AI Agents for Investment Banking: 6 Jobs for Deal Teams

Most of a junior banker's week goes on research: building comps, pulling precedent transactions, drafting buyer lists, checking every number on a pitch page, logging the week's deals and pulling aftermarket data for an ECM pitch. AI agents for investment banking can do that research in minutes, provided they read the primary record (10-Ks, 10-Qs, 8-Ks, prospectuses and deal announcements) and return every figure with its source.

This guide follows six jobs from a deal team's week. We ran each one as a Valyu DeepResearch task on November 11, 2026, then checked the key figures at the source on SEC EDGAR and in the deal announcements.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for investment banking do

An AI agent for investment banking is a language model that can search and read the documents bankers work from, such as SEC filings, merger announcements, prospectuses and market data, and return cited findings. It doesn't replace the banker's judgement on valuation or the client relationship. It does the research behind the pitch book, faster and with a source for every number.

Your own agents can call Valyu's search when they need one figure. For pitch book work, DeepResearch does the job properly: it plans the research, reads the full filings, computes the figures and returns a cited brief in the format you ask for.

Figure 1. Six deal-team jobs we ran as DeepResearch tasks on November 11, 2026, with the sources used, what came back and what the team does next.

1. Trading comps: LTM revenue from the filings, not a data feed

The problem: an analyst preparing a pitch for Dynatrace needs a trading comps table: LTM revenue, growth, margins, market cap, enterprise value and EV/revenue for Dynatrace and its observability peers.

What we ran: a DeepResearch task in standard mode that named the peers and spelled out how to build LTM revenue. Six companies' worth of filings is more reading than a one-company question, so we used standard rather than fast.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
task = valyu.deepresearch.create(
    query="Build a trading comparables set for Dynatrace (NYSE: DT) as of November 2026.",
    mode="standard",
    research_strategy="Use these peers: Datadog, Elastic, PagerDuty, ServiceNow and Cloudflare. For each company, LTM revenue is the latest 10-K fiscal year revenue minus the prior year's year-to-date revenue plus the current year-to-date revenue from the latest 10-Q. Use the latest 10-Q for margins, cash and debt, the cover page for shares outstanding, and stock data for the latest close. State the period and source for every figure.",
    report_format="A comps table with company, ticker, fiscal period, LTM revenue, growth, gross margin, operating margin, market cap, EV and EV/revenue, then the median and mean, then three notes on comparability.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-income-statement-US", "valyu/valyu-balance-sheet-US", "valyu/valyu-stocks"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited brief in markdown

What came back (in 7.0 minutes): a full comps table with the LTM arithmetic shown for every company, priced at the November 10, 2026 close:

The comparability notes covered the LTM windows that end two months apart, the large zero-coupon convertible notes that move EV depending on how they're treated, and the mix of pure-play observability and broader platforms in the set.

What we checked at the source: Dynatrace's 10-Q (quarterly revenue of $579,918 thousand, 286,944,133 shares and $1,411,242 thousand of exchangeable notes), Datadog's 10-Q (nine-month revenue of $3,324,970 thousand and 361,042,545 shares across both classes) and Cloudflare's 10-Q (nine-month revenue of $2,118,329 thousand). All match.

The catch: one outlier moves the average. The peer median is 6.7x, but the mean is 13.5x, pulled up by Cloudflare's 39.9x. Quote the median, and say which peers are in it.

The decision: show Dynatrace at 5.4x against a peer median of 6.7x, with Cloudflare footnoted or excluded, and present the highest GAAP operating margin in the set as the reason it's worth more than its growth alone suggests.

2. Precedent transactions: count which deals disclosed revenue

The problem: a team pitching a cybersecurity sale needs precedent transactions: every deal over $1 billion since January 2024, with value, consideration and multiple.

What we ran: a DeepResearch task that told it not to estimate undisclosed revenue.

task = valyu.deepresearch.create(
    query="Precedent transactions: acquisitions of cybersecurity companies valued at $1 billion or more, announced from January 2024 to November 2026.",
    mode="fast",
    research_strategy="Use SEC filings (8-Ks, merger proxies, 10-Ks of acquirers) and news. For each deal give announcement date, acquirer, target, public or private target, headline enterprise or equity value as announced, consideration (cash, stock or mix), target revenue or ARR if disclosed, the implied multiple only when both figures are disclosed, and status (closed, pending, terminated). Do not estimate undisclosed revenue.",
    report_format="A precedent transactions table sorted by date, then the median disclosed revenue multiple and a short note on which deals lack disclosed revenue.",
    search={"included_sources": ["valyu/valyu-sec-filings", "ap/news", "web"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 3.7 minutes): 23 deals with value, consideration, status and segment, each deal labelled pure-play cybersecurity or security-adjacent. The largest:

What we checked at the source: the Proofpoint and Varonis announcement ($6.3 billion, all cash) and ServiceNow's Armis announcement ($7.75 billion, $340 million of ARR). Both match.

The catch: only 4 of the 23 deals disclosed the target's revenue or ARR, so the median multiple of about 19.6x rests on four deals. Three are measured on ARR and one on revenue, and most of the multiples come from deal trackers rather than the parties. DeepResearch listed the 19 deals without disclosed revenue rather than estimating it, and called the median indicative.

The decision: use the set as value precedents, show the four multiples individually with their basis, and don't put a single precedent multiple on the valuation page.

3. Buyer list: evidence for each name

The problem: a sell-side team with a mandate for a US veterinary practice management software company (about $60 million in ARR) needs a buyer list, with evidence that each buyer fits.

What we ran: a DeepResearch task that asked for proof, not just names.

task = valyu.deepresearch.create(
    query="Build a buyer list for a sell-side mandate: a US provider of practice management software for veterinary clinics, about $60 million in annual recurring revenue.",
    mode="fast",
    research_strategy="Identify strategic buyers (public and private companies already selling software or services to veterinary practices or animal health) and financial sponsors that own veterinary or vertical-software platforms. For each, cite evidence of fit: existing veterinary software products, acquisitions in veterinary or practice-management software since 2020 with dates, and for public strategics the relevant 10-K language. Mark each buyer strategic or sponsor.",
    report_format="A buyer list table with buyer, type, rationale, evidence and source, ranked by fit, then the top five with one line on why each would pay up.",
    search={"included_sources": ["valyu/valyu-sec-filings", "ap/news", "web"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 4.1 minutes): 22 buyers ranked by fit, each with evidence and a source. The top five:

1. IDEXX Laboratories. Its 10-K lists four practice management systems ("cloud-based ezyVet, Animana, and IDEXX Neo, and on-premises Cornerstone"), and it bought ezyVet in 2021.

2. Covetrus. Owned by CD&R since 2022, it runs the AVImark and Covetrus Pulse systems.

3. Patterson Companies. Owned by Patient Square since April 2025, it took full ownership of the NaVetor system and relaunched it in January 2026.

4. Mars. Family-owned, so there's no fund clock, and it bought Heska, including its software, in 2023.

5. Instinct Science. It bought ScribbleVet in January 2026 to build a practice management platform.

The practice consolidators (Mission Pet Health, NVA, PetVet and others) were listed as channel partners rather than buyers, since none has bought practice software.

What we checked at the source: IDEXX's 10-K for 2025. The practice management quote matches the filing word for word.

The catch: one comparison site listed Cornerstone as a Covetrus product. DeepResearch noticed that IDEXX's own 10-K describes Cornerstone as an IDEXX product and called the site's claim an error. It also marked Henry Schein as uncertain, because its 10-K describes its practice software as dental, and said the sponsor-to-platform links come from a secondary compilation.

The decision: run the process with IDEXX, Covetrus, Patterson and Mars as the first tier, and confirm sponsor ownership before outreach.

4. Pitch book fact-check: the date is part of the fact

The problem: a draft company page on Snowflake has five statements. Before it goes to the client, someone has to check each one against the filings.

What we ran: a DeepResearch task with the draft statements in the query.

task = valyu.deepresearch.create(
    query="Fact-check these statements from a draft pitch book page on Snowflake (NYSE: SNOW): 1) Snowflake is headquartered in San Mateo, California. 2) Snowflake was founded in 2012. 3) Product revenue grew about 30% in fiscal 2026. 4) Snowflake had more than 700 customers with trailing 12-month product revenue above $1 million at the end of fiscal 2026. 5) Its net revenue retention rate was 126% at the end of fiscal 2026.",
    mode="fast",
    research_strategy="Use the 10-K for the fiscal year ended January 31, 2026 and later 10-Qs. For each statement, mark it Correct, Incorrect or Outdated, give the correct figure with the period, and quote the filing. If the latest 10-Q updates a figure, give both.",
    report_format="A table with statement, verdict, correct figure, period and filing quote, then corrected wording for the pitch book.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-income-statement-US"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 2.9 minutes): a verdict on each statement, with the filing quote and corrected wording:

What we checked at the source: the 10-K on EDGAR. The Menlo Park address, the San Mateo impairment, the July 23, 2012 incorporation date, the 733 customers and the 125% retention rate all match.

The catch: 126% is a real Snowflake figure, just not for the date on the page. A reviewer who searched "Snowflake net revenue retention 126%" would find it and tick the box. The research strategy asked for the period behind every figure, which is what exposed the error.

The decision: change the headquarters to Menlo Park, and either use 125% at January 31, 2026 or re-date 126% to July 31, 2026.

5. Deal log: what was announced, and what kind of deal it is

The problem: a coverage banker needs a log of semiconductor M&A announced since October 1, with values, structure, timing and approvals.

What we ran: a DeepResearch task that excluded rumours and asked which deals an 8-K confirms.

task = valyu.deepresearch.create(
    query="Which mergers and acquisitions involving US-listed semiconductor companies were announced between October 1 and November 10, 2026?",
    mode="fast",
    research_strategy="Use 8-K filings and news. Include deals where a US-listed semiconductor or semiconductor equipment company is the acquirer or the target. For each, give announcement date, acquirer, target, deal value, consideration, expected close and regulatory approvals required. Exclude rumours unless confirmed by a filing or a company press release, and say which are confirmed by an 8-K.",
    report_format="A deal log table with date, acquirer, target, value, consideration, expected close, approvals and source, then a three-sentence read-out for a coverage banker.",
    search={"included_sources": ["valyu/valyu-sec-filings", "ap/news", "web"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 3.6 minutes): five transactions with terms, timing and approvals:

What we checked at the source: the Marvell and Lumentum terms (a fixed 4.500 exchange ratio and $115 billion enterprise value) and onsemi's $600 million for Ferric. Both match.

The catch: three of the five need care. Applied Materials' Besi purchase is a minority stake capped at 20% with no board seats, not an acquisition. Intel's 10-Q puts the Classiq price at about $1.9 billion plus contingent payments, while press reports said $2.2 billion or $2.3 billion; the filing is the figure to use. And deals that closed in the window but were announced earlier, such as Skyworks and Qorvo, were listed separately so they weren't counted twice.

The decision: focus the coverage conversation on the two large deals, Marvell and Lumentum (all stock, with a long road to close) and onsemi's cash bid for Synaptics, and log Besi as a stake, not a deal.

6. IPO aftermarket: how the year's big listings have traded

The problem: an ECM team pitching an IPO mandate wants to show how 2026's large US tech IPOs have traded since pricing.

What we ran: a DeepResearch task on prospectuses and stock data.

task = valyu.deepresearch.create(
    query="How have the largest US technology IPOs of 2026 traded since pricing?",
    mode="fast",
    research_strategy="Find US technology IPOs priced in 2026 that raised at least $500 million. For each, use the final prospectus (424B4) for the initial price range, final offer price, shares and gross proceeds, and stock data for the first-day close and the latest close in November 2026. Compute pricing versus range, first-day return and return from offer price to latest close.",
    report_format="An ECM table with company, ticker, pricing date, range, offer price, proceeds, first-day return and return to date, then three observations for an ECM pitch.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-stocks", "ap/news", "web"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 9.5 minutes): full pricing and aftermarket data for the two largest listings, and pricing terms for five more:

It also gave pricing terms for SK Hynix's US listing, Stripe, Quantinuum, Fervo Energy and X-energy. Its observation for the pitch: the year has rewarded size and liquidity, and Cerebras shows how little a first-day jump says about later returns.

What we checked at the source: Cerebras's pricing announcement ($185.00 a share) and SpaceX's final prospectus ($135.00 a share). Both match.

The catch: scope is a judgement call. Renaissance Capital classes SpaceX as industrials, not technology, so say which definition the table uses. fast mode gives the headline comparison; for aftermarket returns on all seven listings, name them and use standard mode, as we did for the comps.

The decision: use SpaceX and Cerebras as the headline comparison, and rerun the table in standard mode with all seven listings named before it goes in the pitch.

What the six jobs cost

A fast task costs a fixed 25 credits and a standard task 100, about $0.16 and $0.63 at pay-as-you-go rates of 1,600 credits for $10, whatever they search and read. The six jobs cost about $1.41 together. Use fast for one question on one company, and standard when a task needs several companies' filings, as comps do.

Go further: a full company profile in one task

The six jobs each answer one question. A deal team's actual output is a document: the company profile, the comps book, the buyer universe. A single DeepResearch task in standard or heavy mode writes these, with files such as an xlsx comps table and a docx profile.

task = valyu.deepresearch.create(
    query="Prepare a sell-side company profile and valuation pack for Dynatrace.",
    mode="heavy",
    research_strategy="Use the latest 10-K and 10-Qs for business, segments, LTM financials, cash and debt. Build trading comps against Datadog, Elastic, PagerDuty, ServiceNow and Cloudflare, with LTM revenue built from the filings. Use merger announcements for observability and monitoring software deals since 2023, with disclosed multiples only. Cite every figure.",
    report_format="A company profile: Business, Financials, Ownership, Trading comps, Precedent transactions, Potential buyers, then 10 questions for management. Every figure cited with its period.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-income-statement-US", "valyu/valyu-balance-sheet-US", "valyu/valyu-stocks", "ap/news", "web"]},
    deliverables=["An xlsx comps and precedents workbook with sources", "A docx company profile"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six jobs: build LTM revenue from the filings, use disclosed multiples only and cite every figure with its period. To run the same profile across a coverage list, the DeepResearch Batch API runs the tasks in parallel and calls one webhook when they finish.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows for deal teams each come with set inputs, a recommended mode and file deliverables:

To run one from code, pass its slug as workflow_id instead of a query:

task = valyu.deepresearch.create(
    workflow_id="ib-company-profile",
    workflow_params={"company": "Dynatrace"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

To browse the full library, open Workflows in the Valyu dashboard, or see running a workflow in the docs.

Where AI agents stop in investment banking

The jobs above show what DeepResearch can do from public records. Four limits shape how to use it:

• Private deals disclose little. Most acquisitions of private companies don't disclose revenue, so precedent multiples rest on a few deals. No research can add what the parties didn't publish.
• Valuation is a judgement. The agent builds the comps and precedents. Which peers belong, which multiple applies and what premium is fair are for the deal team.
• Averages hide outliers. One 39.9x multiple moved the peer mean from 6.9x to 13.5x. Decide which peers belong before quoting a summary figure.
• Client knowledge stays with the banker. Who the board will talk to, what the founder wants and which buyer is serious come from relationships, not filings.

How to set up an investment banking research agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp and ask for the same research in plain English. To build your own agent, follow how to build a financial research agent with Valyu. For related work, see our guides to writing a first-look brief on an acquisition target, AI for due diligence and searching SEC filings in Claude and ChatGPT.

If you already run your own agents, they can call Valyu's search for single lookups, such as one company's share count. For anything that goes in a pitch book, use DeepResearch.

Three habits made the jobs above reliable:

• Ask for the period behind every figure. It caught the retention rate attached to the wrong date.
• Tell it how to compute. Spelling out the LTM formula and naming the peers gave a comps table we could check line by line against the 10-Qs.
• Quote the median with the mean. One outlier can double an average, as Cloudflare did in the comps.

FAQ

What are AI agents for investment banking?

AI agents for investment banking are language models that search and read the documents bankers work from, such as 10-Ks, 10-Qs, 8-Ks, prospectuses, merger announcements and market data, and return cited findings. Research agents such as Valyu DeepResearch plan the work, compute figures from the filings and write a brief: a comps table, a precedents list, a buyer list or a fact-checked company page.

How do investment banks use AI?

Investment banks use AI to speed up pitch book research: building trading comps, pulling precedent transactions, drafting buyer lists, checking pitch pages against filings, logging deals and pulling IPO aftermarket data. In a November 2026 test, six Valyu DeepResearch tasks did one of each for 225 credits, about $1.41.

Can AI build trading comps?

Yes. In our test, a standard-mode DeepResearch task with the peers named and the LTM formula spelled out returned a full comps table for six companies in 7 minutes, and the figures we checked matched the 10-Qs. Valyu's Comparable Companies workflow (ib-comps-analysis) builds the peer set and the table from a company name.

Can AI check a pitch book?

Yes. Give it the statements from the page and ask for a verdict, the correct figure, the period and the filing quote for each. In our test on a Snowflake page, DeepResearch found that two of five statements were wrong: the headquarters address and a retention rate attached to the wrong date. Both corrections matched the 10-K.

What is the best AI tool for investment banking?

Investment banking AI tools fall into three groups. Market data platforms, such as Capital IQ, FactSet and PitchBook, hold structured company and deal data. Banking-focused AI platforms, such as Rogo and Hebbia, search licensed documents inside their own products. Valyu DeepResearch researches primary records, such as SEC filings, prospectuses and announcements, and returns cited briefs with xlsx and docx files through an API or inside Claude and ChatGPT.

How much does AI research for investment banking cost?

With Valyu DeepResearch, a fast-mode task costs a fixed 25 credits, about $0.16, a standard task 100 credits, about $0.63, and a heavy-mode report 500 credits, about $3.13. In a November 2026 test, six investment banking jobs cost 225 credits, about $1.41, in total. Free credits are included every month.

How accurate is AI for investment banking research?

In our test, every key figure we checked matched the primary source: share counts, LTM revenue and debt from the 10-Qs, deal terms from the announcements, and the Snowflake 10-K. The points that need care are in the data itself: multiples on different bases, deals that disclose no revenue and LTM windows that end on different dates. Ask for the period and basis behind every figure.

Start with one job

Pick the job that takes your team the most time on every pitch, such as the comps table or the buyer list, and run it as a DeepResearch task on a company you already cover. Compare the brief with your own work.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For the buy side of the same research, see AI for Due Diligence.