Wharton came out with a survey this week on AI adoption that pushes back on a couple of narratives that have emerged recently :1) businesses are struggling to get ROI out of AI and 2) the technology is shrinking opportunities for recent college graduates. The business school and GBK Collective surveyed about 800 “senior decision makers” across industries including banking and finance.

This study challenges an MIT survey that said about 95% of corporate GenAI pilots fail to deliver measurable business results. JPMorgan’s Dimon said earlier this month AI is already paying for itself. Other studies this week point to measurable benefits: AI adoption among financial services compliance professionals jumped 26 points to 71%, according to a new industry survey.

It’s only been three years (!) since ChatGPT went mainstream and AI has quickly permeated the corporate world and adoption is accelerating. Many business tasks don’t demand critical thinking, a skill AI (still) lacks.

JPMorgan and Citigroup are letting employees use AI to help write performance reviews—a task we’re all glad to give to the bots.

Anthropic, the AI startup backed by Amazon and Google, announced updates to its Financial Services offering, including a beta preview of an Excel plugin that lets users analyze and interact with their spreadsheets through its chatbot Claude.

LLMs, by themselves, perform poorly at math, and therefore are not well suited for Excel. That’s because language models are built to predict the next token, not perform calculations. (And also why most AI-for-finance startups have focused on chatbot-driven research.)

Anthropic has been working to connect these two worlds: LLM probability and deterministic code.

It released Model Context Protocol last year to help these different systems work together, a standard that’s been quickly adopted by competitors including OpenAI and Google.

Anthropic also rolled out what it calls “Agent Skills,” pre-built modules that handle finance tasks like building discounted cash flow models, running comps, analyzing earnings, and writing coverage reports across its apps.

Anthropic also announced new data integrations.

A year ago, lots of commentary focused on how LLMs were unreliable. Folks were using them as a Swiss army knife, rather than just another tool. Now, things have moved toward integrating a probabilistic system into deterministic workflows.

Buy, Sell, and Hold stock recommendations have been part of Wall Street research for decades. They caught on in part because they’re simple and easy to measure: you have a price target and a time horizon, and later you can check whether that call actually made money.

Macro forecasts are harder to track. It’s trickier to judge whether an analyst’s outlook on inflation was actually right—and if it was, whether it was right for the reasons the analyst gave.

In theory, you could track all this yourself, but it’s a manual, time-consuming task—one that large language models are well suited for because they excel at organizing messy data.

Menos AI, a San Jose startup, ranks analysts on both their calls and reasoning. Its Voice Scoring system uses large language models trained on financial data and agentic workflows to evaluate whether an analyst’s explanations proved correct, producing a “voice score.”

The company is piloting Voice Scoring with several global-macro hedge funds, per the company. Menos AI has raised $5.2 million in seed capital and was founded last year by William Wu, Chris Yang, and Xiang Pan.

AI can now extract and connect data at a cost that makes new indicators like Menos AI’s voice scores economically viable.

Researchers tracked day-to-day changes in macroeconomic sentiment in Wall Street Journal stories using AI to identify and categorize growth and inflation narratives to show that news-based sentiment mirrors key economic indicators, according to research from the Bank for International Settlements.

The paper uses OpenAI’s GPT-4.1 and GPT-5 mini models to analyze 25 years of Wall Street Journal coverage of the U.S. economy. By analyzing about 200,000 articles, the models generated daily “growth” and “inflation” sentiment scores and broke them down into demand- and supply-side drivers such as fiscal policy, commodity prices, and supply disruptions.

Researchers found that the LLM-derived sentiment indices closely track traditional indicators such as payrolls, inflation, and industrial production.

“The LLM-based approach bypasses statistical inference and directly draws on human-curated narratives,” the paper said, adding that the method “provides a real-time, high-frequency quantitative assessment of the state of the economy.”

The BIS, sometimes called the central bank of central banks, has released the resulting indices and plans to update them over time.

As for current economic conditions, the AI models suggest that recent articles point to slower growth, with more coverage of weak spending and tighter financial conditions, along with renewed inflation pressures from government actions and energy costs.

A big issue with AI forecasting is that we don't know what data went into training the model, which opens the door to look-ahead bias—testing it on data it already “knows.” So as the BIS releases more data, we’ll be able to see how the models perform on information it’s definitely not seen.

Authors: Byeungchun Kwon, Taejin Park, Phurichai Rungcharoenkitkul, and Frank Smets

LLMs can quantify economic sentiment by analyzing news narratives. The BIS study shows their results closely track traditional indicators like payrolls, inflation, and industrial production.

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.

Is there a conference I missed? Reach out: [email protected]