Inside Anthropic’s Wall Street Strategy
An interview with Jonathan Pelosi, Head of Financial Services at Anthropic
OpenAI rival Anthropic has focused on selling its models to large enterprise customers. In July, the company launched Claude for Financial Services, a domain specific platform built for regulated finance and run by its frontier language models.
Early users span hedge funds, insurers, and sovereign wealth funds. Bridgewater has used Claude to help researchers query internal documents and data. AIG has applied it to underwriting and risk analysis. Norway’s sovereign wealth fund, NBIM, uses it to work through policy and investment material at scale.
I was happy to speak with Jonathan Pelosi, who leads Anthropic’s financial services effort, about how firms are actually using the product. Here is what readers will learn from the conversation.
How Anthropic is tailoring large language models for regulated financial workflows
What Claude for Financial Services is designed to do in day-to-day finance tasks
How skills and Model Context Protocol connect models to firm-specific workflows and data
Why confidence in using LLMs inside financial institutions has increased
Where adoption may expand next, and what is slowing it down
This interview has been edited for length and clarity.
Matt: A lot of people in markets are asking, “What does this AI mean for me?” Everyone keeps talking about AI. It feels like we should get used to it.
Jonathan: I almost liken it to the internet. We talk to the internet so much you do not really refer to it that way anymore. It is more like a layer of technology the world uses. Increasingly it will not be “AI here” or “AI there.” It will just be technology people use.
Matt: Tell me about Claude for Financial Services.
Jonathan: The idea is: how have we taught Claude to be specifically good at financial tasks versus things like generating images or videos. We dedicate resources to enterprise workloads, with financial workflows as a primary vertical. Practically, that means it is very good at building discounted cash flows, financial analysis, or building a leveraged buyout model.
One step further is determinism. In financial services, people do not have the luxury of inconsistent outputs. If you and I have a specific LBO model we use at our firm, you do not want Claude making a different one every time. You build a skill, like a deal teaser skill. It is basically a set of prompts and files packaged together. Now every time it takes that action, it produces a deterministic output. It does it the same way every time.
A practical example: if I am JPMorgan and I am creating PowerPoint presentations, I do not want Claude guessing what the presentation should look like. I upload brand guidelines as a skill. Every time it generates a deck, it uses the color schemes, font, logo, and format. That is helpful because you do not want it to keep inventing formats.
In a regulated world, you want the power of a large language model and deterministic outputs that matter in your workflow.
Matt: Is that under the umbrella of Model Context Protocol?
Jonathan: Not exactly. The next layer of the pyramid is MCP. Think of that as data connections. You can crawl the internet and build internal knowledge spaces through projects, but you might want to query FactSet data, PitchBook data, or other third-party services and data providers. Through MCP, Model Context Protocol, we develop native connections, so it can plug into those.
Now when you ask Claude a question, it can crawl Snowflake data, crawl FactSet data, and cite it in a way that provides context and citations. It shows exactly where it is pulling from, for example on an income statement housed in FactSet.
What it generates is a PowerPoint file following the template you want. You can download it and open it, and it follows the specific design template.
You can give it specific instructions: “Screen this deal teaser against our investment criteria from SharePoint, flag any risks, build a quick leveraged buyout model assuming the following, and do that with the leveraged buyout skill.”
It reads the skill and matches the information. It builds a fully functional model from scratch, with formulas and links.
We also launched Claude for Excel as a plugin. You can use Claude within Excel files. You can prompt: “Add a few rows below the upside case and downside that is 10 times lower than the base case,” and it will build within Excel. You can add a sensitivity table, and it will generate it in Excel.
It also creates investment memos. That uses Claude’s artifacts feature and can be downloaded as a PDF.
This shows all three layers: the power of the model, skills and capabilities tailored to financial tasks, and MCP connecting to the data sources customers rely on. Everyone needs context, whether third-party datasets or internal repositories. Instead of querying public 10-Ks, you can query an internal database just as easily if you connect it.
Matt: I am curious about how you take an LLM, which is probabilistic, and then MCP is retrieving from deterministic structured sources. In Excel, it is that intersection between the two.
Jonathan: There are things where you do not want a probabilistic output. Excel is a great example. Formulas have to be exact.
What we found is that because Claude is so good at code, it is also very good at knowing when to write code like Python, which works well with Excel. It knows how to use tools thoughtfully. Instead of doing a calculation loosely, it can call a calculator tool and put the numbers in so there is no gray area. That calculator tool is deterministic.
That bridging between the model’s horsepower and deterministic outputs has become a force. Historically you would not have confidence in it. Now you do, and it can cite everything it pulls. It might say it pulled something from page four, paragraph three, and you can double check.
Matt: A year ago people complained LLMs couldn’t do math but now they can connect to calculators. The issue is getting them to talk to each other. So you want it to predict the token to do the tool call.
Jonathan: Exactly. It is very good at knowing which tools to call.
Skills go one step further. When we interview individual contributors, analysts, junior bankers, researchers, they have specific workflows that need to be done a specific way.
You can build custom skills at the employee level, the team level, and the org level. Org level can be something like: every output has to follow company values and brand guidelines. There is comfort knowing whatever email or messaging is generated adheres to guidelines.
These Lego blocks let a company take datasets via MCP and skills, and move them around in a way that becomes “Claude for Bridgewater” versus a broad platform not tailored to their needs.
Matt: I am curious how conversations have evolved in the last year. It seems like the technology is moving faster than the industry can scale it. How have the conversations changed?
Jonathan: Confidence has gone up, and for good reason.
People were skeptical because they used it personally and got burned. A year ago, someone might use ChatGPT on their phone for a mortgage calculator and it would be wrong. If that happens, they shelve it and do not take it seriously for work.
We are as much in change management as in technology. It is reminding and showing people how fast it has improved to rebuild confidence. Sometimes you need to show someone 50 workflows working correctly for them to trust it.
A metaphor is driverless cars. Even if you publish research showing it is as safe as a human, people are reluctant. It is only when it starts hitting something like 10 to 20 times safer that a tipping point happens. In San Francisco, Waymo is everywhere now.
It is similar here. People got burned once, they see a couple wins, still reluctant. Now they are seeing it knock it out of the park repeatedly, and that is where we are. People are giving it their undivided attention.
Matt: It reminds me of cloud computing. No one wanted to put data in the cloud. And that’s obviously changed. Workflow-wise and enterprise-wise, where do you see the next 12 months?
Jonathan: The biggest bottleneck will be enterprise bureaucracy, legacy systems, and change management. It is not the technology.
Someone gave a great analogy: it is like if you have a Ferrari, but you live in a school zone with a 20 mile an hour speed limit and you cannot leave. You cannot do much with it.
In the next 12 months, adoption increases dramatically. I do not see a large retail bank running purely on AI. I see humans moving up the value chain. Customer service is an example. A high percentage of airline customer service is now done through text and bots, but there is still a big customer service team for higher-value, more complicated queries.
In finance, a team might have needed to hire five more analysts to evaluate more deals. Now the existing team can do more. Headcount asks slow down, and people become more productive.
At the same time, some roles are more susceptible.
Matt: There is a lot of white-collar manual labor. You are pointing and clicking at things. Are you looking at forward deployed engineers, teams that sit with customers and help build real applications?
Jonathan: For some of the highest-priority enterprise customers, we have teams to help them build and bring applications to life when they do not have the people, speed, or time to do it themselves. Forward deployed engineers is the term most people use.
The earliest adopters are startups and digital natives. We encourage large banks and insurers to look to those firms for inspiration, while recognizing they are regulated and cannot do everything a tech startup can do. You can look to a digital-first insurer or bank for signal on what is possible because they are better equipped to adopt it.
Matt: Anything you think is underhyped?
Jonathan: Skills are dramatically underhyped. We launched it under the radar, but the implications are big.
One of the biggest barriers to widespread adoption in financial services is the ability to make workflows deterministic. Skills fills that gap. MCP has gotten a lot of attention. Skills will get more attention because it lets institutions take workflows they have honed over decades and combine them with the intellectual horsepower of large language models. That brings adoption to life faster with practical use cases.
Matt: I was talking with Man Group. The gist was they can take the quant process end to end and do more, so they have more ideas and can test more strategies in parallel.
Jonathan: Exactly. You can develop, test, and run strategies in parallel. You start to have a virtual fleet of quants. The individual becomes the orchestrator, aiming those agents in the right direction.