AI in Investment Banking: Use Cases by Deal Stage

In a demonstration reported by CNBC in September 2025, JPMorgan's internal AI platform built a five-page presentation with news, earnings and a peer comparison in about 30 seconds. That is work a team of analysts used to do overnight. AI in investment banking is no longer a pilot story told at conferences: it is a production tool sitting inside the largest banks, and the gap between the banks that have organized around it and the ones still debating it is widening every quarter.

This guide is written for managing directors, COOs and heads of strategy at investment banks, M&A boutiques and corporate finance advisory firms. It does not list tools. It walks through the deal lifecycle stage by stage, shows where AI creates measurable value and where it creates risk, and closes with a self-assessment, a cost model and a 30/60/90 day plan you can actually run.

I have spent more than twenty years building companies and helping leadership teams implement new technology. The pattern I see in banking is the same one I have seen everywhere else: the technology is rarely the bottleneck. Process design, data access and governance are.

What AI in Investment Banking Actually Means in 2026

The phrase covers three very different things, and confusing them is the first mistake most firms make.

1. Productivity AI for bankers. Large language models that draft, summarize, search and reformat. This is where the 2025 and 2026 rollouts at the bulge bracket banks have concentrated.
2. Analytical AI inside deal work. Models that screen targets, flag anomalies in data rooms, extract terms from contracts and pressure-test valuation assumptions.
3. Agentic workflows. Systems that chain several steps together, such as pulling comparable transactions, populating a model template, drafting the summary page and routing it for review, with a human approving each handoff.

Each layer has a different risk profile, a different data requirement and a different payback period. A boutique that buys a general chatbot licence and calls it an AI strategy has done the first layer badly and skipped the other two.

Why the economics are unusually strong

Investment banking is an industry where the most expensive people spend a large share of their time on formatting, searching and rework. Deloitte's analysis of the top 14 global investment banks, published in July 2023, estimated that generative AI could lift front-office productivity by 27% to 35% by 2026, with the investment banking division gaining the most, an average of 34%, ahead of equities and fixed income. You can read the full analysis in Deloitte's report on generative AI in investment banking.

Treat that number as a ceiling projected in 2023, not a guarantee. It does, however, explain why every major bank moved from experiments to firmwide deployment within two years.

What the Largest Banks Have Already Done

It helps to look at what the leaders have actually shipped, because it shows where the industry consensus is forming.

JPMorgan. According to CNBC's September 2025 report on JPMorgan's AI blueprint, about 250,000 employees have access to the bank's LLM Suite, excluding branch and call center staff, and roughly half use it every day. The platform connects to models from several providers and is updated about every eight weeks. The pitch deck demonstration is the example everyone quotes, but the strategic point is different: the bank treats AI as infrastructure, not as a tool each desk buys on its own.

Goldman Sachs. In June 2025 Goldman launched its GS AI Assistant firmwide after a pilot with about 10,000 employees, as reported by Fortune. The assistant runs behind the firm's firewall and gives employees access to several commercial models for tasks such as summarizing complex documents, drafting and translation.

The workforce signal. In January 2025 Bloomberg Intelligence surveyed chief information and technology officers at 93 large banks. On average they expected a net 3% reduction in headcount over three to five years, which would mean up to 200,000 roles across global banks, concentrated in back office, middle office and operations. Nearly a quarter expected cuts of 5% to 10%. The coverage is summarized by Bloomberg Law.

What this means for everyone else

Three conclusions follow for mid-market banks and boutiques:

• Client expectations are being reset by the bulge bracket. If a large bank can turn a first draft of a pitch book in minutes, a client will not pay for a boutique to take three days on the same deliverable.
• The moat is not the model. JPMorgan and Goldman both route across several commercial models. The models are available to you too. The advantage sits in proprietary data, templates, review processes and the speed of iteration.
• Governance arrived before regulation. Both banks built secure, internal environments first and opened access second. That order matters more than any single use case.

Where AI Creates Value Across the Deal Lifecycle

The most useful way to plan AI in investment banking is to map it onto the deal process you already run. Below is the lifecycle most advisory firms will recognize, with the realistic contribution of AI at each stage.

Stage 1: Origination and coverage

Coverage bankers spend a large part of their week tracking clients, sectors and triggers. AI is well suited to this because the task is mostly reading.

• Trigger monitoring. Summarizing filings, earnings calls, press releases and management changes for a coverage list, then flagging events that historically precede a transaction: a new CFO, a covenant amendment, a strategic review announcement.
• Target screening. Turning a qualitative thesis ("founder-owned industrial services companies with recurring maintenance revenue") into a structured screen across databases you already license.
• Meeting preparation. A two-page brief on a company, its recent news, its peers and the questions worth asking, produced in minutes rather than hours.

The value here is not that AI finds deals humans would miss. It is that a coverage banker can maintain a deeper relationship with more clients without adding analysts.

Stage 2: Pitching

This is where the time savings are largest and most visible. A pitch book is mostly assembly: market overview, trading comparables, precedent transactions, positioning, process options and a fee page.

AI can draft the narrative sections, populate standard pages from approved data sources, reformat into the house template and check that numbers are consistent across pages. What it cannot do is decide the angle. The managing director's judgment about why this buyer, why now and why this firm remains the product. AI simply removes the hours between that judgment and a reviewable draft.

A practical rule I recommend: every AI-generated page carries a source trail. If a number appears on a slide, the reviewer must be able to click through to where it came from. Without that, review time grows and the productivity gain disappears.

Stage 3: Due diligence

Data rooms are where AI moves from convenience to genuine analytical value. Typical uses:

• Contract extraction. Change-of-control clauses, exclusivity, termination rights, unusual indemnities and assignment restrictions pulled across hundreds of agreements into a single table.
• Financial anomaly detection. Flagging revenue recognition patterns, customer concentration shifts, working capital swings and one-off adjustments that deserve a question.
• Q&A management. Drafting answers from the data room on the sell side, and clustering and prioritizing questions on the buy side.

The diligence side of this is covered in more depth in my guide to AI for due diligence, which includes a checklist for data room reviews.

Stage 4: Valuation and modeling

Here caution is warranted. Large language models are not reliable calculators and should not be trusted to build a discounted cash flow model from scratch. Where they help:

• Populating templates from cleaned historical financials.
• Sensitivity narratives. Explaining in plain language which assumptions drive the valuation range, for the client or the committee.
• Model review. Checking formulas for hard-coded values, broken links and inconsistent time periods.
• Comparable selection support. Proposing a peer set with reasons, which the banker then accepts, edits or rejects.

The rule is simple: the model owner remains a human, and AI is a reviewer and a drafter, never the source of a number that goes to a client without a check.

Stage 5: Marketing materials and outreach

On the sell side, the confidential information memorandum, teaser and management presentation are long, repetitive documents built from the same facts. AI drafts them quickly and keeps them consistent. Buyer outreach lists can be ranked using past behavior, strategic fit and financial capacity, then reviewed by the deal team.

CNBC's coverage also reported that JPMorgan is training its system to help draft the long confidential memos the bank prepares for prospective M&A clients. That is the direction of travel for the whole industry.

Stage 6: Execution and closing

During execution, AI helps track process milestones, summarize bidder marks, compare draft agreements across rounds, maintain the issues list and prepare board materials. None of this is glamorous. All of it consumes junior hours at the worst possible time, late in a deal when fatigue drives errors.

Stage 7: Post-deal and knowledge capture

The least exploited stage. Every closed deal produces lessons about buyers, valuations, objections and terms. Most firms lose this knowledge in email folders. A searchable internal knowledge base, built from closed deals with proper access controls, compounds in value every year. My guide to AI for knowledge management covers how to structure this without creating a compliance problem.

Use Cases Ranked by Value and Risk

Not every use case deserves the same priority. The table below is the ranking I use with leadership teams. Value reflects time saved and quality improvement. Risk reflects regulatory, confidentiality and reputational exposure.

The pattern is clear: start internally, where errors are caught before a client sees them, and move outward only once review processes and audit trails are proven.

If you want an outside view on where your firm sits on this table and which three use cases to start with, you can request a consultation through the contact page on this site. I work with leadership teams on exactly this sequencing.

The Junior Banker Question

Every conversation about AI in investment banking eventually reaches the same question: what happens to analysts and associates?

The honest answer is that the job changes before it shrinks. The Bloomberg Intelligence survey points to back office, middle office and operations as the most exposed functions, not the front office. In the deal team itself, AI removes the hours that were never the point of the job: reformatting, chasing numbers, rebuilding the same page in a new template.

The real risk is the apprenticeship model

Banking has always trained people through repetition. An analyst who builds fifty comparable tables learns what a sensible multiple looks like. If AI builds those tables, how does the next generation develop judgment?

Firms that handle this well do three things:

1. Redesign training explicitly. Analysts still build models by hand during their first months, then use AI as a reviewer rather than a replacement.
2. Shift review upward. Juniors spend more time checking, questioning and explaining AI output, which is closer to the skill a senior banker actually needs.
3. Measure what juniors learn, not only what they produce. Staffing and reviews include judgment-based criteria, not just volume.

Hours, not headcount, is the first lever

For most boutiques, the realistic outcome over the next two years is not fewer analysts but fewer 100-hour weeks, faster turnaround and more deals per team. That is a recruiting advantage in an industry with a well-known retention problem.

Regulation, Compliance and Confidentiality

Investment banks operate under some of the strictest information rules of any industry. AI does not change those rules. It creates new ways to break them.

FINRA and US securities rules

In June 2024 FINRA published Regulatory Notice 24-09 on generative AI and large language models. It does not create new obligations. It reminds member firms that FINRA's rules are technology neutral and continue to apply, that generative AI could touch virtually every area of a firm's regulatory obligations, and that firms should evaluate tools before deploying them, whether built internally or bought from a vendor.

In practice that means:

• Communications with the public drafted by AI are held to the same content standards as any other communication.
• Recordkeeping applies to AI-assisted communications and the channels they pass through.
• Supervision must cover AI use, which requires knowing who uses which tool for what.
• Vendor due diligence applies to AI features embedded in software you already license, not only to new tools.

Material non-public information and information barriers

This is the risk that should keep a COO awake. An AI system with access to a shared document store can surface information from one deal team to another, across an information barrier, without anyone intending it. Retrieval systems do not understand walls unless you build them in.

Minimum controls:

• Permission-aware retrieval. The AI can only search documents the user is already entitled to see.
• Deal-level segmentation. Each live mandate has its own workspace, with access granted and revoked like any other wall-crossing.
• No training on client data by external providers, confirmed contractually.
• Logging of prompts and outputs for supervisory review, retained according to your recordkeeping policy.

European firms and the EU AI Act

For firms operating in the European Union, the EU AI Act applies alongside sector rules. Most deal-team uses, such as drafting, summarizing and searching, do not fall into the high-risk categories. Credit scoring of individuals does, and banks that also lend need to track that. After the amendments of Regulation (EU) 2026/1744, obligations for high-risk systems listed in Annex III apply from 2 December 2027, and transparency obligations apply from 2 August 2026. Article 4 asks providers and deployers to take measures to support the development of AI literacy among their staff, without imposing a specific level.

For a broader view of how to structure oversight, see my guide to AI agent governance.

A Playbook for Boutiques and Mid-Market Banks

The bulge bracket banks have budgets for internal platforms and dedicated AI teams. A 30-person M&A boutique does not. The good news is that it does not need them. The core ingredients are available off the shelf: enterprise versions of the main models, secure document workspaces and integrations with the data providers you already pay for.

Build, buy or configure

For a boutique, the answer is almost always configure. Building proprietary models is unnecessary. Buying a single vertical tool for everything creates lock-in. The practical path is an enterprise licence for a general model with strong data protection terms, connected to your document management system with permission controls, plus one or two specialist tools for data room review where volume justifies it.

My analysis of this trade-off in general terms is in the build vs buy decision guide. The banking version adds one requirement: any tool that touches client documents must pass the same vendor review as your virtual data room provider.

Your templates are your moat

The single highest-return investment for a boutique is not a tool. It is turning the firm's best work into reusable, AI-ready assets:

• A library of approved pitch book pages with clear structure and source fields.
• Sector primers maintained centrally and refreshed quarterly.
• A precedent transaction database cleaned and tagged by your own team.
• Standard diligence question lists by sector and deal type.

When these exist, AI produces drafts that look like your firm. When they do not, AI produces generic material that takes longer to fix than to write.

The sponsor relationship

Many boutiques live on private equity relationships. Sponsors are adopting AI quickly in their own processes, from sourcing to portfolio monitoring, as I describe in the AI for private equity guide. A banker who understands how the sponsor's deal team now works, and can deliver materials in the formats and timelines their tools expect, has an edge in winning the mandate.

What I Have Learned From Other Sectors

I have not run an investment bank. I have led AI and growth projects in sectors where the same structural problem appears: skilled people spending their time on assembly work instead of judgment. Two results illustrate the lesson that matters for banking.

A sports distribution company I worked with grew sales by 30% after we rebuilt its marketing around AI-driven targeting and content. The tools were not exotic. The gain came from redesigning the process so that people spent their time on decisions and relationships, and the machine handled volume.

A hotel I worked with moved from 9 million to 10 million in revenue. Pricing and demand analysis improved, but the real shift was that the team stopped reconciling spreadsheets and started acting on the numbers faster.

The translation to banking is direct. The firms that capture the 27% to 35% productivity range Deloitte projected will not be the ones with the most tools. They will be the ones that redesign the deal process around review, source trails and reusable assets.

Self-Assessment: Is Your Firm Ready?

Score each statement from 0 (not true) to 2 (fully true). Be honest. The value of the exercise is in the gaps.

Reading your score:

• 0 to 8: Foundation stage. Do not roll out tools yet. Start with policy, environment and permissions.
• 9 to 16: Pilot stage. You can run controlled pilots on internal use cases while closing governance gaps.
• 17 to 24: Scale stage. Expand to client deliverables and diligence, and start measuring productivity per deal team.

A broader version of this diagnostic for any business is available in my AI readiness assessment guide.

What It Costs and How to Measure the Return

Costs vary widely, so think in components rather than a single figure.

Cost components

1. Licences. Enterprise seats for a general model, typically priced per user per month, plus any specialist tools for diligence or document review.
2. Integration. Connecting the AI environment to your document management system, CRM and data providers with permission controls. This is usually the largest one-off cost.
3. Governance. Policy drafting, compliance review, logging configuration and vendor due diligence.
4. Content preparation. Building the template library and cleaning precedent data. This is internal time, often underestimated.
5. Training and change management. Short, role-specific sessions, repeated, with office hours in the first months.

A simple ROI model

The metric that matters is banker hours released per deliverable, multiplied by the share of those hours redeployed to revenue-generating work.

Example for illustration, using round assumptions you should replace with your own data:

• A deal team produces 40 pitch books a year.
• Each takes 60 analyst and associate hours today.
• AI drafting and checking cut that by a third, releasing 20 hours per book.
• That is 800 hours a year for one team, before counting CIMs, diligence and board materials.

Whether those hours become more pitches, better quality or fewer late nights is a management decision. All three have value. What does not work is measuring adoption by logins. Measure turnaround time, error rates found in review and mandates won per team.

What to track from day one

• Turnaround time from kickoff to first reviewable draft.
• Number of review cycles before a deliverable goes to the client.
• Errors caught by senior review, by type.
• Share of bankers using the environment weekly.
• Compliance incidents and near misses.

For finance leaders who want to connect this to budgeting, my guide to AI in FP&A shows how to build the business case into the annual plan.

The 30/60/90 Day Roadmap

This is the sequence I recommend for a mid-market bank or boutique starting from a low base.

Days 1 to 30: Foundations

1. Appoint an executive sponsor at managing director level and a working lead from operations or the COO office.
2. Draft and approve an AI use policy with compliance, covering permitted tools, data classes, client consent and recordkeeping.
3. Select and contract an enterprise AI environment with no-training terms and logging.
4. Configure permission-aware access for one practice group.
5. Baseline the metrics: turnaround times and review cycles for pitch books and CIMs.
6. Pick three internal use cases, such as meeting briefs, research summaries and formatting checks.

Days 31 to 60: Controlled pilot

1. Run the pilot with one or two deal teams, with weekly check-ins.
2. Build the first version of the pitch book template library with source fields.
3. Introduce AI drafting for pitch books on live mandates, with mandatory senior review.
4. Train analysts on verification: how to check numbers, sources and claims.
5. Log every issue found in review and fix the templates, not just the output.
6. Brief the managing directors on early results using the baseline metrics.

Days 61 to 90: Expand and harden

1. Extend to data room contract extraction on one sell-side process.
2. Add a second practice group, applying the lessons from the pilot.
3. Run a compliance review of logs, access patterns and policy adherence.
4. Decide which use cases go firmwide and which stop.
5. Set targets for the next two quarters: turnaround time, review cycles and adoption.
6. Plan the knowledge base from closed deals, with access controls designed in.

At day 90 you should have hard numbers, not anecdotes. That is the point where a leadership team can decide how far to go. If you would like help designing the pilot or reviewing your first 90 days, a consultation request through this site is the fastest way to start the conversation.

Twelve Questions to Ask Any AI Vendor Before Signing

Vendor selection is where many banks lose months. The demo always looks good. The contract and the architecture decide whether the tool survives a compliance review. These are the questions I would put to any provider before a pilot touches live deal data.

Data and confidentiality

1. Is our data used to train or improve your models, in any form? The answer must be no, and it must be in the contract, not only in a policy page that can change.
2. Where is our data stored and processed, and by which subprocessors? You need the full list, with locations, and notice of changes.
3. How long are prompts, files and outputs retained, and can we set the retention period? Your recordkeeping policy should drive this, not the vendor's default.
4. Can we export all logs in a format our supervision team can review? If logs stay locked inside the vendor's interface, supervision becomes manual and fragile.

Access and information barriers

5. Does retrieval respect the permissions in our document management system, in real time? A nightly sync is not enough when a wall is crossed or an analyst leaves a deal.
6. Can we create isolated workspaces per mandate and revoke access instantly? This is the practical test of whether the tool can live inside an information barrier.
7. How are administrator privileges controlled and audited? The people who configure the system can often see everything.

Quality and control

8. Does every answer cite the source document and location? Without citations, review takes longer than writing from scratch.
9. Which models run underneath, and can we choose or switch between them? Lock-in to a single model is a risk when capabilities change every few months.
10. How do you test accuracy on financial and legal documents, and can we see the results? Ask for evaluation on documents similar to yours, not generic benchmarks.

Commercial terms

11. What happens to our data and configurations if we leave? Exit terms matter more than entry discounts.
12. How is pricing structured as usage grows? Per seat, per document and per token models behave very differently once a firm scales from a pilot team to every banker.

A vendor that answers these clearly and in writing is usually one you can work with. A vendor that deflects on the first four questions is not ready for a regulated environment, however impressive the demo.

Seven Mistakes to Avoid

1. Starting with client-facing use cases. Errors in internal drafts cost minutes. Errors in client materials cost mandates.
2. Ignoring information barriers in retrieval. One permissions mistake can create a regulatory problem far larger than any productivity gain.
3. Letting each desk buy its own tools. Fragmentation makes supervision impossible and multiplies vendor risk.
4. Trusting AI arithmetic. Language models draft text well. Numbers need source data and checks.
5. Measuring logins instead of outcomes. Adoption without turnaround improvement is cost without return.
6. Skipping the template work. Without approved assets, AI output is generic and review time grows.
7. Forgetting the apprenticeship. A firm that stops training judgment saves hours this year and loses its next generation of senior bankers.

How AI in Investment Banking Will Evolve Over the Next Three Years

Predictions in this field age quickly, so I will limit myself to trends already visible in what the largest banks have deployed.

Agents with human checkpoints. The move from single prompts to multi-step workflows, such as gathering comparables, populating a template and drafting commentary, is already underway. The winning design keeps a human approval at each step that leaves the firm.

Proprietary data as the differentiator. As models become commodities, the firms with clean, structured records of their own deals, buyers and outcomes will get better results from the same tools.

Regulators asking sharper questions. FINRA has signaled it may issue more specific guidance on particular uses. Firms that already log, supervise and document their AI use will find that transition easy.

Smaller teams, more mandates. The economics push toward leaner deal teams handling more transactions, with senior time concentrated on client judgment and negotiation.

For the venture side of the market, where many of the same tools are being adopted for sourcing and screening, see my AI for venture capital guide. For the lending and retail side of the same institutions, the AI for banking guide covers credit, fraud and customer service.

FAQ

What are the main use cases of AI in investment banking?

The main use cases of AI in investment banking follow the deal lifecycle: coverage and trigger monitoring, pitch book drafting, data room contract extraction during due diligence, model review, drafting of CIMs and teasers, execution tracking and knowledge capture from closed deals. The highest value with the lowest risk usually comes from internal drafting and research summaries, which are reviewed before any client sees them. Client-facing and autonomous uses should wait until governance and logging are proven.

Will AI replace junior investment bankers?

AI changes the junior role before it reduces it. A Bloomberg Intelligence survey of 93 large banks in January 2025 found the most exposed functions are back office, middle office and operations, with an average expected net headcount reduction of about 3% over three to five years. In deal teams, AI removes formatting and assembly work. The bigger risk is to training: firms need to redesign how analysts build judgment when machines do the repetitive work.

How much productivity can AI add in investment banking?

Deloitte's 2023 analysis of the top 14 global investment banks projected front-office productivity gains of 27% to 35% by 2026, with the investment banking division averaging 34%. That is a projection, not a guarantee. Actual gains depend on template quality, data access and review processes. A practical way to measure your own return is banker hours released per deliverable, tracked against a baseline taken before the rollout.

Is it safe to put client data into AI tools?

It can be, under strict conditions. Use an enterprise environment with contractual no-training terms, keep data within approved infrastructure, enforce permission-aware retrieval that respects information barriers, and log prompts and outputs for supervision. Public consumer chatbots are not appropriate for client or deal information. FINRA Regulatory Notice 24-09 reminds firms that existing obligations on communications, recordkeeping and supervision apply fully to generative AI.

How should a boutique M&A firm start with AI?

A boutique should start with policy and environment, not tools. Approve an AI use policy with compliance, contract an enterprise AI environment with no-training terms, and pilot three internal use cases with one deal team for 30 days. In parallel, build a library of approved pitch book pages and clean precedent data. After 90 days, decide what to scale based on measured turnaround times and review cycles rather than impressions.

How long does it take to see results from AI in investment banking?

Internal use cases such as meeting briefs and formatting checks show results within the first month. Pitch book drafting with proper templates usually shows measurable turnaround improvements within 60 to 90 days. Data room extraction and knowledge bases take longer because they depend on integration and data preparation. A structured 30/60/90 day plan with a baseline makes the results visible to the leadership team.