Claude for Investment Banking: 21 Agents for Deal Execution
The question is not whether Claude can do banking work. It is how much of a live deal you are willing to run through a system that cannot sign anything. These 21 agents answer that by splitting the deal into six stages, giving each agent one job, and never letting the agent that built a number be the one that checks it.
Every live deal produces a chain of evidence that somebody has to stand behind. In fiscal year 2024, according to the 47th Annual Hart-Scott-Rodino Report issued by the Federal Trade Commission and the Department of Justice Antitrust Division, companies notified the agencies of 2,031 transactions under the HSR Act, approximately one quarter of which were valued at more than a billion dollars. Behind each of those filings sat a model, a data room, a committee memo and a named person who signed for the numbers.
This pack is 21 agents, one job each, mapped to the six stages a deal actually moves through: Originate, Value, Model, Diligence, Execute and Communicate. What separates it from a folder of prompts is the control architecture wrapped around it. A builder never reviews its own work. Every input carries a source, an as-of date and a status. Nothing leaves the system until a named human releases it.
If what you want is breadth instead, meaning capital raising, syndication, negotiation and post-close mechanics, this is the wrong pack and the routing table below sends you to the right one.
Quick answer: which Claude agents do you need for a deal?
Work backwards from the stage the deal is standing in rather than from the document you have been asked to produce. The stage decides what evidence exists, and the evidence decides which agent can honestly run.
For investment banking work in Claude, the 21-agent deal engine covers one live M&A process end to end across six stages: Originate for the sector screen, buyer list and thesis; Value for the DCF, trading comparables, precedent transactions and LBO; Model for the three-statement build and its independent audit; Diligence for the data room index and earnings quality; Execute for the CIM, management presentation and buyer Q&A; and Communicate for the committee memo and recommendation. For breadth beyond M&A, including capital raising, negotiation and post-close mechanics, use the 60-agent deal flow set instead. For a single method applied once rather than a stack of agents, use the 10 Claude M&A skills. For the slides at the end, Oria or a design tool.
How we chose the 21 agents
Five rules decided what went into the pack and what was cut.
- One job per agent. Anything that claimed to cover origination through closing was cut and split. Each of the 21 names the stage it serves, states its inputs, and refuses the work either side of it. A bounded worker you can audit beats a clever one you cannot.
- The builder is never the reviewer. Model Audit Review is read-only by construction. It returns exceptions to the agent that built the model rather than silently repairing them, and a material change forces a fresh review. The same separation runs through Diligence and Execute.
- Every output has to carry its evidence. Each agent labels an input reported, calculated, assumed or not found, and attaches a source ID and an as-of date. An answer that arrives without its lineage cannot be checked by the next reviewer, so it does not count as an output here.
- We ignored anything that sends, files or releases. No outreach sequences, no filings, no counterparty contact, no trading. The distributed defaults produce drafts and logs. The moment an agent can act on the outside world, the control question stops being about quality and starts being about authority.
- We ignored benchmark scores. One number did change the design, though. FinanceBench, built by researchers at Patronus AI, Contextual AI and Stanford University and published as arXiv:2311.11944, found that GPT-4-Turbo used with a retrieval system incorrectly answered or refused to answer 81% of questions in its reviewed sample. That is why these agents return their evidence rather than only their answer.
Key takeaways
- Pick by stage, not by deliverable. The same target file supports a coverage pitch, a committee memo and a bid. Each stage needs a different agent even though the underlying company has not changed at all, because what has changed is what you are allowed to assert.
- Count the agents you will actually run. Most deals use six or seven of the 21 in a given week. Installing all of them is cheap; running all of them is not. Pick the stage you are in, run those agents, and leave the rest uninvoked rather than half-fed.
- An agent is only as good as the evidence you hand it. Each one expects specific primary material: filings, one market-data cut, an adjustment schedule, a data room index. Without that material you get a plausible-sounding template with nothing verifiable inside it.
- Expect a working artefact, not a finished document. What comes back is a screened set, an exception list, a risk register or a page brief. That is the state an associate hands upward, not the state a client, a lender or a board ever sees.
- Know whose signature closes each stage. Valuation weighting, the price you advise, the fairness of consideration and anything disclosed under securities rules remain with named people. The pack makes that explicit per stage rather than leaving it implied.
The six stages compared, in deal order
One row per stage, against the five decision rules above, in the order a deal moves. The slide layer sits where it actually sits, between building the execution materials and putting a recommendation in front of a committee.
| Stage | Agents | What it needs from you | What comes back | Whose signature closes it |
|---|---|---|---|---|
| Originate | Sector Screen Mapping, Buyer & Acquirer Lists, Thesis & Angle Development | A sector definition, a screening rule set and approved sources | A ranked candidate universe, a counterparty list with evidence, a falsifiable thesis | The coverage banker who takes the pitch |
| Value | DCF Valuation, Comparable Companies, Precedent Transactions, LBO Analysis | Filings, one market-data cut with an as-of date, a peer definition | Four transparent valuation views, each input labelled and sourced | The valuation owner, not the agent that built it |
| Model | Three-Statement Build, Sensitivity Tables, Scenario Toggle Design, Model Audit Review | A driver set, historicals, and accounting conventions stated up front | A forecast backbone, coherent ranges, and an exception list from a read-only reviewer | The modeller, after the reviewer returns exceptions |
| Diligence | Data Room Synthesis, Earnings Quality Review, Risk Flag Register | Data room access, the adjustment schedule, the management representations | A source ledger, a test of the adjustments, a consolidated risk register | Legal, tax and accounting advisers, who this does not replace |
| Execute | CIM Drafting, Management Presentation, Process Letter Drafting, Buyer Q&A Management | Reviewed numbers and a cleared fact base, with source IDs attached | Drafts in which every claim points at a source ID, and a Q&A log | The deal captain, who releases nothing until it is checked |
| Slides | Not an agent. Oria, or a design tool such as Gamma for visual work | Approved page briefs: headline, message, chart spec, source IDs, caveat | Dense corporate slides held to the firm template, or a visual deck elsewhere | The banker whose name is on the page |
| Communicate | Deal Memo Drafting, Committee Recommendation, Committee Q&A Prep | The cleared analysis, the alternatives considered, the open conditions | A structured decision case, the conditions, and the hardest questions first | The investment or commitment committee |
Which Claude agents originate a deal and build the buyer list?
Three agents run the origination stage. Sector Screen Mapping takes a market definition and a screening rule set and returns a ranked candidate universe with the rule that admitted each name attached. Buyer & Acquirer Lists builds the counterparty universe from disclosed transaction history, stated strategy and capacity, and tiers it by fit rather than by size. Thesis & Angle Development turns cleared research into a thesis written so it can be falsified, with the counterarguments listed beside it.
None of the three contacts anybody. They surface candidates, evidence, gaps and the case against, which is the part a coverage banker usually has to assemble by hand the night before a pitch. The falsification requirement matters more than it sounds: a thesis that cannot be wrong cannot be tested by the committee that has to approve pursuing it.
The trade-off: origination is the stage most dependent on data you may not have. These agents read what you give them. If your firm runs a market-data terminal, they work downstream of its exports rather than instead of them, and a screen built on stale exports will look just as confident as one built on fresh ones.
How do you build a DCF, comps and an LBO with Claude agents?
Four agents cover the valuation stage: DCF Valuation, Comparable Companies, Precedent Transactions and LBO Analysis. Each builds one view and keeps it transparent. Every input needs a source, an as-of date, a definition and a status label, and each multiple is recalculated rather than copied. Comparable Companies screens candidates against the criteria you approved, on one market-data cut, and stops if source dates conflict instead of quietly reconciling them.
The four views are deliberately not merged into a single number. What comes back is four sets of analytical support with their assumptions visible, which is the raw material for a football field and a written view on weighting. The weighting itself is a judgement, and the pack treats it as one.
The discipline that matters most here is the as-of date. A comparable set assembled on Tuesday and a precedent set pulled on Thursday will disagree, and the disagreement is invisible by the time it reaches a page. Comparable Companies therefore stops on conflicting source dates rather than silently picking one, and Precedent Transactions records the cycle each deal closed in, because a multiple paid in a different rate environment is a different multiple.
The trade-off: these agents reason about a valuation, they do not construct the workbook. If you need the cells, the circularity and the debt schedule built and audited in Excel, use Claude for Excel and the Excel modelling pack, then bring the outputs back here.
Can Claude build and audit a three-statement model?
Four agents, and the fourth is the point. Three-Statement Build lays the forecasting backbone against a stated driver set and stated accounting conventions. Sensitivity Tables and Scenario Toggle Design produce ranges that stay coherent with each other rather than three unrelated cuts that happen to share a tab. Model Audit Review then reads all of it as an independent challenge.
Model Audit Review is read-only by design. It does not repair the builder's work, because a reviewer that edits becomes a second builder and the audit trail disappears. It returns an exception list, the builder resolves it, and a material change triggers a fresh review. That loop is slower than a single agent fixing its own output, and slower is the intended behaviour.
What the reviewer checks is narrower than a full model audit, and more useful for being narrow: whether the three statements tie, whether the balance sheet balances under every scenario toggle rather than only the base case, whether a driver has been hard-coded somewhere downstream of where it is declared, and whether each sensitivity table is still keyed to the inputs it claims to flex. Those four failures account for most of the model errors that survive to a committee room.
The trade-off: the loop costs turns and tokens, and on a small internal analysis it is overhead you do not need. Run the builder alone on something reversible. Turn the reviewer on the moment the output is going to be shown to somebody who will act on it.
Can Claude read a data room and flag earnings quality?
Three agents. Data Room Synthesis builds the source ledger, meaning an index in which every document has an ID that later claims can point back to. Earnings Quality Review tests whether the adjustments are supported and whether the earnings convert to cash. Risk Flag Register consolidates reviewed issues into one place with an owner and a status against each.
These are evidence and triage tools. They do not complete legal, tax, accounting or quality-of-earnings diligence, and the agent files say so in their own instructions. The value is in the ledger: once every number in the CIM and the memo can be traced to a document ID, the argument in the committee room moves from whether a figure is right to whether the source is good.
The ordering is deliberate. Data Room Synthesis runs first because nothing downstream can cite what has not been indexed. Risk Flag Register runs last because a risk with no source ID behind it is an opinion. Earnings Quality Review sits between them and is the one worth running twice: once on the vendor adjustments as presented, and once on the same adjustments with the supporting schedules attached, because the gap between those two runs is usually the finding.
The trade-off: this stage is where the confidentiality question bites hardest, and it is not one a download page can answer for you. Data room material is usually under an NDA that names permitted recipients. If you do not know your firm's documented position on uploading it, the correct default is not to.
Can Claude draft a CIM and a management presentation?
Four agents handle execution: CIM Drafting, Management Presentation, Process Letter Drafting and Buyer Q&A Management. All four are gated. They run only after the controlling evidence and numbers have been reviewed, and every claim they write points at a source ID from the Diligence ledger. Buyer Q&A stays in draft state until an authorised human approves and releases it, because an answer to a bidder is a representation, not a document.
Process Letter Drafting exists to make bids comparable: instructions, basis and deadline stated once so that what comes back can be scored on more than price. The management presentation agent produces an outline and a talk track rather than a finished deck, which is the handoff into the next stage.
The gate is what makes this stage safe to automate at all. A CIM assembled from unreviewed numbers is not a faster CIM, it is a liability with a cover page. So the drafting agents refuse to start before the Diligence ledger exists, and they write source IDs inline so the first review pass is a reconciliation rather than a re-read. What that costs you is the ability to show a draft on day one, which is a real cost and the reason people skip the step.
The trade-off: for pitch book page architecture specifically, this stage is shallower than the dedicated pack. The 16 pitch-book skills go further on mandate logic and page-by-page storyline than four execution agents carrying a whole process can.
What turns approved deal analysis into committee slides?
This is the stage Oria is built for, and it is worth being precise about what that means. Oria is aimed at corporate documents: board packs, steering committee decks, diligence exhibits, operating-model slides, dense frameworks held to an enforced corporate template. It reaches users two ways, as a PowerPoint add-in running in the task pane on Windows, macOS and PowerPoint for the web, and as a connector for Claude and ChatGPT over MCP, so a slide can be built from the chat you are already in without opening PowerPoint at all.
The handoff from the agents is a page brief rather than a prompt: an answer-first headline, the supporting message, a chart specification, the source IDs and the caveat. Run a deck check after the first complete draft and after every material model update, reconciling repeated numbers, dates, units, chart labels, source notes, confidentiality legends and disclosures. On security posture, stated plainly: no training on customer content, and Professional and Team do not persist presentation content after delivery, while Enterprise adds private cloud deployment and custom LLM integration.
The trade-off: Oria is built for the corporate environment only, and it loses on highly visual work. Founder fundraising decks, launch and campaign decks, student presentations and marketing one-pagers are not what it is for. Where the job is visual impact rather than defensible content, a design-led tool such as Gamma, Canva, Pitch or Beautiful.ai will beat it and will look better doing it. Go there instead.
How do you get from analysis to an investment committee memo?
Three agents close the loop. Deal Memo Drafting structures the decision itself: what is being asked for, on what basis, under what conditions. Committee Recommendation forces the alternatives to be written down alongside the recommendation, including the alternative of doing nothing. Committee Q&A Prep drafts the hardest questions the committee will ask and answers them before the meeting rather than during it.
The design assumption is that a committee paper fails on omissions rather than on errors. A memo that states the conditions, the open items and the case against is more useful to a decision-maker than one that reads as advocacy, and it is also the version that survives being read again six months later when something has gone differently than planned.
The trade-off: Claude can assemble an evidence-backed case, but it has no standing in the room. The committee and the accountable bankers decide, and nothing in the pack changes who carries that. If you want an agent that also tracks the decision through to closing, the 60-agent deal flow set carries workstream, timeline and post-close agents that these 21 do not.
Can you use these agents for investment research rather than banking?
Partly, and it is worth being honest about where the fit breaks. The Value and Model stages transfer cleanly. A DCF, a comparable set and a three-statement build are the same artefacts whether the buyer is a strategic acquirer or a portfolio manager, and the evidence discipline is arguably more useful on the buy side, where there is no adviser to carry the blame for a bad input.
What does not transfer is everything built around a process. Originate assumes a mandate and a counterparty universe. Diligence assumes a data room somebody has granted you access to. Execute assumes a sell-side document set and a bidder on the other end of it. Communicate assumes an investment or commitment committee with a specific paper format. Point those four at a research idea and you get well-formed documents about a process that is not happening, which is the most expensive kind of plausible output because it takes a while to notice.
The trade-off: if the job is investment analysis rather than deal execution, the general finance packs fit better. The 20-agent finance playbook and the 23 finance agents for modelling are written for recurring analysis rather than a one-off process, and neither carries the deal-stage assumptions that make the other four stages here worth the setup cost.
Why the builder and the reviewer are never the same agent
The separation of roles in this pack is not an AI idea. It is an old complaint about deal work, made most directly by Aswath Damodaran of NYU Stern, who argued on his blog Musings on Markets in December 2012 that "the deal making has to be separated from the deal analysis", and that allowing the dealmaker to also be the deal analyst is, in his words, "a recipe for bad deals and we have no shortage of those." His remedy was a third party with no stake in whether the deal closes, paid only to assess it. In a stack of agents that third party is cheap to create, which is one of the few things automation genuinely makes easier rather than only faster.
The regulatory picture points the same way. The Financial Conduct Authority and the Bank of England reported from their 2024 survey of artificial intelligence in UK financial services that 75% of firms are already using AI with another 10% planning to within three years, that 55% of AI use cases involve some automated decision-making, and that only 2% are fully autonomous. Adoption is close to universal; unsupervised autonomy is a rounding error. The same survey found that 34% of firms claimed complete understanding of the AI they use, against 46% with only partial understanding, which is the gap a source ledger is meant to close.
So the control rules are deliberately boring. Use only firm-approved environments and entitled sources. Treat nonpublic deal information as confidential and potentially MNPI, and keep each deal inside its own permission boundary. Preserve an evidence ledger and exact versions. Separate reported facts, calculations, assumptions and judgements. Never invent a missing value. Never let a builder be its only reviewer. Never parallelise agents that write to the same workbook or deck. Escalate conflicting sources, unstable models, unclear permissions, legal or accounting questions and unsupported claims.
How do you prompt Claude for investment banking work?
Write the prompt as a staffing note rather than a request. State the deal ID and the audience. Define the objective. Name the approved sources and the as-of date. Specify currency, units, periods and accounting conventions. List the constraints. Describe the artefact you expect back. Define the checks it has to pass. Then tell Claude when to stop.
A worked example: ask the Comparable Companies agent to screen candidates against five approved criteria, use one market-data cut, label every input reported, calculated, assumed or not found, recalculate each multiple rather than lifting it, and stop if source dates conflict. That is reproducible next week by somebody else. "Find me good comps" is not, and the two prompts cost the same to write.
The pack ships a reusable work-order template for exactly this, plus the rule for when a stable procedure should be promoted from an ad hoc prompt into a saved Claude Skill: after the team has run it manually, reviewed the ordinary and the edge cases, captured the templates and definitions, and assigned an owner and a version. Anthropic describes a Skill as a folder of instructions, scripts and resources that Claude loads when it is relevant, which is the right shape for a settled method and the wrong shape for live deal data. Keep the deal data out of it.
If you need X, go to Y
Plenty of people arriving on a Claude investment banking query do not want this pack at all. Anthropic now ships its own finance agent templates, including a pitch builder, a model builder and a valuation reviewer, running as plugins in Claude Code and Claude Cowork with vendor data integrations behind them. If your firm already has that relationship, start there rather than here.
| If you need | Go to | Why |
|---|---|---|
| Breadth past M&A: capital raising, negotiation, post-close mechanics | The 60-agent deal flow set | Reaches IPO readiness, roadshow, syndication, earnouts, escrow and antitrust. These 21 do not. |
| A single-turn procedure rather than a stateful agent | The 10 Claude M&A skills | A skill applies one method once. No handoffs, no reviewer, far less to install. |
| Pitch book pages specifically | The 16 pitch-book skills | Page architecture and mandate logic, at a depth the Execute stage here does not reach. |
| Sponsor-side diligence on a buyout | The private equity diligence pack | Written from the buyer side, where the diligence question is the investment question. |
| The workbook itself, built and audited in Excel | Claude for Excel, or the Excel modelling pack | These agents reason about a model. They do not construct one cell by cell. |
| Vendor-integrated templates with market data behind them | Anthropic's own finance agent plugins | If your firm already has the vendor relationship and the data integration, start there. |
| A corporate or committee deck held to an enforced template | Oria | Our own product. Built for dense corporate documents and nothing else. See the block above. |
| A visually striking fundraising, launch or campaign deck | Gamma, Canva, Pitch, Beautiful.ai | Design-led tools win outright where the job is visual impact rather than defensible content. |
| Charts wired to a live workbook | think-cell | A charting add-in with data links, not a writing or reasoning tool. |
For the wider finance picture rather than M&A specifically, the 20-agent finance playbook and the 23 finance agents for modelling cover valuation and planning work outside a deal, and the full skills library lists every pack on this site.
Frequently asked questions
Can Claude do investment banking work?
It can do the parts of the work that are reproducible from evidence you supply: screening a sector, assembling a peer set, building a forecast backbone, indexing a data room, drafting a memo against reviewed numbers. It cannot carry responsibility. The valuation you advise, the price you recommend, the fairness of consideration and anything disclosed under securities rules stay with named people who sign for them.
Are these Claude Skills or Claude agents?
Agents. Anthropic defines a Skill as a folder of instructions, scripts and resources that Claude loads when it is relevant, while a subagent in Claude Code is a bounded worker defined in .claude/agents/ that runs in its own context window with its own tools and permissions. These 21 files are subagents, because the deal work needs handoffs and an independent reviewer rather than a single procedure applied once.
What is the difference between this 21-agent pack and the 60-agent set?
Scope and posture. These 21 cover one live M&A process end to end with a control architecture wrapped around it: builder and reviewer separated, a source ledger, stop conditions, a human release gate. The 60-agent deal flow set is a breadth catalogue that also reaches capital raising, syndication, negotiation, earnouts, antitrust and post-close mechanics, with no separation of roles. Roughly half the jobs overlap.
Can the agents edit my model or send client materials?
No. The distributed defaults are read-oriented and produce proposed analysis, logs and drafts. They do not send communications, publish documents, contact counterparties, trade, or approve valuation, legal, tax, accounting, compliance or release decisions. Every agent writes to a file you read. If you want one to write into a live workbook, that is a change you make deliberately and review before it touches a deal file.
Is it safe to upload a CIM or data room documents to Claude?
That is a question for your compliance function and your NDA, not for a download page. Check two things before pasting anything: what your firm agreed with the vendor about training and retention, and whether the material is under an NDA naming permitted recipients. Treat nonpublic deal information as potentially MNPI and keep each deal inside its own permission boundary. If you do not know your firm position, assume it is no.
How do I install the 21-agent pack in Claude Code?
Download the ZIP, unpack it, and copy the agent files into .claude/agents/ in the project folder you run deals from, or into ~/.claude/agents/ to make them available everywhere. Each file is Markdown with YAML frontmatter naming the agent, its description and its tools. Claude Code picks them up on the next session. The work-order template and routing rules travel in the same ZIP.
Run the first controlled workflow
Do not start on a live mandate. Pick one reversible internal task, write the work order out in full, freeze the inputs, run the builder and the reviewer as separate agents, and put a named human at the approval gate. Then look at what the reviewer caught. That exception list is the honest measure of whether the control architecture is doing anything, and it is a far better test than any benchmark score you could quote at a committee.
Claude helps the team find, test and structure the answer. Oria helps turn the approved answer into complex slides. Neither of them signs the memo.