00:00:00Analysts do this statically in one Excel sheet that they refresh manually every week, every quarter.
00:00:06Instead of doing that, BCI has instead used our artifact feature to connect directly to SMP and
00:00:12FACSAT datasets so that the artifact is a live dashboard of how these metrics compare against
00:00:18each other. And with one simple prompt to Claude, you can easily update it. And these artifacts are
00:00:23also shared with their managing directors who are directly interfacing with these platforms as well.
00:00:29So I think we're really seeing not just acceleration of work, but a way for the work
00:00:35to actually be transformed. Hey, my name is Alexander Brickin, and I lead our applied AI
00:00:40engineering team for financial services. Today, we're going to be talking to you about Claude
00:00:44for finance, and I'm joined by my colleague, Nick. Hey, my name is Nick Lin, and I lead product for
00:00:50Claude for financial services. I'm also a recovering investment banker and private equity investor.
00:00:57A lot of these problems we're about to talk about are very near and dear to my heart.
00:01:01So very excited, Alexander. Awesome. So Nick, my first question for you is,
00:01:04how do you feel about the shift in the AI landscape for financial services these days?
00:01:09You know, I've been with Anthropic for a little bit over a year and a half now.
00:01:13That was before Claude 3. So I think the enterprise AI landscape has changed significantly,
00:01:20especially in the past few months. What I am really noticing is that there is a fundamental shift from
00:01:27curiosity, observing from the sidelines, to actually starting to build and deploy into production.
00:01:34Now, as we all know, coding is one of the first products and first domains within AI with really
00:01:40strong product market fit. I think we're starting to see this really extend to other verticals as
00:01:45well, including finance. For example, Inbin or the Norwegian Southern Wealth Fund, one of our largest
00:01:52customers, they have about 9,000 portfolio companies. What they've done is they've built
00:01:57integrations on their own with things like model context protocols so that all of their portfolio
00:02:03managers are querying these integrations every single day to get insights into their portfolios.
00:02:09So I think we're really starting to see analysts spend a lot less time on the mundane, manual,
00:02:16tedious parts of the work and start to focus on what they really care about, you know, which is
00:02:21building relationships, meeting with their customers, and actually understanding the
00:02:25business models of the companies they're investing in. Yeah, that really resonates from my standpoint
00:02:30as well as an applied AI person. Whenever I go and interact with customers, a lot of the time
00:02:36last year, let's say, they would start with building an AI chat feature. Like they'd have a bunch of
00:02:42models represented and they would select one, maybe a random business user, and they would try to work
00:02:47with it and just chat with it. Eventually, now we've seen things like MCP come out where the chat
00:02:54has become so much more powerful. You can interact with the systems you care about. And I think that's
00:02:58really exciting specifically for finance because often there are just so many product surfaces
00:03:03that folks have to interact with. If you give a model a tool these days, often the model is
00:03:09intelligent enough to know what that tool does given the tool description and the tool name.
00:03:14But equally, the model has certain primitives baked into it, like the security that we try to bake into
00:03:20the way the model interacts with the world. So we train our models to be helpful, harmless, and honest.
00:03:25And often that's a reflection of the data that they interpret and the output that it basically
00:03:30corresponds to. So I think that's probably what you're referring to as well in that the model
00:03:34is generally intelligent. And so if you give it these different layers, you can really see some
00:03:38cool results. You know, safety is something that you touched upon. That is so foundational to everything
00:03:44we do. It's about securely deploying these solutions into enterprise environments. It's about making sure
00:03:50that the models can accurately answer the questions with the right level of understanding of those
00:03:55problems and fidelity. And third is actually giving our users the trust, the verification, the auditability
00:04:04to understand these results. So I think we think about all three of those components of safety.
00:04:08Yeah, I mean, speaking of, right, Anthropic was founded on the principles of AI safety. It was a
00:04:14research org from scratch. I'm curious, how have we gone from being a research org to releasing a
00:04:21distinguished product in financial services? In my mind, Anthropic really aims at building models
00:04:28that can be safely deployed to solve the most complex and difficult problems in the world, right?
00:04:36We're state of the art when it comes to code. 0.5% of the world's population are software engineers.
00:04:43So that is just one sliver of these really complex, difficult problems we can really start solving,
00:04:50right? They really exist everywhere else in the world. Code is so foundational to every single
00:04:55part of a company, right? It is how a company is run. So that means that cloud is really great at
00:05:03interacting with more complex systems, being able to expose thinking and its logic, and that's why
00:05:09it's great at finance as well, right? Finance are complex problems deployed into regulated verticals
00:05:14that need verification, auditability, and ultimately accuracy really matters. Financial analysts these
00:05:21days spend a lot of time getting down to like the pixel perfect level of, let's say, a PowerPoint deck
00:05:26or an Excel model, right? You can't get anything wrong. And it's funny now that we're in this
00:05:31paradigm where models can do something similar, but using the capabilities they have to write
00:05:37really structured logic. And so that's actually what we've found language models to be good at,
00:05:41what we've trained them on. And that ability to do that, it feels like it's just being abstracted
00:05:47into so many other domains, like creating Excel spreadsheets or like creating PowerPoints.
00:05:52And so, yeah, it's just been like super kind of striking, at least to me, to see how many domains
00:05:59the like logic and reasoning of these models actually ends up touching.
00:06:03Ultimately, these are digital systems that we interact with every single day, right?
00:06:07The fact that Claude is great at code gives it a flexible skill and a shortcut to do all of these
00:06:14really cool interesting things, right? Our file creation feature that was launched a few weeks ago
00:06:19that enables Claude to create Excel documents in PowerPoint is essentially Claude accessing
00:06:24a virtual machine within which it can run Python code at scale to edit, analyze, and create Excel
00:06:33documents and create these perfect DCF models, which I think is super exciting for us, right?
00:06:37So I think there's a lot of other domains that code can start really unlocking.
00:06:42What's different to Claude for finance versus other products on the market in the financial services?
00:06:47You know, there are three verbs, I think, about a lot that governs what I want to build
00:06:52for Claude for finance. And these are retrieve, analyze, and create. Starting with retrieval,
00:07:00many of the research agents on the market has seen, you know, quite a lot of maturity, right?
00:07:04Large language models are fantastic at digging into large pools of data and gathering insights.
00:07:10It can read in a 5,000 probably times faster than humans. But what we want to do with finance is
00:07:15making sure that these systems can connect to all of the core data sources that finance analysts work
00:07:21in. In finance, the ability to uncover insights faster than your competitors and your peers,
00:07:28that's really a key advantage. Now, downstream from that, it's great that we can retrieve
00:07:33this information and connect to it. But the ability to do analysis at scale, either through code or
00:07:40through spreadsheets, is so foundational as well. Financial models themselves, they're not just these
00:07:45beautiful Excel sheets, right? They're a way for finance analysts to inject their own judgment of
00:07:52what the future looks like and what the proper valuation looks like for that company, right?
00:07:57So with that in mind, we want Claude to be really good at understanding these core finance concepts
00:08:02and manipulate systems like Excel and spreadsheets to be able to do that calculation. And then the
00:08:07third part is creation, right? We're all social creatures within the enterprise, right? We do our
00:08:14work to be shared with others. So the outputs themselves in the form of spreadsheets, PowerPoint
00:08:20documents, Word, doing this in a way that is client ready, boardroom ready is really important. So we
00:08:28really want to start pushing Claude's capabilities to be able to do that as well so that it is an end
00:08:33to end agentic autonomous system. That makes a lot of sense. I feel like we build these primitives and
00:08:39then they almost end up snowballing. So you have like the retrieval step, right? You build an MCP
00:08:44server to connect to one system. But then if you take the data from that system, maybe it connects
00:08:49to some other system in a unique way. Like you get data from Snowflake, let's say. You find an ID in
00:08:55there and you need to connect it to your Salesforce instance. You can easily do that with some of those
00:09:00primitives that we've built on the retrieval side. But then it sort of continues to snowball. You get
00:09:04analysis where Claude can write a bunch of code and essentially piece together some of that information.
00:09:11And then finally, the creation is even take that one step further and put it into the environment
00:09:15that someone cares about. Sending that post request back to the API example to a system where an
00:09:22analyst or an operator can see the information that Claude has reasoned through. So let's talk a little
00:09:27bit more about what is actually Claude for Finance. How does it work? What makes it so special? So there
00:09:32are three layers that we think about in our solution. The models, the agentic capabilities, and
00:09:39the platform. Starting with the models themselves. Fundamentally, we are a research lab, right? Everything
00:09:46we do really aims at making Claude the best model for financial services. Now finance presents some
00:09:53interesting challenges to us, right? Code is something that we can test every single day as software engineers
00:09:58and product managers. But there are very few investment bankers within these four walls of
00:10:03Enthropic. So here's where we're really excited to work with early customers like BCI, Pearl at
00:10:10Weinberg, and MBIM to really let us know what are the use cases they really care about? What does good
00:10:16look like? And then help us, much more importantly, uncover those gaps that we can bring back into the
00:10:22research process. The second thing is on the product side, right? Agentic capabilities are essentially
00:10:28the code that we write to enable users to interact with the models. We've built capabilities like deep
00:10:35research. Now we're really investing in being able to embed Claude in all of the core surfaces you work
00:10:41in. Not just Claude for Enterprise, Claude AI, but also the browser extension, Excel, Chrome, and other
00:10:50surfaces that our analysts and enterprise customers work with every single day. The last piece is we
00:10:55want to again build a really flexible platform that can be tailored and deployed very easily for our
00:11:02customers. That's why we've been spending a lot of time with industry partners like S&P, Faxat, Pitchbook
00:11:08to build these integrations so that these agents can be as powerful as possible. So I'm curious, how has
00:11:14adoption been, right? Who's using this? Why are they excited about it? Walk us through that. As I mentioned
00:11:19before, we're really seeing pockets of adoption across the entire industry. I'm often asked, you know,
00:11:25which sub-verticals do you see AI adoption in in finance? I think it's much less about sub-verticals,
00:11:32but much more about the culture that our customers have really engendered, right? Which requires a good
00:11:38combination of top-down encouragement and adoption to lower the barriers, but also a bottoms-up
00:11:46experimentation culture, right? To try all of these tools out there to figure out what makes sense.
00:11:51With that in mind, I think some of the main customers that we've seen strong adoption from,
00:11:56BCI, for example, they've sort of fundamentally transformed the way they work. There are these
00:12:02things called comps analysis that analysts do, which basically means you're comparing comps,
00:12:08financial and operational metrics for all of these different companies to figure out whether
00:12:14they're trading at the right value. Analysts do this statically in one Excel sheet that they refresh
00:12:19manually every week, every quarter. Instead of doing that, BCI has instead used our artifact feature
00:12:26to connect directly to S&P and fact-set data sets so that the artifact is a live dashboard of how
00:12:33these metrics compare against each other, and with one simple prompt to Claude, you can easily update
00:12:38it. And these artifacts are also shared with their managing directors who are directly interfacing
00:12:44with these platforms as well. So I think we're really seeing not just acceleration of work,
00:12:50but way for the work to actually be transformed. Memory is such a fundamental piece of how humans
00:12:57basically exist in the world, right? You have to memorize things to like know where you put your
00:13:00keys last, for example. How are we building that into our models? And why is that important for
00:13:07financial services? The way that we think about how we work with our customers, as I mentioned before,
00:13:12there's very little that we can internally test for these finance use cases, is to, again, work
00:13:18really closely with enterprise customers to understand where things are working with or not,
00:13:22right? And memory systems is something that's really important to allow Claude to understand
00:13:28and maintain contacts across all of these different tools and surfaces that it works in. Claude is in
00:13:33Cloud AI, in Excel, in the browser, interacting with facts at S&P, the ability to understand patterns,
00:13:42understand preferences for that, you know, DCF template that you want Claude to remember. All
00:13:47of these things are really important to just make sure that Claude stays and in turn that
00:13:51continually gets better through his interactions with you. And so like over time, you could imagine
00:13:57someone prompting the model like, "Hey, you got this formula slightly wrong." And then Claude has
00:14:01some way of storing that memory, whether it be a file system or it's implicit, et cetera, which is
00:14:07pretty awesome. I'm excited for that. Or if, you know, the user and analyst really wants to use S&P
00:14:13for a specific piece of EBITDA calculation, Claude should remember those preferences too, just like,
00:14:19you know, a good intern would. Cool. So we've talked a lot about Cloud for Finance. I'm curious,
00:14:24in your opinion, what's next for our product and research orgs in relation to making Claude better
00:14:29for finance? You know, take a step back. Anthropic is enterprise-focused, enterprise-first. The only
00:14:36way for us to deliver outcomes to the enterprise is to focus on specific domains. Finance is one of
00:14:42the most important domains for Anthropic across the entire stack, research, product, and go-to-market.
00:14:50Starting with research, we're finally starting to invest in both specific pre-training and
00:14:56post-training for finance. On the product side, three things I'm really excited about.
00:15:01One is going much deeper into specific subverticals. Private equity has very different needs from hedge
00:15:10funds and insurance firms and investment banks. You want to really start understanding and peeling back
00:15:15the nuances of those workflows and make sure that the components we're building fully serve those
00:15:21workflows. We're also excited about the ability to have Cloud everywhere, right? Not just in the
00:15:27browser, but within Excel, within PowerPoint. On PowerPoint and Excel, I think we still have a lot
00:15:33of room to improve the quality of those outputs. So excited to work again really closely with research
00:15:39and bring these capabilities into the product. On the partnership side, it's really important for us
00:15:44to work closely with the industry. It's been really encouraging to see the fact that MCP servers have
00:15:50only been out for six months and major industry leaders like S&P and Facset have already published
00:15:56functional great versions of their own MCP servers. We want to keep bringing the industry together,
00:16:03including some of the recent announcements we've made. The last piece is working really closely with
00:16:08our enterprise customers, right? Fundamentally, that's how we work together, right, to translate
00:16:15what their needs are and help us build the research and product capabilities to meet those needs.
00:16:21I definitely agree with that, because not everyone comes from a financial services background like you
00:16:25at Enthropic. And so I feel like we learn the most from the customers that we're going deep with,
00:16:30specifically when they're like designing evals, for example. That gives us so much signal about
00:16:35how the model actually works in production. And I think that level of collaboration is what we're
00:16:40going after with Cloud for Finance. I think that's the main thing I would encourage our enterprise
00:16:45customers to think about. You know, evals sound like these mystical concepts, but they're really
00:16:51simple. There are tasks you care about and problems you want to solve and an articulation of what good
00:16:58looks like for those tasks. It's really important for enterprise customers to be thoughtful about
00:17:04these problems rather than thinking about, oh, I need to infuse AI into every part of my business.
00:17:09And that's how we can partner really closely with enterprise customers. We bring those evals
00:17:13directly into the training process, directly into the product pipeline so that we can deliver these
00:17:18capabilities to our customers. 100 percent. Well, thank you so much, Nick. This was
00:17:23fantastic. Appreciate you taking the time. Thanks for having me, Alexander.