Trust is the bottleneck, not the model

Ryan Green: Hello everyone and welcome back to Data Cloud Now. I'm currently in San Francisco where I'm now joined by Cat De Jong, Head of Applied AI at Anthropic. Cat, such a pleasure sitting down with you. Thanks for making the time.

Cat De Jong: Of course, it's super exciting to be back here after my hiatus from before I was at Snowflake.

Ryan Green: Good to see you, welcome back. Yeah. Should we jump into it? Let's go. Let's do it. Cat, you spent years on the data side, including, as you just mentioned, at Snowflake before leading the Applied AI team at Anthropic. How did that shift change? Taking that again, how did that shift change how you think about what it actually takes to get AI working in the enterprise?

Cat De Jong: I mean, I think the first thing I'll say, it's exciting to get to be on both sides of the table. When I was at Snowflake, we spent all of our time doing everything around making sure we're bringing data together all into one place, drive really complex analysis and help solve business problems and processes that we have. And so now that I'm at Anthropic, what we're doing is my team is responsible for helping customers build their first agents. Everything around the data that they need to get access to it, the evals, the custom integrations, the list sort of goes on and on from there. If I take a step back to Snowflake, access to the model felt like the hard thing to have to solve for. But in honesty, everyone here has access to some of the most frontier models, whether it's through Snowflake with CoWork now and CoCo, which I think you guys just rebranded today, or through any of the other providers out there getting access to the models. So access to frontier intelligence is no longer the problem. What the challenge really is, is making sure that you have your data house in order to be able to use the models effectively. You have the right governance in place. You have the right access so people and the models get access to what they need to. You need to think about who's going to be making decisions afterwards with that data. So that's been a bit of a mindset shift of like data is really a big part of it. The other thing I'll share, and we were talking right before this about just how crazy things have been. We are truly living in the exponential right now, which is such an exciting place to be. Model capabilities are jumping exponentially from where they were before. About a year ago, we launched Opus 4. We're now at Opus 4.8 as of last week. And what it's capable of doing is just miles above what it used to be able to do. But when enterprises think about actually adopting AI or adopting Claude, they tend to think about it on a linear scale, which means the gap between what model capabilities are and what enterprises are doing is continuing to get larger and larger. And so what we end up spending time with our customers thinking about is you don't want to be planning for what the model is able to do today, but setting things up, including your data infrastructure and everything else for what the models are going to be capable to do in the next quarter or two quarters. Because if you don't do it now, it's just going to keep running away from you.

Ryan Green: Very much that long-term vision, Cat. And your team helps companies across industries bring Claude into their operations. What separates the companies getting real value from the ones being stuck in pilots?

Cat De Jong: We like to think about this in sort of three different steps when it comes to companies adopting Claude. The first one, we call them our pillars. So our first pillar is employee empowerment. And I think this really needs to be the starting point for anything everyone does. It's getting... Claude into the hands of users that need to learn how to trust it every single day so this could be someone using Claude Code to actually uh complete a coding task it could be using CoCo or it could be a business analyst asking questions of CoWork inside of inside of snowflake um so that way they're getting answers to their natural language questions and so the more and more these people can actually trust the technology see that it actually solves problems and saves them time right they're going to trust it so much more and actually be willing to take on bigger tasks that are going to have even broader business impact to the organization so that's pillar one we then think about pillar two which is faster processes which is how do you think about picking an internal workflow to really re-imagine it into an agentic workflow and so a couple of things we recommend from that is you want to pick one i think sometimes if you pick too many at once it's something It can be a little overwhelming. It's just too overwhelming. There's too much that you have to do. You want to have one owner that drives it and it needs to cross multiple teams so that multiple teams can feel the impact of it. So I think that's the first thing. The second thing is you want to make sure you're bringing the domain experts. Hopefully they're people that you've already convinced with Claude Code or with CoWork or with CoCo that like AI is what they should be thinking about along for the ride because they know the domain so well and they can help when it comes to actually putting this into process. The third is you really need to think through, you don't want to just rebuild it piecemeal. You want to reimagine what the entire process can and should look like.

Ryan Green: Kind of that holistic approach. It's holistic.

Cat De Jong: You don't want to just lift and shift it. It doesn't work. It's the same thing of like when you're, I remember back in the day, you didn't want to lift and shift your on-prem data warehouse into Snowflake. It just didn't get the same as if you reimagined what it could look like in Snowflake. So similar concept. And then the fourth piece, and I touched upon it, it's the data infrastructure, the data governance, making sure that the data can actually get to the model in a way that can be trusted. So that way you can actually use it. And then the third sort of on our path from our pillars, I'll just state is building transformative products. And so I know Fin is an example of this, is that they use Claude through Snowflake to power their support agent. Intercom does that with Fin, which is incredible. And so that's where we start seeing the explosion of business impact afterwards. It's your individual employees. You then take your internal workflow processes and then you build brand new products that never existed before.

Ryan Green: Cat, I love that breakdown. Thank you for walking us through your pillars. And as frontier models continue to advance, where does the real competitive edge come from and what should companies invest in right now?

Cat De Jong: I think there's two things and part of it is one of the things that you all do very well. It's data, it's the governance, it's the structure around it. The model can't work with bad data. So if you put that sort of at the core, you need to have the right data to be able to power anything that you're looking to do from a workflow perspective or from like an agentic perspective. So that's the first thing. but it can't be by itself. You actually need to make sure that you're partnering with the right AI partner, one that actually believes in safety and trust and reliability, because you can't pair a model that you don't trust with your most sensitive data. So I think those two things in parallel have to happen at the same time in order for you to actually be able to build what you need to for your agentic workflows.

Ryan Green: Super clear perspective and central to Anthropic is constitutional AI and it's very much ingrained into everything you do. How do you balance safety with performance?

Cat De Jong: It's core to everything that we do here at Anthropic. There is no trade-off between safety and performance. You can have an incredibly intelligent model, but if you don't trust it and it's not safe, you're never going to actually pair it with your most important data or in the workflows where it's going to have the biggest impact with highly regulated spaces or your customer facing workloads like that needs incredible trust and safety around it because it has such a broad impact. So for constitutional AI, what that essentially means is that instead of having a very robust rules engine, we spent time training Claude on a constitution. So essentially the way that it should be thinking about how it should approach these different decisions. And because of that, it is able to take on more complex tasks when we actually put Claude into production. So you're never going to be able to know every single edge case and Claude has the ability to reason through that and then reach back out to a user when it needs to. I always like to think of the Goldman Sachs example that we've posted relatively recently. About six months ago, my team went and embedded directly with Goldman Sachs to rebuild their Know Your Customer workflow into an agent. And for anyone who knows that, it's highly regulated. Yeah. really expensive. You get it wrong. Like you don't want fraud inside of your bank. And Claude was able to handle that because not because we gave it every single rule it needed to know, but because we gave it the ability to reason through it and then to bring in the experts in that they needed along the way.

Ryan Green: I love that example. Thank you for that, Cat. And how do you bring frontier intelligence to everyday decision makers and what makes it trustworthy?

Cat De Jong: Great question. I'm going to point to, I don't know if you've ever had a chance to meet Mike Krieger. He has an analogy that I absolutely love. So for him, Claude is the brain and it's an incredibly powerful and incredibly smart brain. But for a while, Claude didn't have the right hands to actually do the things that it needed to do with its brain. So it could give you an eloquent and incredibly intelligent response, but you couldn't do anything with it. And so we've actually spent the last couple of years giving Claude its hands or the tools that it needs to connect to the different applications, to the different systems, to get access to the right data. And with that, you can now actually start getting it into the hands of users. Users are never going to go and really don't want to have to use a chat box that's external. You need it to be in the place where they're working every single day. And I think that's what actually sort of like really helps drive it with end users. And then I think the second thing with around trusting it is it all comes down to, again, you have the right data, it's trusted data, it's governed data, and you combine those two things together and it really just helps end users adopt the technology.

Ryan Green: Well, Cat, it's been such a pleasure. I got one final question for you. There's a lot of vision talk about autonomous agents, but you're the one putting them in customer environments. What's real today? What's still say a year out and what excites you most?

Cat De Jong: Yeah, I think what's real today is agents that have scope tasks can take them on end to end. And there's a lot that they can do with that. So it might seem simple, but the number of things that agents can go out and take care of, whether it's a coding task or solving a customer support problem, like I mentioned with Fin, those are two things that have just such a big impact and agents are able to do that. I think one interesting thing I'll call out is that what Claude does is Claude has this tendency to come back to a user and interrupt is probably the wrong word, but ask clarifying questions two times more than an end user is willing to interrupt. And because of that, it's starting to build this trust. It's like trying to figure out exactly when it makes the most sense to get input from those end users. So builds a really trusted environment from there. And I think what's important before anyone can take the next step, like you need to have that trust with the model and understand what it's thinking and it's doing. I think what's still to come in, I don't think it's model intelligence because we know we're on the exponential, but it's giving models full access to across a large number of systems and giving it authority to do different tasks across those systems. And I think a lot of that comes down to we need to make sure we have the right governance in place. It has the right access to the data. It has the right permissions. There's so much around scaffolding that still needs to be put together to make that. And I think what's exciting, I already mentioned Fin and the Intercom team and what they're doing, but just what some customers are accomplishing today is just, it's incredible. I know Simon Data is using Snowflake and Anthropic today, and they're able to uncover problems that they didn't even know that they needed to look for. And I think that's where the next step is going to be. It's going to be really around doing discovery that us as humans just didn't think we needed to do, and it's going to take us to the next level.

Ryan Green: Well, Cat, such a pleasure sitting down with you. Welcome back to Snowflake Summit. I really enjoyed our time together. Thank you. And for the audience watching, I'm Ryan Green with Data Cloud Now. Thank you so much for tuning in. We'll see you next time.