AI Analyst Chat

AI Analyst Chat is an agent that does company research for you. You ask a question, and rather than answering from what a language model already knows, it plans an efficient and validated series of steps to reach the answer, runs its own tools to gather the evidence, checks what came back, and keeps going until the question is resolved. It is an expert on the company you are analyzing, it cites every figure to the passage it came from, and it never makes a number up.

How the Agent Works

The chat is a tool-calling agent, and the model drives the work. It runs a loop: gather the evidence by calling tools, analyze what came back, and decide whether the question is resolved or another pass is needed. Nothing about that loop is a fixed pipeline. The agent chooses which tools to call and in what order based on what you asked, runs independent calls in parallel, and reads each result before it decides the next move. The steps surface in the thread as it works, so you can see which tool it reached for, what each one returned, and how it got to the answer.

The loop is built to be thorough before it is fast. The agent is told to call a tool whenever there is even a small chance it surfaces something relevant, because a missed data point is worse than an extra call. It does not accept a single empty result as the answer. When a tool returns nothing, it is required to try at least two distinct approaches before it concludes a figure is unavailable, which is where the financials-to-document fallback and a second search angle come in. Before it finishes, it reviews what it gathered for completeness and validates any table or chart it produced against the expected schema, so a malformed artifact or a half-answered question does not reach you.

That same standard governs how it talks to you. It corrects a wrong premise rather than going along with it, says so plainly when something genuinely cannot be found, and never fabricates a figure to fill a gap. Accuracy is held above agreeing with you.

The agent runs as of a date. It knows the date the analysis is being run for and resolves "current quarter" or "last year" against that date, so a question asked about a past period does not leak in figures that were not known yet. You can also pin specific evidence: attach a document or an earlier note to your message, and the agent reads each attachment before it answers rather than searching for something it was already handed.

Sources It Reaches For

The agent has a wide set of tools and chooses among them per question. They cover the company's numbers, its documents, the research already run on it, and the context around it. What the agent can read for a given company is set by Company and Data Coverage.

Financials. It reads structured financial reports across reporting periods: income statement, balance sheet, cash flow, and segment data. A financial question does not map straight to a database row, so the agent first works out what you are asking for. It interprets the metric and the periods, resolving "last quarter" against the date it is running as of, treating a full year as its 4Q representation, and summing the constituent quarters when a trailing-twelve-month figure is needed. It then decides whether the structured reports can answer or whether the question is better served by searching the financial documents, and a query the structured reports cannot satisfy falls back to that search rather than giving up. The figures it returns are read as reported. Consensus estimates and expectations come from this financial data, never from the open web.

Primary documents. It has the full document record for the company: annual and interim reports, earnings call transcripts including prepared remarks and Q&A, press releases, and the regulatory filings around an event. It can list what exists for a period, read a specific document, and search across all of them for a fact. Searches span the whole document, not only the summary, so a number buried in the Q&A is reachable. When it searches, an internal pass drops facts that a later disclosure has superseded, so the agent works from the current version of a fact rather than a stale one. Long filings are read in windows across turns rather than truncated.

Deep Research Agents. It runs Deep Research Agents and works from their results. It can launch an agent on the company, list the results already saved for it, read any of them, and fold their findings into its answer. A completed result, such as a guidance ledger or a tariff-impact analysis, becomes material the agent reasons over rather than a separate report you read on its own.

Beyond the company. When a question calls for it, the agent lists related companies by relationship, competitor, supplier, or customer, and it can run a web search for news and context that is not in the filings. Web search is held to general context. It is barred from using the web for estimates, ratings, or price targets, because those vary by source and would undercut the numbers.

Single Company Expert

The agent is deliberately scoped to a single company. The main conversation holds that company's full context: filings, earnings calls, press releases, and management commentary. A typical analyst covers 40 to 60 companies, and one agent stretched across all of them at once would hold none of them in the depth this work needs. Scoping to one company is what gives the agent its depth.

When your question reaches another company, the agent does not answer from general knowledge. It opens a separate expert on that company behind the scenes, asks it the question, and brings the answer back into your thread. You see which company was queried and what it found, as a visible step. Name a ticker and it queries that company. Ask about "competitors" without naming them and it uses the relationship map to decide which names to query, and it queries several in parallel for a comparison.

Asking about competitors fans out to a separate expert per peer, each step visible, with the results returned as one comparison table.

The expert it consults is a full agent in its own right, running the same loop over its own company, but held to a tight brief. It gets only the financials and the primary documents for that company. It cannot run a web search, cannot consult a third company, and cannot touch your notes. It answers the one question, grounded in primary documents with citations, and returns. That constraint is what keeps a cross-company answer as trustworthy as a direct one. Every figure still traces to a filing, and the consulted expert cannot recurse off its own company's record into a chain of secondhand lookups.

Output Artifacts

Alongside its written answers, the agent produces artifacts you can keep and take with you: research notes, tables, and charts.

Research Notes

A Deep Research Agent result opens as an editable note next to the conversation, the chat on the left and the note on the right. As you direct the work, the agent writes to the note instead of leaving the result buried in the thread. Ask it to pull capital allocation into its own section weighted by what management said this quarter, or to add a column to a peer table, and the note changes while the conversation stays a conversation. You can also turn the current thread into a note with the /to-notes slash command, which runs the same agent machinery to synthesize the discussion into a saved report.

The chat on the left, the editable note on the right. Directing the chat updates the note without cluttering the thread.

Notes are versioned. Every change writes a new version and keeps the prior ones, so the note's history stays intact. Editing a note that someone else owns forks a private copy for you first and leaves the shared original untouched. Because the note stays grounded in the same filings, calls, and press releases with its citations preserved, the work compounds across earnings cycles instead of being rebuilt each period. A note exports to PDF or Markdown, and you can copy its Markdown or share a link to it.

Tables

The agent renders two kinds of table. A plain table organizes text and values into rows. A financial table is a structured object: each column carries its own type and formatting, money, percent, multiple, or basis points, with a currency, a magnitude such as millions or billions, and a precision. Subtotal and total rows are set off in bold, columns can be grouped under a banner such as "Actuals," and the table carries a title, a source note, and footnotes. The agent validates the table against its schema before it renders, so a malformed table does not reach you. A financial table exports to Excel, CSV, or Markdown, and the Excel export preserves the number formats so the cells stay numeric.

Charts

The agent picks the chart that fits the question: a bar chart for a peer comparison, a line chart for a trend, and stacked bars, waterfalls, or pies where they suit the data.

A chart rendered inline from a question, with controls to download the image or the underlying data. Charts render inline alongside the written analysis, with a headline that states the insight, axis labels, and a source attribution. You can hover for exact values, and where a series crosses from actuals into estimates the chart marks the boundary. If a number looks off, ask a follow-up and the chart updates. A chart exports as a PNG image, and its underlying data exports to CSV.

The agent handles comparisons and summaries well. Full financial models, custom valuation frameworks, and firm-specific reporting templates still belong in dedicated modeling tools.

Usage is billed in Marvin AI Credits, where one credit typically covers a question-answer pair and Deep Research Agent results folded into a thread are billed separately.