AI Agents Help Investors Parse Bank Results and Track Market Signals

Four AI agents helped one investor work through the latest results from the six largest US banks, a mass of information that included more than 250 pages of financial material and over eight hours of management calls. The tools summarized the quarter, tracked themes across companies, anticipated a likely analyst question and updated a financial model. Their value, as described here, lies in speeding up analysis and organizing evidence; the investment judgment still belongs to the person using them.

The workload will be familiar to anyone who has covered a busy reporting period. In the author’s former role as a research analyst, a company could publish results at 6am, leaving little time to read a document that might run to 75 pages, brief colleagues and institutional clients, revise a model, join a call, ask questions and publish analysis. Several companies might report on the same day, with the cycle starting again the following morning.

That experience also shaped the author’s later work managing money. Familiarity with company disclosures made it easier to separate the figures that mattered from the volume of supporting material. The timing can matter: although earnings release dates represent just 1.6% of trading days, they account for 15% of annual share-price variance, according to a linked analysis. Revisions to expectations can influence stock performance, and the market’s first reading of a company’s results may be incomplete. The author says that, on some occasions, heightened trading activity around announcements created an opportunity to act on their analysis before the market opened or during calls.

Four agents with distinct jobs

The author’s system uses four AI agents, each assigned a different part of the research process. Ernie handles the broad review. It works from a desktop folder for each bank, where the author keeps past results, presentations, call transcripts and the latest annual report. When new results arrive, Ernie compares them with analyst previews to identify where the quarter matches market expectations and where it diverges. After a transcript becomes available, it adds that material and produces a five-page report and a two-page brief.

Thelma follows themes across companies and over time. Its current topics include agentic disruption, AI financing, private credit and the effects of rising rates. For each subject, it collects what management says and compares those comments with the figures. That pairing matters: executives’ descriptions can indicate how a company sees a developing issue, while reported numbers offer another way to assess whether it is affecting performance.

A third agent, the Mike Mayo bot, prepares a possible question for a bank call based on the work of Mike Mayo, a Wells Fargo Securities analyst who has covered banks through more than 140 reporting seasons. The author describes Mayo’s direct questioning as a way to bring the most pressing current issues into focus. The bot’s role is to anticipate one line of inquiry, giving the analyst a prompt for further preparation.

The fourth agent updates the author’s model directly in Excel. The author previously tested AI-assisted modelling in Excel Forever and says the tool has improved since that experiment in February. The account does not specify which model inputs changed in this round, so the practical result is best understood as a faster update process rather than evidence that the model’s assumptions or conclusions are automatically reliable.

What the agents found in the bank results

The agents read the six banks’ results as stronger than the market’s apparent response suggested. Management guidance was broadly higher, particularly on net interest income. Equities trading and banking fees were generally strong enough to offset a softer period in fixed income trading. Consumers continued spending, while credit held up.

Those points offer a useful starting frame, but they do not settle how the banks will perform. Higher guidance on net interest income, for example, is one part of the picture; the underlying drivers and their durability matter too. Trading and banking fees can move in different directions across periods, while consumer spending and credit quality provide separate clues about customer activity and borrowers’ ability to meet obligations. The source does not provide bank-by-bank figures, so the cross-company summary should not be read as a detailed comparison of individual institutions.

The workflow’s main contribution is compression: it turns a large collection of documents and conversations into a set of comparisons, recurring themes and prompts for human review. That can free up time for questions that summaries alone cannot answer, such as whether a change is temporary, whether management’s explanation matches the figures and which assumptions deserve closer inspection. Speed helps most when the analyst can trace an agent’s conclusion back to its source material.

The author says paid subscribers can email for the instructions used to train the agents and for the outputs they produced. The broader takeaway from this account is that AI can organize a dense reporting cycle into manageable workstreams. The useful test is whether those workstreams make the evidence easier to examine, while leaving the interpretation open to challenge.