make sure you're building the right set of things for it.
34:16So as an example, [clears throat] we have a a system in the background um which is called our our rules engine and it um takes in billions of data points at this point in sub 1 second to be able to take down or find any range of bad behavior on the platform. So if you harass someone in a live stream, the LLM will detect it. It will get plugged into the our action
34:46rules engine will look at the history of the seller and based upon other signals programmatically take them down, suspend them, ban them, warn them. If you are not shipping on time, the the rules engine will do the same thing. If you have a high refund rate, the rules engine will do the same thing. And so, um, and look, these are ever changing, you know, problems and issues.
35:09>> Waiting of scoring and things like that will always change. Yeah. Um and and so you constantly have to uh be on top of these things. >> Yeah. Yeah. I mean, you know, if you think about just how many concurrent users you have and how many transactions you process per day. >> Yeah. I I always the the the analog I like for trust and safety is so you know
35:29tens of millions of people use whatnot every >> Yeah. Yeah. >> And so um the trust and safety team is basically the police force and the the legal system of of whatnot. You got to write your policies. You got to be people who enforce your policies. And you know, in a city that has 10, 20 million people, you're looking at a city the size of Tokyo or I mean, maybe a
- ·Trust and safety comprises 40% of Whatnot's employees
- ·System processes billions of data points in sub-1 second
- ·LLMs detect harassment in live streams automatically
- ·Rules engine takes action based on seller history
- ·Equivalent to policing two New York Cities
- ·Platform handles tens of millions of users
- ·Requires police force and legal system infrastructure
- ·Catches: harassment, late shipping, high refund rates
- ·LLM detects bad behavior in real-time
- ·Rules engine evaluates seller history and signals
- ·Programmatic actions: warn, suspend, ban, take down
- ·Sub-1 second response time at scale