Capital Markets Workflow
How to Build Pitch Books With Claude: Profiles, Comps, and Sourcing
The deck still needs someone who knows the client.
Building a pitch book eats analyst hours on repetitive work: pulling comparable companies, drafting profiles from public filings, formatting precedent transaction tables so the formulas are visible to whoever reviews them next. Anthropic's Claude can take on a real share of that first draft, and pitch books with Claude are becoming a normal part of how banks prep for a pitch. This guide covers what Claude can draft on its own, what a human has to verify line by line, and the QC pass that belongs between a Claude draft and a client-facing page.
Pitch books with Claude work as a drafting accelerator, not a finished deliverable. Anthropic's Claude can read public filings and market data to draft company profiles, build comparable company and precedent transaction tables with the formulas shown rather than hidden in a spreadsheet, and write a first-pass situation overview. It cannot confirm deal terms, verify that a filing is current, or know what a specific client wants to see. An analyst still checks every figure against its source and edits for voice before a page leaves the building.
Published
Free lesson · no account neededYour AI Deal Desk: Where Claude Fits in Finance WorkPrivate Equity & Finance with Claude7 minStart the free lessonKey Takeaways
- Claude can draft company profiles from public filings, saving the first pass of research and writing that normally falls to a junior analyst.
- Comps and precedent transaction tables can show their formulas as Claude builds them, so a reviewer can trace every multiple back to its inputs.
- AI pitch book generators work page by page, not deck by deck: profiles, comps, and situation overviews each need separate prompts and separate checks.
- Every number Claude produces still needs to be checked against the original filing or data source before it reaches a pitch book.
- Pitch book automation speeds up the parts of the process that are mostly pattern-matching: formatting, first-draft prose, and pulling known data points.
- The judgment calls, including which precedents are comparable, what story the numbers tell, and what the client needs to hear, stay with the deal team.
What Claude Drafts vs. What a Human Must Verify
| Pitch book component | What Claude can draft | What still needs human verification |
|---|---|---|
| Company profile | Business description, segments, leadership bios from filings | Accuracy against the latest filing, tone for the specific client |
| Comps table | Multiples calculated with formulas shown | Which peers actually belong in the set, outlier judgment calls |
| Precedent transactions table | Implied multiples and deal summaries from a transaction list | Deal comparability, timing relevance, structure differences |
| Situation overview | First-draft narrative and options framing | Whether the narrative matches what the client actually wants to hear |
| Appendix / data pack | Formatting and first-pass data compilation | Source currency and completeness |
Copy and Paste
Prompts You Can Use Today
Draft a company profile from a filing
You are helping draft a one-page company profile for a pitch book. Using the attached 10-K and most recent investor presentation, write a profile covering business description, revenue segments, recent financial performance, and leadership team, in the formatting style of our standard teaser template, and flag any figure you could not confirm against the source documents.
Attach the actual filing text or a clean extract. Don't rely on Claude's memory of a company it may recognize by name.
Build a comps table with formulas shown
Build a comparable company analysis table for the attached list of peer companies, using the financial data in the attached spreadsheet. Show EV/EBITDA and EV/Revenue multiples for the last twelve months and next twelve months, and show the formula and inputs you used for each calculated cell instead of just the output number, so a reviewer can trace every figure.
Asking for the formula, not just the result, is what makes the output checkable.
Draft a precedent transactions table
Using the attached list of precedent transactions with deal value, target financials, and announcement dates, build a precedent transactions table showing implied multiples for each deal, sorted by announcement date, and note which transactions you flagged as less comparable due to deal size, timing, or structure, along with your reasoning for each flag.
Have Claude explain its comparability judgment in writing. That reasoning is what a VP will actually check.
Write a situation overview
Write a one-page situation overview for the first section of a pitch book, covering the client's current strategic position, the rationale for considering this transaction now, and two or three options we might present, based on the attached background memo and public reporting on the client's industry, in a tone appropriate for a senior client audience.
Feed it the internal background memo, not just public info, or the overview will read generically.
Run a consistency QC pass across pages
Review the attached pitch book pages for internal consistency: confirm that every multiple, growth rate, and dollar figure referenced in the situation overview matches the corresponding number in the comps and precedent transaction tables, and list every discrepancy you find along with the page numbers where they appear.
Run this after every draft changes, not just once at the end. Numbers drift as pages get edited separately.
Draft a buyer list from criteria
Using the attached deal criteria memo, draft a preliminary buyer list of companies that fit the stated size, sector, and strategic rationale, grouped into strategic and financial buyers, with one sentence of rationale for each entry and a note on which names you are least confident about and why.
Treat the output as a starting list for the deal team to prune and expand, not a finished buyer universe.
From Anthropic
What Anthropic offers for Investment Banking
Advancing Claude for Financial Services
An October 2025 update adding Agent Skills for comparable company analysis, discounted cash flow models, due diligence data packs, company teasers and profiles for pitch books and buyer lists, earnings analyses, and initiating coverage reports, alongside a beta of Claude for Excel and connectors including LSEG and Moody's.
· Source
Agents for financial services
Ten agent templates released in May 2026, including a pitch builder, meeting preparer, model builder, and valuation reviewer, shipped as plugins in Claude Cowork and Claude Code, with Anthropic's Excel, PowerPoint, and Word add-ins made generally available.
· Source
In Practice
How Investment Banking Teams Use Claude
According to Anthropic's published customer story
RBC Capital Markets
In Anthropic's October 2025 financial services update, RBC Capital Markets' head of AI and digital innovation said Claude stands out at integrating multiple data sources and automating workflows.
According to Anthropic's published customer story
Mizuho
Anthropic's May 2026 announcement of its financial services agents quotes Mizuho's banking COO describing preparation time being turned into time for ideas.
Illustrative walkthrough
A Coverage Team Drafts a First Teaser
- An analyst feeds Claude the target's most recent public filing and asks for a company profile in the team's standard teaser format.
- Claude drafts the profile, including a business description and recent financial highlights, and flags one segment figure it could not confirm.
- The analyst checks that figure against the filing, corrects it, and adjusts the tone for the specific buyer audience.
- A VP reviews the final page alongside the source filing before it goes into the pitch book.
Illustrative walkthrough
Refreshing a Comps Set Mid-Process
- A deal team already has an agreed peer set, but one company's multiple has moved since the last update.
- An associate asks Claude to rebuild the comps table with the updated data and show the formula behind each recalculated multiple.
- Claude flags that the updated multiple now looks like an outlier relative to the rest of the set and explains why.
- The team discusses whether to keep or replace that peer, then finalizes the table for the deck.
The Draftable Pages
What Claude Can Draft for a Pitch Book
Company Profiles and Teasers
Claude can read a filing or data feed and draft a profile covering business description, segments, and recent performance in a template's format.
Comparable Company Analysis
It can build an AI for comps analysis table that shows the formula behind each multiple, not just the resulting number.
Precedent Transaction Tables
Given a transaction list, it can calculate implied multiples and flag deals that look less comparable, with its reasoning attached.
Situation Overviews
It can turn a background memo and public reporting into a first-draft narrative for the opening pages of a pitch.
Data Pack Assembly
It can pull known data points into a consistent format for a due diligence or diligence-adjacent data pack.
Setting Expectations
Is Claude an AI Pitch Book Generator?
The honest answer to 'can AI build a pitch book' is: it can build most of the pages, not the pitch. A pitch book is less a document than an argument: why this transaction, why now, why this buyer or seller fits, and an AI pitch book generator can draft the supporting pages of that argument faster than a human typing from scratch, but it does not know what argument a specific client needs to hear this quarter.
What changes with Claude is where analyst time goes. Instead of spending an evening typing a company profile from a 10-K, an analyst spends that time checking the draft against the filing, adjusting the framing for the specific audience, and thinking about what the comps set is actually saying. The mechanical parts of the page, including formatting, first-pass research, and pulling known figures, move to the model. The judgment parts stay with the person who understands the deal.
This is also why 'pitch book automation' is a better frame than 'pitch book generation.' Automation implies a human is still running the process and checking the output at each step. Generation implies you ask once and get a finished deck. The second framing sets people up to either distrust the tool entirely or trust it too much, and neither is accurate to how this actually works in a deal team.
The Comps Table
AI for Comps Analysis: What Claude Can and Can't Do
A comparable company analysis has two parts: picking the peer set, and calculating the multiples. Claude is genuinely useful for the second part. Given a list of companies and their financial data, it can calculate EV/EBITDA, EV/Revenue, and other standard multiples, and, this matters for review, it can show the formula and the inputs behind each number instead of just outputting a final figure. A reviewer can then trace any multiple back to the exact inputs that produced it, the same way they would check a cell formula in a spreadsheet.
Picking the peer set is a different kind of task. It is a judgment call about which companies genuinely compete for the same capital, customers, or exit multiple, and that judgment depends on context a model does not have: what the client thinks of as its real competitive set, what the market currently rewards, what happened to a peer's multiple last quarter that makes it an outlier right now. Claude can propose a peer set and explain its reasoning, which is useful as a starting point and a sanity check, but the final set is a call for the deal team.
The practical workflow that works well: give Claude the peer list the team has already settled on, along with the financial data, and ask it to build the table with formulas shown. Use it to flag inconsistencies, such as a multiple that looks out of line with the rest of the set, rather than to choose the set in the first place.
The Precedent Table
Precedent Transactions AI: Building the Table With Shown Formulas
Precedent transaction tables are comps analysis with an extra layer of judgment: not just which companies are comparable, but which past deals are comparable, given that deal structures, control premiums, and market conditions all shift over time. Precedent transactions AI work follows the same split as comps work. Claude can take a list of deals with their disclosed terms and target financials and calculate implied multiples cleanly, with the formula shown for each cell.
Where it earns its keep beyond pure calculation is in flagging comparability concerns in writing. Ask it to note which transactions look less comparable due to deal size, timing, or structure, and why, such as a deal from a very different rate environment, a minority stake rather than a control transaction, or a strategic buyer paying a premium for reasons specific to that deal. That written reasoning is exactly what a VP or MD checks when they review the table, so having Claude surface it up front saves a review cycle rather than replacing it.
A caution specific to this table: transaction databases and public disclosures age, and a model working from a static document can't tell you a deal has since been restated or a database entry corrected. Treat the transaction list you feed it as the source of truth, and verify that list against a live data source before the table goes into a pitch.
The Honest Split
Pitch Book Automation: Where It Helps and Where It Doesn't
Helps: First Drafts
Profiles, overviews, and table formatting go from a blank page to a reviewable draft in minutes instead of hours.
Helps: Consistency Checks
Claude can scan a full deck for mismatched figures between the narrative and the tables, a tedious check humans often skip under deadline.
Doesn't: Client Judgment
It doesn't know what story a specific client needs to hear, or what a senior banker knows about that relationship from the last meeting.
Doesn't: Final Accuracy
It doesn't independently confirm that a filing is current or that a deal term it read is the final, signed version.
Confidentiality and Review
The QC Pass Before Anything Reaches a Client
Every page a model drafts for a pitch book needs a review step before it reaches a client, and that review has to happen specifically, not generally. 'Looks right' is not a check. The check is: does every number in this profile match the filing it came from, does every multiple in the comps table trace back to the correct inputs, does the situation overview's framing match what the relationship team actually knows about this client right now.
Confidentiality matters here in a way it doesn't for a lot of other AI use cases. Pitch book work routinely touches material nonpublic information once a mandate is live, and draft pitch materials can themselves be sensitive before a client has agreed to anything. Keep deal-specific or non-public inputs inside whatever enterprise or firm-governed environment your bank has approved for this kind of work, follow your firm's data handling and information-barrier policies, and don't paste confidential deal information into a general consumer AI product. The tool is only as safe as the environment it runs in and the judgment of the person using it.
A useful habit is to run a dedicated consistency pass as a last step, after all the individual pages are drafted and before the deck is assembled: ask Claude to check that every figure referenced in prose matches the corresponding table, and read its findings as a checklist, not as a sign-off. The final sign-off is still a person who is accountable for the deck going out the door.
Getting Started
A Realistic First Workflow
If you're trying pitch books with Claude for the first time, start narrow. Pick one page type, company profiles or a comps table, rather than trying to draft an entire deck in one pass. Feed it real source documents rather than asking it to work from a company name alone, since a model answering from general knowledge about a company is exactly the kind of unsourced draft that needs the most scrutiny.
Build the habit of asking for shown work on anything with a number in it: formulas for calculated cells, named sources for factual claims, explicit flags for anything it's not confident about. This costs a little more output length but it's what turns a draft into something checkable instead of something you have to redo from scratch because you can't tell what to trust.
Once a page type is working reliably, expand to the next one. The teams that get real value from this are the ones that build a repeatable template and review habit for each page type, the same way they already have a template and review habit for the page before AI was part of the process.
Learn to Work Like This With Claude
Reading about pitch books with Claude and actually building a clean comps table with it are different skills. The free Nightschool AI curriculum is hands-on from the first lesson, so you practice the prompting and the checking habits together, not just the theory.
Frequently Asked Questions
Can AI build a pitch book?
AI can draft most of the individual pages in a pitch book, including company profiles, comps tables, precedent transaction tables, and situation overviews, but it can't build the deck's argument on its own. It doesn't know what story a specific client needs to hear, and every number it produces still needs to be checked against the original source before the page is final.
How do investment banks use AI?
Investment banks use AI mainly for first-draft work: pulling company data into profiles, building comps and precedent transaction tables with formulas shown, drafting situation overviews, and running consistency checks across a deck. The judgment calls, such as which companies are truly comparable and what argument to make to a client, stay with analysts, associates, and senior bankers, with every AI-drafted number checked against its source before a page goes out.
What is an AI pitch book generator?
An AI pitch book generator is a loose term for AI tools, like Claude, used to draft the component pages of a pitch book, including profiles, comps, precedent tables, and overviews, rather than a tool that produces a finished deck in one step. The more accurate way to think about it is as a drafting accelerator that still needs a human review pass on every page.
Is AI good at comps analysis?
AI is strong at the calculation side of comps analysis, computing multiples accurately and showing the formula behind each one, and weaker at the judgment side, which is deciding which companies genuinely belong in the peer set. Give it an already-agreed peer set and financial data, and it can build a clean, checkable table quickly.
Can AI build precedent transaction tables?
Yes, if you give it the underlying transaction data. It can calculate implied multiples for each deal and flag which transactions look less comparable due to timing, size, or deal structure, with its reasoning written out. It can't independently source a transaction list or confirm that disclosed terms are current, so that data still needs to come from a verified source.
What is pitch book automation?
Pitch book automation refers to using AI to speed up the repeatable, pattern-matching parts of building a pitch book: formatting pages, drafting first-pass prose, pulling known data into tables, and checking a deck for internal consistency. It's automation of drafting steps within a human-run process, not a replacement for the process.
Is it safe to use AI with confidential deal information?
Only inside an environment your firm has specifically approved for that kind of data, following your information-barrier and data handling policies. Pitch book work can involve material nonpublic information once a mandate is live, so confidential or deal-specific inputs shouldn't go into a general consumer AI product.
How do you check an AI-drafted comps table for accuracy?
Ask the AI to show the formula and inputs behind every calculated cell, not just the output number, then trace a sample of those cells back to the source financial data by hand. Also have it run a separate pass checking that every multiple in the table matches any figure referenced in the surrounding prose.
Keep going