How Claude's AI Financial Analyst is Changing Investing (Insider Unveils)

Analysts spend 60-70% of their time on mundane rote work: rebuilding models from scratch, relinking cells, checking for circular references, and creating PowerPoint presentations with formatting—tasks Claude can automate.

“our core focus area right now, as I mentioned before, is really starting to peel away all of this mundane wrote work that probably analysts spend 60 to 70% at least of their time doing. Yeah. Right. you know, rebuilding the model from scratch every single time and relinking all these cells, making sure that you have all of the circular references, you know, all checked out and error-free and again all of the PowerPoint creation as well, which I think mostly is just about formatting and translation and making sure that the font sizes and the colors really match, right?”

For deployed enterprise customers, Anthropic provides standard support through teams of customer success managers and applied AI that support daily workflows, with more intentional design partnership for customers building specific new product capabilities.

“for all of our deployed customers, we have four teams of customer success managers and applied AI that really supports their daily workflows. So that really just, you know, is our standard deployment and support model for all of our enterprise customers. I would say for you know building specific product capabilities we are much more intentional about who are the customers we're hoping to target and how to bring them into the product development life cycle.”

Fundamental Labs built an Excel agent called Shortcut on top of Claude Opus that reconstructs the entire Excel application interface in the browser and replicates Excel capabilities while building Claude intelligence into the product.

“One of the customers that we work with really closely at Fundamental Labs, they've built an Excel agent called Shortcut on top of Opus. They've actually reconstructed the entire Excel application interface in the browser and has honestly done a tremendous job of replicating a lot of these capabilities and building clause intelligence into into the product itself.”

Software developers represent only 0.5% of the global workforce, while finance and financial services employ a much larger population, making finance a more impactful vertical for AI deployment to reach more workers.

“0.5% of the world are software developers, but I think coding is a fantastic starting point for us to think about how to solve some of these harder problems, right?”

Claude Sonnet 4.5 outperforms Claude Opus 4.1 by five full percentage points on the FinBench Agent benchmark, with Sonnet 4.5 achieving 55% accuracy compared to Opus's approximately 49% after focused training over a few months.

“sonnet 45 outperforms our opus 4.1 model which previously topped the charts by five full percentage points... I think it's 55% I think 03 is at like 48% right now... Exactly. and Opus is just about a little bit like 49 or something like that.”

Enterprise AI deployment requires safety considerations across three dimensions: data security (bulletproof systems preventing data leakage), accuracy assurance (ensuring produced answers are correct for specific use cases), and auditability/trustability (enabling humans to trust accurate answers through citations and audit trails).

“A lot of what we haven't talked about is how do you actually deploy these solutions safely within the enterprise and safety has a few different layers, right? Number one is making sure that from a data security perspective, the systems we're building is bulletproof. Second is understanding that the answers that are being produced is accurate for use cases. Third is making sure that humans have a way of actually trusting these accurate answers with auditability and citations.”

Microsoft has announced that Claude is now available inside the Office suite and users can toggle between different foundational model companies, giving Microsoft leverage to distribute Claude to billions of Office users.

“I think Microsoft recently announced SATA said that Claude is now available inside of the office suite and they're building AI tools and you can toggle between different foundational model companies.”

Anthropic deployed Claude to Deloitte's 450,000 employees globally approximately one year and a half ago as one of the first Claude for Enterprise customers, and has now rolled out across all employees despite consulting firms having disparate needs across integration, management consulting, and implementation arms.

“Deoid was actually one of the first claw for enterprise customers we signed maybe about a year and a half ago... and you know consulting firms are really interesting because one they have multiple arms right... They have integration arm, they have management consulting, they have four deployed engineers... And they have 450,000 employees and management consulting is very different from implementation consulting... but you're live with all 450,000 now. Exactly.”

Anthropic's goal is for Claude to be the backbone model for financial services regardless of how enterprise customers adopt it—whether through Anthropic's own application, third-party UIs (e.g., Abbvie's AI financial agent), or custom integrations.

“You know, again, Anthropic is ultimately a research lab. What we really care about is delivering the highest quality model intelligence to the industries that we really care about. And my goal is for claw to just be the backbone for the financial services industry regardless of how our enterprise customers want to adopt.”

A major obstacle to AI adoption in financial services is that many institutions have not built great data lakes, making it necessary to invest in data infrastructure before applying AI capabilities.

“but it's very important a lot of financial services institutions have not built great data lakes, right? And so like, oh, I need to have an AI strategy. You're like, well, do you have you talked to Snowflake yet? And like, have you built? And so there is this huge impetus and some of the big winners I think right now are going to be the folks who are helping those organizations just build the data layer before they can even apply intelligence on top.”

Design partnerships don't need to be highly programmatic; they just require regular touchpoints with customers, such as weekly standups, to share ideas and get validation.

“I think you know there's a misconception that design partnerships need to be very programmatic. It doesn't really need to be. Yeah. As long as you have regular touch points with your customers. So I have a weekly standup with BCI for example where I share all of these ideas in my head and it's really important for us to just get validation from you.”

Deploying AI solutions safely within enterprise requires three layers: data security (bulletproof systems), accuracy (ensuring answers are correct for use cases), and auditability/citations (allowing humans to trust and verify answers).

“A lot of what we haven't talked about is how do you actually deploy these solutions safely within the enterprise and safety has a few different layers, right? Number one is making sure that from a data security perspective, the systems we're building is bulletproof. Second is understanding that the answers that are being produced is accurate for use cases. Third is making sure that humans have a way of actually trusting these accurate answers with auditability and citations.”

The primary problem Anthropic's product strategy aims to solve is change management: even if AI capabilities are fully functional today, they don't matter if users have to significantly change their behavior to use them.

“Yeah, for sure. You know, the way that we think about building claude into the enterprise is really hoping to solve the problem of change management right a adoption is happening really quickly Even if these model capabilities, product functionalities are, you know, fully capable, fully functional today, it doesn't matter if our users have to significantly change their behavior to use these model capabilities.”

Analysts should focus on understanding markets, understanding business models, and spending time with founders and companies to understand business models, rather than on manual data manipulation.

“But what should analysts really think about and focus on? And what are they passionate about? They're passionate about understanding markets. They're passionate about understanding business models. And they're passionate about spending time with founders and their investy companies to understand those business models much more. So we want to really start freeing up the time to do all of those things that actually really matter.”

A lot of similar research use cases across customers include understanding and picking up trends within large datasets that are easy to miss, and spotting these insights is where alpha (investment outperformance) comes from, not just from faster data processing.

“I think a lot of the similar research you know use cases that we see right really understanding picking up trends within these large data sets that's really easy to miss and that's ultimately where alpha comes from right it's not just about processing the data faster”

Published benchmarks serve as a starting point for evaluating AI capabilities, but enterprise customers should not rely solely on them; instead, they should develop their own internal versions based on problems they're hoping to solve.

“I think it's also important to think about what benchmarks really serve in terms of their purpose, right? you see a lot of publish published benchmarks that are just more academic representations of what these things could do in theory. I would always encourage all of my enterprise customers to use those benchmarks as reference, but really think a lot more about what problems they're hoping to solve internally in developing their own versions.”

Anthropic's goal for AI agents in finance involves three verbs—retrieve, analyze, and create—which map to research/data gathering, qualitative and quantitative analysis, and output creation in formats like documents, spreadsheets, and presentations.

“You know, our job at Anthropic is building virtual agents or collaborators that are fully autonomous and that can own decisions and projects end to end. There's three verbs I like to think about in terms of what these agents could do. very similarly to what we can do. Yeah. As knowledge workers, as podcast hosters, as product managers, right? And that's retrieve, analyze, and create. Right. Everything starts with the research that we have to do and the data we have to gather. Yeah. Downstream from that, we do qualitative, quantitative analysis on that data. And ideally, we create outputs that can be shared with others in the form of word documents, spreadsheets, and powerpoint documents as well.”

Anthropic uses a 'research, product, and customer flywheel' to improve capabilities: product features give customers access to model capabilities, customer feedback identifies where things work or fail, and feedback is brought back to research teams to improve models.

“Yeah. And I think the way that we think about how we improve these capabilities in the future, there's a term we use internally quite a lot, the research product and customer flywheel. Again, ultimately anthropic or research lab. Yeah. My belief is that the product features that we build give our customers access to these model capabilities. Not everyone is a developer that is comfortable talking to Claude just you know bare bones within the terminal, right? You have to build tools and integrations and surfaces to make these model capabilities actually useful for you. So we want to do that. That's my main job, right? Build these model capabilities and product features so that my enterprise customers can interact with cloud and they let me know where things are working and where they're not. And we bring all of this feedback back to our research teams.”

Finance is 10% of global GDP and represents a massive industry where AI is barely scratching the surface in terms of problems that can be solved, making it an ideal vertical for Anthropic to focus on initially.

“Finance is 10% of GDP. It's a massive industry where I think we're just barely scratching the surface in terms of the problems that we can solve”

BCI integrated Claude and artifacts (Claude's feature that connects to information and uses code to render data in real time) against S&P and Factset data sources to build live artifact dashboards that allow managing directors to interact directly with analysis rather than asking analysts to rerun calculations.

“There's a feature within claw.ai called artifacts. Artifacts are essentially a way to connect to information and allow claw to use what it's the best at, which is coding. Claude uses code, renders code in real time to display a ton of information. So instead of using these static Excel sheets as comm sheets, BCI's integrated against S&P and fact and built live artifact landscapes, yeah, for themselves, for their managing directors to interact with even their MDs are talking to these artifacts on a daily basis instead of having to ask analysts to rerun calculations.”

Norges has seen big adoption across public market use cases where they analyze structured data within Snowflake, having invested in building up a high-quality data lake with ingested core data sources—work that is necessary before AI can generate alpha.

“We see a big um set of adoption across their public market use cases where they are going much deeper into structured data within Snowflake. I think Nor has also done a really good job of they spend a lot of time making sure their data lake is to your point up to par and ingesting all of the core data sources.”

Many financial services institutions have not built effective data lakes, and organizations implementing AI strategies need to address this foundational data infrastructure problem before they can apply AI intelligence on top.

“A lot of financial services institutions have not built great data lakes, right? And so like, oh, I need to have an AI strategy. You're like, well, do you have you talked to Snowflake yet? And like, have you built? And so there is this huge impetus and some of the big winners I think right now are going to be the folks who are helping those organizations just build the data layer before they can even apply intelligence on top.”

Investment banking and financial analysts spend 60-70% of their time on mundane rote work—rebuilding models from scratch, relinking cells, checking for errors, and creating PowerPoint presentations with formatting and alignment—rather than higher-value work.

“Our core focus area right now, as I mentioned before, is really starting to peel away all of this mundane wrote work that probably analysts spend 60 to 70% at least of their time doing. Right. you know, rebuilding the model from scratch every single time and relinking all these cells, making sure that you have all of the circular references, you know, all checked out and error-free and again all of the PowerPoint creation as well, which I think mostly is just about formatting and translation and making sure that the font sizes and the colors really match, right?”

Coding skills translate well to financial services because both domains require understanding logic, parsing data, and executing things in a structured and logical way—capabilities trained into Claude as a foundational model.

“You know, the world knows that we're fantastic at coding. 0.5% of the world are software developers, but I think coding is a fantastic starting point for us to think about how to solve some of these harder problems, right? Coding is so foundational to every single company out there, right? And these are complex systems where we really have to understand logic, parse data, and do things in a structured and logical way. So, A lot of what we trained into claude as a model we believe can also be really foundational in solving some of these harder problems across other industries as well.”

Coding is fantastic as a starting point for solving harder problems because coding is so foundational to every company and requires understanding logic, parsing data, and doing things in a structured and logical way, which are skills that transfer to other complex domains like finance.

“You know, the world knows that we're fantastic at coding. 0.5% of the world are software developers, but I think coding is a fantastic starting point for us to think about how to solve some of these harder problems, right? Coding is so foundational to every single company out there, right? And these are complex systems where we really have to understand logic, parse data, and do things in a structured and logical way. So, A lot of what we trained into claude as a model we believe can also be really foundational in solving some of these harder problems across other industries as well.”

Anthropic's goal is for Claude to be the backbone for the financial services industry regardless of how enterprise customers want to adopt, with the expectation that enterprise is not winner-take-all.

“My goal is for claw to just be the backbone for the financial services industry regardless of how our enterprise customers want to adopt. Yeah. Right. As we all know, enterprise is not winner take all.”

Finance shares similar characteristics to code in that it involves very complex systems, operates in regulated industries, requires understanding logic and audit trails, and demands accuracy.

“finance is probably the first vertical that we've really started tackling from a top down approach. So research product and go to market. We're excited about it because again finance is 10% of the world's GDP and I think it shares a lot of similar characteristics to code right very complex systems in regulated industries where understanding the logic having audit trails is really important to be able to trust these systems right and accuracy is ultimately extremely important”

55% performance on FinBench might seem low—if Claude gets the diluted share count wrong 40% of the time, that's not very useful—but these tools are the worst they will ever be, and Anthropic has only worked on them for 3 months.

“I think on one hand you might listen this way oh 55% is not very good if I'm going to ask you know claude to go find the like fully delu share count for a public company and 40% of the time it's wrong like how useful is that but the flip side is these are the worst the tools have ever been or will ever be and you've really only been working on improving them for three months.”

BCI's managing directors now interact directly with Claude-generated artifacts for comps analysis instead of requesting analysts rerun calculations and create updated spreadsheets, changing the workflow between MDs and analysts.

“BCI's integrated against S&P and fact and built live artifact landscapes, yeah, for themselves, for their managing directors to interact with even their MDs are talking to these artifacts on a daily basis instead of having to ask analysts to rerun calculations.”

Anthropic demoed their first AI financial analyst in July that built a fully auditable discounted cash flow model live on stage in two minutes without needing to be prompted to create an investment memo with proper citations.

“In July, Anthropic demoed their first ever AI financial analyst. In two minutes, it built a fully auditable discounted cash flow model live on stage. And here's where it gets powerful without needing to be prompted. She asked Claw to create an investment memo, all properly cited.”

Anthropic wants to close the loop across the entire value chain so that agents can be fully functional and autonomous co-workers within enterprise customers.

“we really want to start closing the loop across this entire value chain and really make sure that our agents can be fully functional autonomous um sort of co-workers within our enterprise customers.”

BCI has a strong top-down motion encouraging experimentation and adoption, with their main champion Ben creating sophisticated prompts for teams to follow, suggesting that driving adoption requires both top-down leadership and bottom-up user design.

“And I think one thing that's really encouraging for folks like BCI is that there's a strong top- down motion as well that encourages experimentation and adoption. So, you know, Ben, our main champion at BCI, spends a lot of time just creating these really interesting, sophisticated prompts for their team to follow. So, I'd really encourage all of my enterprise customers to think about how to drive that bottom up adoption.”

Success with public market use cases comes partly from Norges having built a mature data lake and ingesting core data sources, which is necessary for coherent public equities investment strategy.

“I think Nor has also done a really good job of they spend a lot of time making sure their data lake is to your point up to par and ingesting all of the core data sources. That's where of course on the public side you need to have a cohesive strategy around”

Design partnerships with customers don't need to be programmatic—regular touchpoints (e.g., weekly standups with BCI) suffice for sharing ideas and getting validation, allowing discovery of use cases customers find that founders never anticipated.

“I think there's a misconception that design partnerships need to be very programmatic. It doesn't really need to be. Yeah. As long as you have regular touch points with your customers. So I have a weekly standup with BCI for example where I share all of these ideas in my head and it's really important for us to just get validation from you. Yeah. I think the beauty of these AI systems is that it is not deterministic right you can have a set of hypotheses of the problems you're hoping to solve but your customers might find completely different use cases for these systems. So I think getting in front of your customers as early as possible to really share your ideas and getting feedback even with designs, even with mocks and prototypes is something that I would really encourage.”

The research-product-customer flywheel is Anthropic's internal model: research creates model capabilities, products expose those capabilities through tools and integrations, customers use and provide feedback, and that feedback informs research priorities for the next iteration.

“there's a term we use internally quite a lot, the research product and customer flywheel... My belief is that the product features that we build give our customers access to these model capabilities. Not everyone is a developer... you have to build tools and integrations and surfaces to make these model capabilities actually useful for you. So we want to do that... we bring all of this feedback back to our research teams.”

Anthropic is fundamentally a research lab focused on deploying models as safely as possible to solve the most complex and hardest problems where getting things wrong has real consequences.

“We're fundamentally a research lab that's really focused on deploying our models as safely as possible to solve what I would say are the most complex and hardest problems where I think getting things wrong have real consequences. That's why safety really matters, right?”

Anthropic does not currently build vertical-specific models; instead, they maintain belief that cross-learning across domains is valuable and that being excellent at code translates to being excellent at finance and other domains, so a general horizontal model layer is more valuable than vertical specialization.

“But ultimately, you know, I also get this question a lot. Are you all building vertical specific models? Mhm. Right now we're not. Our very firm belief is that there's a lot of cross learning across all these different domains and you know being great at code translates really nicely to being great at finance and you know these analytical capabilities in finance also translate nicely to consulting and other functions as well. So I think this set of flexibility of being a full horizontal layer is really important for us.”

Downstream from research, financial analysts must put information into spreadsheets to build discounted cash flow models, perform sensitivity analyses, and present findings in investment memos or pitch decks, but these downstream steps aren't captured by current benchmarks.

“Yeah, you have to put it into a spreadsheet to build a discounted cash flow model, do a lot more sensitivity analyses, and then present your findings in the form of a pitch deck or investment memo to be shared with others. Right? I think all of the steps downstream there isn't really a good benchmark to capture performance there.”

FinBench Agent benchmark, while useful as a starting point, only captures retrieval and research tasks (the least interesting but foundational bucket of work) and does not measure analytical or creation capabilities that are more valuable in real-world financial work.

“FinBench agent really only covers the first bucket, which is the retrieve and research research... It's asking, you know, entry-level finance analyst questions like, you know, what does [earnings data] look like for Apple and how did that grow over the past 10 years... very much retrieve the data, do a little bit of like, you know, funging with it, but it's mostly just pure research as opposed to”

Claude's Excel agent can solve complex optimization problems like backing out revenue targets: when asked to increase revenue growth from 15% to 18%, Claude can intelligently manipulate multiple interrelated inputs (like store count and same-store sales growth for Chipotle) to achieve the target by holding less-controllable variables constant and adjusting more-controllable ones.

“You know, one of the things that I frequently do in finance is having the models backs off what an outcome needs to be, right? So say that I need to increase, you know, revenue growth rate from 15% to 18%. Yeah, revenue is and that sounds simple, but like imagine a single cell in Excel. Well, it might actually the trail might go to like 15 or 20. So it's a pretty complex optimization problem to understand. how all these things tie together. So anyways, exactly right. So very simple example, revenue growth is broken down by the number of stores that you have. This is for say Chipotle and the same store sales growth, right? Even within that example, Claude needs to manipulate these two inputs so that the end output is 18%. It did exactly that. It held same store growth constant so that it can vary what's probably more within control for Chipotle which is the number of restaurants and it backs off to get to 18%.”

Anthropic focuses on building domain-specific language expertise in the model intelligence layer, similar to how coding has specific programming languages, while finance has accounting and other domain-specific frameworks.

“you have a domain specific language you know for coding you have coding languages for here you have accounting and a bunch of other specific”

Within six months of MCP's existence, major financial data providers (S&P, Faxet, Pitchbook) have published functional MCPs and are receiving strong customer feedback, which is remarkable given that getting even an API from these organizations historically took many years.

“MCP as a concept has existed for six months and even within these six months major players like S&P, Faxet, Pitchbook have published functional working MCPs that are you know getting really good feedback from our customers. That is remarkable to think about how many of these like just getting an API period from some of these organizations took many many years and then they had like an XML API and getting that upgrade to a modern like restful JSON hasn't happened for some of these people and then MCP servers come along and in six months”

BCI became excited about Anthropic's platform primarily because of its flexibility—the ability to connect different integrations and MCP servers and tailor specific workflows for different asset classes (privates vs publics).

“What really got BCI excited is the flexibility of our platform, right? We're able to connect to a number of different integrations and MCP servers. We're able to tailor specific workflows for privates versus publics.”

Downstream from retrieval and analysis agents, the next wave of progress is analytical agents, spreadsheet agents, and PowerPoint agents that can close the loop across the entire value chain and function as fully autonomous co-workers within enterprises.

“I think research and retrieval agents have been the most mature in the market and has obviously seen great product market fit. But downstream from that, analytical agents, spreadsheet agents, PowerPoint agents very early. Yeah, we really want to start closing the loop across this entire value chain and really make sure that our agents can be fully functional autonomous um sort of co-workers within our enterprise customers.”

While 55% accuracy on FinBench is not excellent, it represents the best current performance and only three months of focused work, meaning these are 'the worst the tools have ever been or will ever be' and substantial improvement is ahead.

“I think on one hand you might listen this way oh 55% is not very good... but the flip side is these are the worst the tools have ever been or will ever be and you've really only been working on improving them for three months.”

Flexibility and basic model capabilities (1M token context window, web search) are sufficient to derive value from Claude without requiring extensive MCP integrations; manual uploads and public data sources enable utility even for early adopters.

“I don't actually think a lot of integration work really needs to go in to derive value from these systems because they're so flexible and so powerful as long as you know what problems you're hoping to solve, right? Even without MCP integrations, these systems are great at looking at public data sources... We have a 1 million token context window which is some of the largest in the market as well. So a lot of these core foundational capabilities in the model are extremely useful for our customers.”

Public benchmarks serve as a reference but are insufficient; enterprises should develop their own internal benchmarks based on their specific problems and use cases rather than relying solely on published academic benchmarks.

“I think it's also important to think about what benchmarks really serve in terms of their purpose, right? you see a lot of publish published benchmarks that are just more academic representations of what these things could do in theory. I would always encourage all of my enterprise customers to use those benchmarks as reference, but really think a lot more about what problems they're hoping to solve internally in developing their own versions.”

Financial analysts should focus on high-value activities: understanding markets and business models, and spending time with founders and portfolio companies to build deep understanding—activities that are passion drivers and competitive differentiators.

“What should analysts really think about and focus on? And what are they passionate about? They're passionate about understanding markets. They're passionate about understanding business models. And they're passionate about spending time with founders and their investy companies to understand those business models much more. So we want to really start freeing up the time to do all of those things that actually really matter.”

Norges is distinguished by being very technical with technical champions who are builders, and they have constructed their own internal Snowflake MCP server even before Snowflake announced their own MCP, allowing portfolio managers to query approximately 9,000 different portfolio company information daily.

“the beauty of Nores is that they're really technical and they have a technical set of champions who are builders within our ecosystem which is fantastic to see... What Noris has done is that they've constructed their own internal workflows. For example, they built their own Snowflake MCP even before Snowflake has announced their MCP server. And I believe daily today they're having their portfolio managers query probably 9,000 different portfolio company information.”

Anthropic believes it is doing well on accuracy and auditability, but data security (compliance processes) still takes 6-12+ months of enterprise sales cycles, slowing widespread deployment.

“I think we think a lot about all three of those components as a safety research lab. And I would say we're getting pretty good at two and three, but number one, it still takes time, you know, to go through these compliance processes, you know, rip out your existing solutions and just think about that usual 6 to 12 month plus enterprise sales cycle.”

Anthropic views its approach to AI as fundamentally open-ecosystem focused, with MCP as an open-source protocol and active encouragement of partners like S&P, Faxet, and Pitchbook to publish their own functional MCPs.

“You know, I think the really interesting thing about our approach to AI is that we want to foster and build an ecosystem around anthropic. So everything we do is open, right? MCP as a concept is an open source protocol. We really want to encourage the world to think about how to connect systems to AI, right? Because of that, MCPs can be built in a few different ways.”

For all deployed customers, Anthropic has dedicated teams of customer success managers and applied AI specialists who support daily workflows as standard deployment and support model.

“So for all of our deployed customers, we have four teams of customer success managers and applied AI that really supports their daily workflows. So that really just, you know, is our standard deployment and support model for all of our enterprise customers.”

Anthropic is rolling out their financial analyst to BCI of Canada (managing $200 billion) and Norges of Norway (managing over $2 trillion, the largest sovereign wealth fund in the world).

“And now they're rolling out the financial analyst to two of the world's largest sovereign wealth funds. BCI of Canada, which manages over 200 billion, and Norges of Norway, which manages over two trillion, making them, in fact, the largest sovereign wealth fund in the world.”

Anthropic is also rolling out Claude to all 450,000 of Deloitte's global employees.

“In October, Anthropic also announced that they're rolling out Claude to all 450,000 of Deoid's global employees.”

Nick Lynn worked for most of his career in finance and fintech before joining Anthropic, spending approximately 75% of his time doing manual data analysis, PowerPoint creation, and formatting tasks rather than substantive financial work.

“Yeah, we didn't really cover this, but I am what I like to call recovering investment banker and private equity investor. So before anthropic, I spent my entire career in finance and in fintech. So, a lot of these problems we're hoping to solve are just so near and dear to my heart because I spent, you know, probably 75% of my time just doing this manual data analysis, you know, PowerPoint creation, making sure that the text boxes really match the same exact shade of blue, right?”

BCI (Canadian Sovereign Wealth Fund) has been a key design partner for Anthropic, closely partnering for months to validate problem identification and solution approaches.

“Yeah, for sure. BCI has been a fantastic what we call design partner for us. As I mentioned before, everything we do, we cannot exist without very close partnership with our enterprise customers. So we share ideas as early as possible with folks like BCI so that we get feedback and validation on hey these are the problems we're solving and this is the right approach to solve it. So they've been really closely partnered with us for quite a few months now.”

BCI is attractive as a design partner because it manages different investment strategies with slightly different requirements that are hard to satisfy with generic AI solutions, and the platform's flexibility in connecting integrations and MCP servers allows them to tailor workflows for different asset classes.

“I would say what's really interesting about BCI is that we know they're not the largest sol right in a UN but they move really nimly. So they've also started to recognize a lot of problems internally where they manage a large number of different strategies. So every single team has slightly different requirements that's really hard to satisfy with any generic AI solution. I think what really got BCI excited is the flexibility of our platform, right? We're able to connect to a number of different integrations and MCP servers. We're able to tailor specific workflows for privates versus publics.”

BCI has $200 billion in assets but only 200 employees, meaning they have a high data-to-staff ratio and need AI tooling like MCP servers to analyze investment data spread across the world.

“When I think about a sovereign wealth fund like BCI being an interesting design partner, one thing I immediately think of is, you know, they have $200 billion in assets, but they only have 200 people. So it's not a lot of people, but they actually have a tremendous amount of data they need to analyze for all of their investments, which are spread around the world. And so you need that tooling to your point, you need to have all these MCP servers, etc. to actually get some of the analysis.”

Integration work required to use Claude is minimal because the models are flexible and powerful with public data access, but Anthropic wants to extend capabilities with MCP integrations to access internal data and APIs.

“Yeah, for sure. I don't actually think a lot of integration work really needs to go in to derive value from these systems because they're so flexible and so powerful as long as you know what problems you're hoping to solve, right? Even without MCP integrations, these systems are great at looking at public data sources. Yeah. Right. And identifying trends and you can upload a lot of things into cloud as well. We have a 1 million token context window which is some of the largest in the market as well.”

Anthropic is not just a chat application, set of APIs, or code builder—it is a full platform with product services, API, cloud code, and application service layers to allow customers to adopt based on their specific needs.

“Ultimately Anthropic we're not just a chat application. We're not just a set of APIs. We're not just cloud code. I think the beauty of working with Anthropic is that we have all of these product services for you to really adopt based on your needs. Yeah. Right. And our API and our cloud code makes it really flexible to build your own internal solutions as well. Whereas our application service makes it super easy to adopt with very little integration effort.”

Claude.ai has a feature called 'artifacts' that allows Claude to connect to information and use its coding abilities to render live code in real time to display information, enabling dynamic dashboards instead of static Excel sheets.

“There's a feature within claw.ai called artifacts. Artifacts are essentially a way to connect to information and allow claw to use what it's the best at, which is coding. Claude uses code, renders code in real time to display a ton of information. So instead of using these static Excel sheets as comm sheets, BCI's integrated against S&P and fact and built live artifact landscapes,”

Anthropic's mission is fundamentally about deploying models as safely as possible to solve the most complex and hardest problems where getting things wrong has real consequences, which is why safety matters.

“We're fundamentally a research lab that's really focused on deploying our models as safely as possible to solve what I would say are the most complex and hardest problems where I think getting things wrong have real consequences. That's why safety really matters, right?”

Claude tweaks approximately 18 inputs in the backend of a spreadsheet model to achieve a target outcome, whereas an analyst manually ripping apart an Excel model would find it extremely time-consuming without proper architecture.

“So you see like this one little thing in the cell change, but really it's tweaked like 18 inputs in the back end to figure out exactly what that looks like, which if you're an analyst and you're like ripping apart an Excel model and you're trying to do that and it's not already set up in the right architecture, you're like, 'Oh man, this is going to take a long time.'”

Everything starts with research and data gathering, which flows downstream to qualitative and quantitative analysis, and finally to creating shareable outputs.

“Right. Everything starts with the research that we have to do and the data we have to gather. Yeah. Downstream from that, we do qualitative, quantitative analysis on that data. And ideally, we create outputs that can be shared with others in the form of word documents, spreadsheets, and powerpoint documents as well.”

The retrieve phase (research and data gathering) is the most mature in the financial AI market and has achieved product-market fit, while analyze (spreadsheets, models) and create (presentations) phases are very early and represent substantial opportunity.

“I think research and retrieval agents have been the most mature in the market and has obviously seen great product market fit. But downstream from that, analytical agents, spreadsheet agents, PowerPoint agents very early. Yeah, we really want to start closing the loop across this entire value chain”

Approximately six months ago, research found that about 1% of bankers had access to AI tools at work, and Anthropic estimates today (present conversation) that access remains in the single digits—indicating early-stage adoption despite accelerating deployment.

“I researched what percentage of bankers have access to AI tools at work and the nearest answer I could get is 1%... Yeah... Where do you think we are today... You know, I would probably still say it is in the single digits. Yeah.”

Norges built their own Snowflake MCP before Snowflake announced their own MCP server, demonstrating their technical sophistication and ability to construct custom integrations.

“What Noris has done is that they've constructed their own internal workflows. For example, they built their own Snowflake MCP even before Snowflake has announced their MCP server.”

Norges' portfolio managers query approximately 9,000 different portfolio company information points daily using their Snowflake MCP integration with Claude.

“And I believe daily today they're having their portfolio managers query probably 9,000 different portfolio company information.”

Norges has a much larger organization than BCI (approximately 2,000 people with at least 100+ engineers), a larger public market team with more technical quants and traders, and uses their engineering resources to build customized workflows on top of Claude.

“It is interesting thing about BCI about 200 employees. I don't know the size of their engineering team but I imagine it can't be more than 20. It's probably more like 10 or something whereas NO is 2,000 people. They probably have a at least a hundred engineers somewhere in there, right? That's a sizable team for you guys to work and collaborate with and learn from. And I think Nores's public uh team is also a lot larger. So more technical quants and traders as well who are really empowered to build into their own workflows.”

Six months ago, approximately 1% of bankers had access to AI tools at work, and Lynn estimates this is still in single digits today, indicating that financial services AI adoption is still at very beginning stages despite recent deployments.

“something else I want to touch on, I guess, with Deote too is thinking about I six months ago I researched what percentage of bankers have access to AI tools at work and the nearest answer I could get is 1%. About six months ago. Yeah. Where do you think we are today in terms of number of bankers, consultants, you know, large enterprise companies that have access to AI? You know, I would probably still say it is in the single digits. Yeah. Yeah,”

Research and retrieval agents have achieved the most maturity in the market and have clear product-market fit, while downstream analytical agents, spreadsheet agents, and PowerPoint agents are very early.

“Yeah, I think I would probably go back to the three verbs that we talked about, right? I think research and retrieval agents have been the most mature in the market and has obviously seen great product market fit. But downstream from that, analytical agents, spreadsheet agents, PowerPoint agents very early.”

Deloitte was one of Anthropic's first Claude for Enterprise customers, signed approximately a year and a half ago, and consulting firms are interesting as customers because they have multiple business arms with disparate needs (integration, management consulting, implementation consulting, accounting).

“Yeah, for sure. So, Deoid was actually one of the first claw for enterprise customers we signed maybe about a year and a half ago already and you know consulting firms are really interesting because one they have multiple arms right? Yeah. They have integration arm, they have management consulting, they have four deployed engineers. So um I think their needs are quite disparit across the whole organization.”

Financial services industry needs Anthropic's approach because it combines model capability improvements, product/UI innovation, and ecosystem partnership across three sustained dimensions rather than one-time product launches.

“I think all three of those components we're continuing to work on and I'm excited about what's next”

The FinBench Agent benchmark, published by VoxAI, covers only the first and least interesting but foundational bucket (retrieve and research), asking entry-level finance analyst questions like basic data retrieval from SEC filings.

“So, you know, the one benchmark that's picked up a lot of steam in the industry, published by our good friends at vows.ai is called finest agent. Finest agent really only covers the first bucket, which is the retrieve and research research. Yeah. Right. It's asking, you know, entry-level finance analyst questions like, you know, what does just look like for Apple and how did that grow over the past 10 years? I pulled a few sample questions. So, like, what is the total number of comto shares repurchased by Netflix? what is the percent of revenue that AWS derived in each year in the three-year category?”

Norges has approximately 2,000 employees, which is roughly 10x larger than BCI, and therefore likely has a larger technical team and more public market quants and traders who can build custom workflows.

“it is interesting thing about BCI about 200 employees. I don't know the size of their engineering team but I imagine it can't be more than 20. It's probably more like 10 or something whereas NO is 2,000 people. They probably have a at least a hundred engineers somewhere in there, right? That's a sizable team for you guys to work and collaborate with and learn from.”

For research and retrieval use cases, Norges' public team gets the most use, mostly from public datasets sometimes augmented by third parties; privates use cases come next, then deeper analysis work.

“And if you think that first bud get the research like definitely like the NOR's public team is getting the most use and those are mostly public data set sometimes augmented by third parties but then it's like okay the next is probably the privates within that and then you start to really get deep into the analyze thing.”