Can an AI agent do your financial research and analysis?

Can an AI agent do financial research?

Part of it, as an assistant. An AI agent for financial research can search filings and research libraries, extract figures and terms, and draft cited primers and summaries. An analyst still checks the numbers that matter, builds the model and owns the conclusion. Treat it as a fast junior who shows its sources, not as an analyst you can leave alone.

Two firms describe this in production, both through their own or their vendor's announcements. Morgan Stanley says its AskResearchGPT lets Investment Banking, Sales & Trading and Research staff search and summarize its research, more than 70,000 reports published a year. That release dates from late 2024; it stays here because it is the firm's own description of a production research tool, and no newer first-party account was found. Hebbia says Oak Hill Advisors, a credit investor, uses its platform for credit agreement analysis, industry research and earnings updates.

The accuracy evidence is a benchmark, not a field study. Vals AI's Finance Agent benchmark asks questions that target the analytical depth of a second- or third-year investment banking analyst, answered from public-company filings. In version 2.1, updated November 11, 2026, the top model scored 69.47%. That is useful help and not yet a finished answer.

This page is not financial, tax or accounting advice.

What the agent does and what stays with you

Agent

0of 8

Agent with review

6of 8

• Find and extract figures and passages from filings and transcripts
• Search and summarize an internal research library
• Turn long documents such as credit agreements into structured data
• Draft industry primers, company primers and competitive landscapes
• Compare a company with peers or precedent deals
• Draft the summary or client note

Human

2of 8

• Build and maintain the financial model
• Form the investment view and sign off the conclusion

The dividing line is between finding and judging. The sources here show agents searching, extracting, summarizing and drafting. None shows an agent owning a recommendation or a model.

Morgan Stanley's description draws the same line. Findings from AskResearchGPT go into an email draft "ready to be modified and customized before sharing" with clients, with links to the cited research. At Oak Hill Advisors, according to Hebbia, the tool turns dense legal agreements into structured data so analysts can spend time on deal analysis. Hebbia reports analysts cut review times by up to 75% on some workflows and puts the firm's return at 6x; these are vendor figures, and the method is not published.

AlphaSense's documentation shows what a research run looks like. Its Deep Research mode "may run 50, or even 100 different searches", regularly cites 100 or more sources, takes 5 to 30 minutes, and keeps a step-by-step log of how it built the report. That log is what makes review possible.

The benchmark shows where to be careful. Vals AI reports the leading score per category. Earnings Analysis led at 93.0%, General Quantitative at 87.8% and Disclosure Analysis at 83.3%. Comparables led at 65.0%, Financial Modeling at 57.0% and Precedents at 49.3%. The original paper, on an earlier version, reported the best model at 46.8% accuracy at an average cost of $3.79 per query. Scores rose between versions, but the questions and grading changed too, so the two numbers are not directly comparable.

A worked example: one question

This is an illustration, not a customer story. Assume, for this example, that you ask the agent whether a listed retailer beat its own pre-tax margin guidance last quarter.

The agent finds the earnings release and the prior quarter's guidance on EDGAR, extracts the guided range and the reported margin, and computes the difference in basis points. Its pack shows both figures with links to the paragraphs, the calculation, and one open question: the release reports an adjusted margin, and the guidance did not say which basis it used.

The analyst opens both links, confirms the figures, and checks the earnings call transcript for the basis. It was the adjusted figure, so the agent's comparison stands. The analyst adds a line on what drove the beat and sends the note. The agent found and calculated; the analyst decided what it meant.

What you need to provide

Name the sources the agent may use, and the ones it may not. Give it read access to a filings source such as SEC EDGAR, the research or data platforms you already license, and the internal folders in scope. It touches financial data about companies and, in deal work, confidential documents. It should not touch material non-public information unless your compliance team has approved the tool for that.

Write down your house conventions: fiscal-year handling, adjusted versus reported figures, units and rounding. Benchmark questions turn on these details. Vals AI says its questions expect sector conventions that are not stated in the question and often hinge on footnotes or accounting policy disclosures.

Give it two or three past research packs your team considers good, and keep its access read-only. The agent has no business editing your model or sending client notes.

If you are a FINRA member firm, FINRA's 2026 oversight report says its rules continue to apply when firms use GenAI, and that GenAI use can implicate rules regarding supervision, communications, recordkeeping and fair dealing; under Rule 3110 a member firm must have a reasonably designed supervisory system tailored to its business. For AI agents, the report lists where to have "human in the loop" oversight and how to track agent actions and decisions among the considerations. Ask compliance or counsel which obligations apply to your firm and role.

Job brief template

Use this as the brief for a vendor, an operator or your own team. Replace the bracketed values with your own.

The source-link rule does most of the work. If every figure points to a page an analyst can open in under a minute, checking is cheap. If figures arrive without sources, you are paying for a second round of research.

Risks and human checkpoints

The main risk is not an obviously wrong answer. It is a plausible number from the wrong line, period or basis. FINRA's report describes hallucinations as inaccurate or misleading information presented as factual, and lists auditability as an agent risk because multi-step reasoning can be hard to trace. Insist on source links and the research log, and sample against them.

Watch for hedging. Vals AI changed its v2.1 grading to fail answers that give two candidate values without committing to one. Your acceptance criteria should do the same.

How to run a trial

Run the agent beside your analysts before relying on it. The length is your call. Our suggestion: cover at least one earnings season for the companies in your slice.

1. Pick a slice: 10 companies you already cover, or one recurring task such as earnings updates.
2. Give the agent read-only access to the sources in the brief. Your team works as usual.
3. Each week, compare the agent's packs with your team's work on the same questions. Count figures that differ from the source.
4. Plant a known trap, such as a quarter where adjusted and reported margins diverge, and check the agent flags the basis.
5. Time the review. Hours spent checking the agent are the cost you trade against search and extraction time.
6. Record which task types fail. Expect comparisons and models to need more rework than extraction, as the benchmark suggests.

1. 1

   Pick a slice

   10 companies you already cover, or one recurring task such as earnings updates

2. 2

   Give the agent read-only access

   To the sources in the brief; your team works as usual

3. 3

   Each week, compare the agent's packs with your team's work

   Count figures that differ from the source

4. 4

   Plant a known trap

   Such as a quarter where adjusted and reported margins diverge; check the agent flags the basis

5. 5

   Time the review

   Hours spent checking the agent are the cost you trade against search and extraction time

6. 6

   Record which task types fail

   Expect comparisons and models to need more rework than extraction

If the agent meets the brief's criteria across the trial, let it prepare that slice for real and widen the scope one task type at a time. Keep the model and the conclusion with your analysts.

For structuring the test, see how to run a paid trial with an AI agent, and for the brief, how to write a job description for an AI agent. Related pages: the financial analyst role, FP&A and forecasting, month-end close, deep research reports, competitive intelligence and the finance overview.

Questions buyers ask

Is a benchmark score the accuracy I will get? No. Vals AI's questions have one correct answer and come from public filings. Your questions may be vaguer and your sources different. Measure your own error rate in the trial.

Does a more expensive model mean better research? Not on this benchmark. Vals AI notes that mid-tier models sit a few points below the leaders at a fraction of the cost; its leaderboard shows $4.83 per test question for the top model and $0.18 for one scoring 67.54%. Those are model costs in Vals AI's test setup, not product prices.

Can it replace a junior analyst? The firms cited here describe tools that speed analysts up, not replace them. Hebbia says the tool helped Oak Hill Advisors' newer analysts "stand on the shoulders of seasoned associates", which is support for juniors, not a substitute.

Sources

1. 1
   Finance Agent Leaderboard and Methodology (v2.1, updated Nov 11, 2026)

   Vals AIVendorAccessed Nov 14, 2026

2. 2
   Finance Agent Benchmark: Benchmarking LLMs on Real-world Financial Research Tasks

   arXiv (Bigeard, Nashold, Krishnan, Wu)VendorAccessed Nov 14, 2026

3. 3
   Deep Research: Conduct In-Depth, Multi-Step Research, Autonomously

   AlphaSense Help CenterPrimaryAccessed Nov 14, 2026

4. 4
   Morgan Stanley Research Announces AskResearchGPT Powered by OpenAI

   Morgan StanleyVendorAccessed Nov 14, 2026

5. 5
   Oak Hill Advisors Unlocks 6x ROI with Hebbia

   HebbiaVendorAccessed Nov 14, 2026

6. 6
   2026 FINRA Annual Regulatory Oversight Report (PDF; GenAI: Continuing and Emerging Trends, pp. 25–28)

   FINRAPrimaryAccessed Nov 14, 2026