AI for Investment Banking: Comps, Pitch Books, CIMs and Diligence Without Breaching the Wall

Where AI saves banker hours, and where it creates risk

Investment banking runs on documents produced under deadline: pitch books for a Monday meeting, a CIM before launch, a fairness opinion backup, a model that has to tie out. Most of the hours behind them are spent finding numbers, formatting pages and checking each other’s work. Those are exactly the tasks language models and agents now handle reasonably well.

The risk sits in the inputs. A coverage banker’s pitch uses public data. An execution team’s CIM uses the client’s confidential management accounts. A buy-side diligence team reads another company’s data room. The same assistant can be harmless on the first and a wall breach on the third, so the useful question is not whether a bank should use AI but which task goes to which deployment.

This guide covers the main workflows, then the controls: information barriers, supervision and records, and model risk. Regulatory and vendor pages were checked on 22 October 2026.

AI for investment banking analysts: comps, precedents and filing summaries

Comparable companies and precedent transactions

A comps pull has three steps: choose the peer set, gather financials and multiples, and calendarise and normalise them. AI helps with each, as long as the numbers come from a data provider rather than the model’s memory.

• Peer set. A model can propose candidates from business descriptions in filings and explain why each fits or does not. The senior banker still decides the final set.
• Data gathering. Through a connector to the bank’s licensed data provider, an agent can fetch revenue, EBITDA, net debt and share counts for each peer and fill the template.
• Normalisation. Fiscal year calendarisation, currency conversion and adjustments for one-offs are rules a spreadsheet tool applies better than a model reasoning in text.

Precedent transactions follow the same pattern with deal announcements, merger proxies and press releases as sources. The model’s value is reading the announcement and extracting consideration, structure and stated multiple, with a link to the source.

Earnings and filing summaries

Summarising a 10-K, 10-Q, earnings call transcript or annual report is the lowest-risk, highest-volume use in the bank. Public information carries no confidentiality issue, and a good summary saves an analyst an hour per company. Ask for a fixed structure: guidance changes, segment trends, margin drivers, capital allocation and management’s tone on M&A. Then check every number against the filing, because models still misread tables.

AI for investment banking presentations: drafting the pitch book

Pitch books are where analysts lose nights, and where Office add-ins have made the most visible progress since 2025. A model can draft situation overview pages, buyer universe slides, valuation summary text and process timelines in the bank’s PowerPoint template, pulling figures from the comps and the model.

What still needs a person:

• Story. Why this client should sell now, or which buyers will pay most, is the judgement the client is paying for.
• Number consistency. Every figure on a page has to match the backup. An agent can check that, but a person signs off.
• Brand and compliance. Pitch materials are communications with the public for a broker-dealer, so the bank’s review process applies whether a person or a model wrote the page.

A practical setup is a library of approved pages and language, with the model drafting only from that library and the deal’s own backup files. That keeps tombstones, league table claims and credentials consistent with what compliance has already approved.

Writing the CIM and supporting sell-side diligence

On a sell-side mandate the bank writes the confidential information memorandum from the client’s own material: management presentations, budgets, customer data and the quality of earnings report. AI can produce a first draft of each section from that material and keep the numbers tied to the source.

Sell-side diligence preparation also benefits. Before launch, a team can ask a model to read the data room the way a buyer would, list missing documents, flag contracts with change-of-control clauses that buyers will ask about, and draft answers to the questions buyers usually raise. During the process, a buyer Q&A log can be triaged and answered from the data room, with each draft answer citing the document behind it.

All of this material is client confidential, and much of it is covered by the engagement letter and the bank’s own confidentiality obligations to the client. On buy-side mandates, the bank is also bound by the NDA the client signed with the seller. Review for these workflows looks like the legal due diligence use case: extraction with citations and a named reviewer for each finding. For credit-side work inside the same bank, the credit memo agents guide covers a similar drafting pattern.

Auditing the financial model

Model review is a good fit for agents that use tools. A language model reading cells as text will miss things; an agent that opens the workbook, walks the formula graph and runs checks will not.

Useful automated checks:

• hard-coded numbers inside formula rows, and formulas that break the pattern of their row;
• links to external workbooks and references to stale tabs;
• balance sheet balance, cash flow tie-out and circular references;
• historical figures that differ from the filed financials;
• sensitivity tables that do not recalculate, and scenario switches that change nothing.

The model’s job is to explain each issue in plain language for the associate who owns the model. The fix stays with the associate, and the reviewer sees both the issue list and the change log.

Information barriers when the assistant can read everything

A bank keeps private-side deal teams apart from sales, trading and research because those teams must not act on material nonpublic information. Section 15(g) of the Securities Exchange Act requires every registered broker-dealer to establish, maintain and enforce written policies and procedures reasonably designed to prevent misuse of MNPI by the firm and its associated persons.

An AI assistant is a new way through that wall. If deal documents and research notes go into one shared index, a public-side user can retrieve a passage from a live mandate without anyone noticing. The model does not know who is over the wall. Only the retrieval layer can enforce it.

Design points that follow:

1. One collection per live deal, with membership tied to the deal team list that the control room already maintains.
2. Retrieval filtered by the user’s entitlements before anything reaches the model, never after.
3. No shared conversation memory across private and public sides.
4. Logged wall-crossings and logged retrieval, so surveillance can see which passages each user retrieved.
5. Deal codenames carried into prompts and logs, so watch-list and restricted-list reviews can search them.

Our guide to information barriers for relationship manager assistants covers the same design for commercial banking teams in more depth.

Supervision and records: FINRA 3110 and SEA Rule 17a-4 in brief

Investment banks that are, or sit inside, registered broker-dealers carry FINRA and SEC obligations into any AI tool their bankers use.

• Supervision. FINRA Rule 3110 requires a supervisory system reasonably designed to achieve compliance. In Regulatory Notice 24-09, published on 27 June 2024, FINRA said its rules are technology neutral, created no new obligations, and listed Rule 3110 and Rule 2210 among those that apply to generative AI.
• Records. SEA Rule 17a-4(b)(4) requires broker-dealers to preserve communications they send and receive relating to their business as such, including inter-office memoranda. Whether a given prompt or output is such a communication depends on its use, but a draft that becomes a client email or a pitch page is plainly in scope. Logging prompts and outputs with the model version makes the question answerable.
• Agents. FINRA’s 2026 Annual Regulatory Oversight Report, released in December 2025, added risks specific to AI agents, including agents acting beyond the user’s intended authority and multi-step actions that are hard to trace.

The FINRA AI guidance overview goes through these obligations rule by rule. The practical takeaway for a deal team is short: use only the tools the firm has approved, and assume anything typed into them may be reviewed.

Model risk at bank-owned investment banks: SR 26-2

Banks that own their investment bank also answer to prudential supervisors. On 17 April 2026 the Federal Reserve, the OCC and the FDIC issued SR 26-2, Revised Guidance on Model Risk Management, which supersedes SR 11-7 from 2011 and SR 21-8 on BSA/AML models. The letter says it is expected to be most relevant to banking organisations with more than $30 billion in total assets.

The guidance places generative and agentic AI models outside its scope because they are novel and rapidly evolving. It then says a banking organisation’s broader risk management and governance practices should guide the controls for tools it does not cover. In other words, there is no model validation template for a pitch-drafting assistant, but the bank is still expected to govern it. Traditional statistical models and non-generative machine learning stay inside SR 26-2.

Claude for investment banking: what Anthropic offers

Claude comes up often in banking searches because Anthropic has built a dedicated financial services offering. From Anthropic’s own announcements:

• July 2025. Claude for Financial Services launched with Claude for Enterprise, Claude Code, connectors to data providers including FactSet, S&P Global, PitchBook, Morningstar and Daloopa, and implementation partners. The announcement said customer data is not used to train Anthropic’s generative models by default.
• October 2025. Claude for Excel entered beta as a research preview, with new skills for comparable company analysis, discounted cash flow models, due diligence data packs and earnings analyses.
• May 2026. Anthropic released ten agent templates. Five are aimed at research and client coverage (pitch builder, meeting preparer, earnings reviewer, model builder, market researcher) and five at finance and operations. Excel, PowerPoint and Word add-ins became generally available, with Outlook listed as coming soon, and new connectors included SS&C Intralinks and Third Bridge. The templates run as plugins in Claude Cowork and Claude Code, or as cookbooks for Claude Managed Agents, which Anthropic says log every tool call for compliance review.

Two facts matter for a bank’s risk review. Claude runs under Anthropic’s commercial terms as a cloud service, directly or through cloud platforms, and there is no version that installs on the bank’s own servers. Retention, zero data retention eligibility and audit features differ by plan, which our guide on whether Claude is safe for confidential information sets out.

Choosing AI tools for investment banking

Most banks now evaluate four kinds of tool:

• Enterprise assistants with financial data connectors. Strong on public-company research and Office work. Data leaves the bank for the provider’s cloud under commercial terms.
• Office-native copilots. Convenient inside Excel and PowerPoint, with the same cloud processing question.
• Data room AI. Useful for diligence when the client or seller has enabled it, since processing stays under the data room contract.
• Self-hosted platforms. Models, retrieval and logs run inside the bank, so deal walls, retention and surveillance follow the bank’s own systems.

The “best” choice usually combines two of these. The table below is a starting point for that split.

How VDF AI fits an investment banking coverage and execution team

VDF AI installs on servers the bank controls, on-premises, in a private cloud or air-gapped. Models, retrieval indexes, conversation history and audit logs remain inside the bank’s network, and with external model APIs switched off, deal documents are processed only by local open-weight models. VDF AI holds no FINRA, SEC or prudential approval; it runs inside the bank’s own supervisory system, model governance and vendor review.

The information barrier design above maps onto the platform directly. Each mandate can have its own knowledge collection, and permission-aware retrieval checks the user’s role before any passage reaches a model, so a public-side analyst’s query cannot surface a private-side document. Governed agents can run comps pulls, model checks and Q&A triage with tool permissions set per role, and a workflow can stop at a human approval step before anything goes to a client. Each turn writes a record of the user, the model version, the retrieved sources and any tool calls, which the bank can export to its surveillance and archive systems.

The VDF AI Router lets a bank register local model endpoints from Ollama, vLLM or its own servers, and add a cloud model only for task types its policy allows, such as public filing summaries. On-premises sign-in uses Microsoft Entra ID natively, with other identity providers working through an SSO-aware reverse proxy, and role-based access control is included on every plan. The finance and banking solution, the finance agents catalogue and the finance executive brief show the wider setup. This article is general information and not legal or regulatory advice.