Bloomberg reported this week that OpenAI has 100+ bankers at $150 an hour to help train AI on how to build financial models. The goal: automate the tedious work done by junior bankers.

The group includes former employees from JPMorgan, Morgan Stanley, and Goldman Sachs. According to Bloomberg, they write prompts and build financial models for different transaction types, including restructurings and IPOs.

They’re instructed to build their models in Excel and to follow standard Wall Street conventions, from margin sizing to italicizing percentages. A reviewer checks each submission, and participants are required to fix any issues before the work is added to OpenAI’s systems.

There’s a common assumption that AI will lead to mass layoffs, one that is not backed up by data. This story is not about the machines taking investment banking jobs. It’s about codifying how deals are put together on Wall Street by having experts teach AI via reinforcement learning.

“Nobody writes financial documents better than highly trained analysts at investment banks,” said Raj Bakhru, general manager at BlueFlame AI, (now part of Datasite). He added that reinforcement learning from such data could make AI outputs “much more useful” for first drafts of models and pitch materials.

As for OpenAI’s motives, Bakhru said OpenAI appears more focused on enhancing its base models with banking-grade data than on launching a dedicated product for bankers.

Derek Snow, founder of Sov.ai, says that much of the knowledge these models need isn’t stored in public datasets but “in xls and a bit in IB brains.” Hiring bankers, he said, helps expose AI systems to the kind of proprietary, unstructured information that can’t be scraped or sourced from the open web.

OpenAI’s rivals are also building out their financial offerings. Anthropic’s Claude for Financial Services lets banks and asset managers use its LLM for deal modeling and compliance. Perplexity Finance offers live earnings-call transcripts and AI summaries of SEC filings.

Reinforcement learning — I’m very much oversimplifying — is basically a giant game of hot and cold. The data helps tell the model when it’s getting “warmer” or “colder” to the right answer. It doesn’t explain why it’s the right answer.

So while AI companies like to market their models as “thinking,” they’re not really reasoning in the syllogistic sense. (All men are mortal; Socrates is a man; therefore Socrates is mortal). These models are mimicking thinking, not doing it.

AI is coming for investment banking tasks. The tech is starting to automate the tedious point-and-click white-collar work that has historically taken junior analysts 100+ hours a week.

Last week I mentioned Citi’s AI Outlook report, which included a surprising use case: using the tech to test your investment idea, kind of like a contrarian analyst for hire. And this week, Bloomberg reported that Viking Global Investors has its own GPT and is using it for exactly that, among other use cases.

You know those weird company names on your credit card statement? You’ll eat at a restaurant called Piccola Cucina, but the charge shows up as Realto Group. You’re about to call the bank to report fraud, but you google and see that Piccola Cucina = Realto Group.

Now imagine being JPMorgan. The bank processes 50 million transactions a day and needs to determine, within milliseconds, which are legitimate and which are fraudulent.

Historically, rule-based systems handled about 80% of those transactions. Programmers had to write thousands of hard-coded rules: when you see x, do y. That worked for known merchants but broke down with new or messy names.

Large language models flip that approach: they learn patterns from data instead of following hand-coded instructions. But running a massive general-purpose model like GPT-4 would be too slow and expensive for real-time banking systems.

So JPMorgan researchers built a small decoder-only Transformer—the same architecture behind ChatGPT, but scaled down to 1.7 million parameters and trained only on transaction data, according to new research.

By tailoring the model to messy payment text, the proprietary Transformer raised transaction coverage from 80% to 94%, correctly identifying 14% more transactions the old system missed. That improvement reduced false fraud alerts and saved about $13 million annually.

The same approach could apply to other domains with messy, structured text—normalizing names or metadata across news, filings, social media, or trading records. For instance, it could reconcile inconsistent entity names (“AMZN,” “Amazon.com Inc,” “Amazon.com Services LLC”) or standardize broker, venue, and trade-type data.

Small language models (“Small” is relative; there’s no agreed definition) can improve accuracy if its training data is constrained on a specific output. They show that targeted training can outperform larger models in domain-specific tasks.

Finster AI, a platform for investment banks and asset managers, raised $15 million across its Seed and Series A rounds. The Series A was led by FinTech Collective and the Seed was led by Peak XV, with ongoing participation from Hoxton Ventures.

The company will use the funding to expand in New York, grow its London HQ, and build new data and technology partnerships. Finster says it works with several Tier-1 global banks and asset managers.

Founded in 2023 by former Google DeepMind researcher Sid Jayakumar, Finster builds AI tools designed for regulated finance, emphasizing data governance, accuracy, and MNPI security. Its platform tailors models and workflows for research, trading, and client engagement.

The company has recently added banking veterans, Chris Andrews, former Global COO of Research at Morgan Stanley, joins as COO in New York; and Veeral Manek, previously GM of Wealth & Trading at Revolut.

I spoke with Sid last year in an AI Street 5 Minutes With interview where much of the conversation centered on trustworthy AI.

Finster’s raise shows that investors continue to back specialized AI—especially platforms solving compliance and data-trust challenges.

Advertise on AI Street to reach a highly engaged audience of decision-makers at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email me ([email protected]) for more details.

The following is from my Sunday Edition, where I cover practical use cases of AI. If you’d like to sign up for that issue, sign up here.

AI may eventually surpass the average Wall Street analyst, but right now general models like ChatGPT perform investing tasks poorly.

A third-party benchmark shows how far they need to go to match an average Wall Street analyst.

Vals AI, a San Francisco startup that evaluates AI models across industries ranging from healthcare to finance, found that the best models today score around 55 percent accuracy on a test designed to mimic the work of an entry-level financial analyst. Most don’t even reach 50 percent.

The benchmark measures how well AI agents handle real-world finance tasks. Vals AI collaborated with investment bankers, hedge fund analysts, and professionals from a major global bank to help create the 537 test questions.

They cover nine different types of financial tasks - everything from simple information lookup to complex financial modeling. Each question was reviewed for accuracy.

AI struggles with multi-step, complex tasks, but performs far more reliably on narrow, single-step use cases. So prompts should focus the model on one clear action at a time, not broad requests like “What is your analysis of X?”

*Recently added events in bold