How Investment Bankers Use AI to Manage Deal Flow in 2026

Investment bankers use AI for deal flow in three places: compressing research and first drafts, keeping relationship memory current without manual logging, and catching the follow-ups and stalled deals that slip under process pressure. Judgment stays with the banker, including valuation, which buyers to call, and whether a mandate is worth taking.

Which deal teams should read this before buying AI

If you run mandates at a boutique or lower-middle-market advisory firm, somewhere between two and thirty bankers, and someone on the team keeps asking which AI tool to try next, this is written for you. It assumes you already have a pipeline, an inbox and a calendar, and that the question is what to change, not whether AI exists.

It is also useful for an operations lead or chief of staff asked to write the firm's first AI policy, since two sections below are templates you can adapt.

It is not for bulge-bracket teams with a dedicated AI or data-science function; the surveys we cite describe your peers better than they describe ours. It is not a sourcing buyer's guide either. If your core problem is finding companies you have never heard of, start with AI deal sourcing for investment banks, which covers the data vendors in depth.

What the 2026 surveys actually show

Three data sets are worth knowing, and each has a limit that matters more than its headline.

Bain & Company. In Bain's Global M&A Report 2026, 45% of M&A executives reported using AI tools in M&A during 2025, more than double the prior year, and about a third said they were using it systematically or redesigning processes around it. The sample was 300 M&A executives, which reads as corporate and large-cap. "Used AI tools" is also a low bar: it counts a partner pasting a CIM into a chatbot the same as a firm that rebuilt its sourcing workflow. Deloitte's generative AI survey of large corporate and PE dealmakers points the same direction, and we walk through what Deloitte's respondents actually use AI for in our CRM comparison.

SS&C Intralinks. The most deal-specific data comes from SS&C Intralinks' AI in M&A Dealmaking benchmark study, a Q1 2026 survey of 400 senior deal professionals designed and analyzed by Reuters Insights, split into equal quotas across private equity, corporate acquirers, advisory firms and investment banks, venture capital and law firms. Its findings:

• 49% of dealmakers say AI is fully integrated across most deal stages and another 41% report partial integration; only one in ten are still in pilots.
• One third reported time savings of 21 to 30 percent in due diligence, and fewer than six percent reported savings above 50 percent in any phase.
• 64% of analysts and associates use AI regularly for valuation work, against 44% of partners and MDs; 12% of partners and MDs do not use it for valuation at all.
• 80% experienced an AI-related security or accuracy incident or near miss in the past 12 months. Access-control lapses led at 48%, followed by hallucinated outputs leading to inaccurate diligence at 40%. Among advisory firms and investment banks specifically, those figures rose to 60% and 53%.

Two caveats. Intralinks sells a data room with AI features, so it has a view on where AI should live. And a quota sample of senior professionals is not a census of boutiques.

Stanford HAI. For the wider business picture, the Stanford AI Index 2025 reports that 78% of organizations used AI in 2024, up from 55% the year before, and that the cost of running a model at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024. The second number matters more than it looks: AI that was too expensive to run on every email and call two years ago is now cheap enough to run on all of them, which is why capture-based features have spread so fast.

Where we disagree with the survey framing

Intralinks reads its own data as "the market demands specialization, not consolidation": 49% of respondents prefer multiple specialized AI tools, while 20% want a single integrated platform. For a firm with an IT team and a procurement function, that preference makes sense.

For a boutique we think it is the wrong default, and the same report supplies the reason. The leading incident type, access-control lapses at 48%, grows with every tool that holds a copy of deal data. A four-banker firm running five AI tools has five vendors' retention terms to read and five places a CIM can leak. It also has five tools that each know a slice of the relationship and none that knows all of it. Specialized tools win on depth per task. At boutique scale, we would trade some of that depth for fewer places confidential data lives and one record of what was said to whom. The report's authors concede part of this, noting that the next efficiency gains will likely come from connecting tasks into continuous workflows rather than adding more tools.

The constraint is no longer access to AI. It is trust in the data going in, and control over where it goes.

What is AI deal flow management?

AI deal flow management is the use of AI models to capture, organize, summarize and act on the activity behind a firm's pipeline of mandates and targets, from first outreach through close. What makes a CRM genuinely AI-native is where the AI sits in the workflow, not whether it exists.

AI across the deal lifecycle

The table below maps what AI does well at each stage, what a banker still has to check, and the kind of tool that typically does the work. Examples illustrate a category, not an endorsement, and several categories overlap.

The "verify" column never goes empty. Read down it and a pattern shows: AI is most reliable where it summarizes the firm's own logged activity, and least reliable where it predicts something about the outside world.

Where AI helps, and the research on why

The best evidence on AI and professional work does not come from M&A surveys. It comes from controlled experiments, and three of them map closely onto a deal team.

The frontier is jagged. In an experiment BCG ran with its own consultants, described in How People Can Create and Destroy Value with Generative AI, around 90% of participants improved their performance on creative ideation with GPT-4, converging on a level 40% higher than those working without it. On a business problem-solving task outside the model's strengths, participants using GPT-4 performed 23% worse than those working without it. The lesson for a deal team is uncomfortable: the tasks where AI hurts do not announce themselves, and a confident answer looks the same in both cases. A buyer-list brainstorm sits inside the frontier. Deciding which of two LOIs is actually better for the seller, with a rollover and an earnout on one side, does not.

Juniors gain most. An NBER study of 5,179 customer support agents, Generative AI at Work by Brynjolfsson, Li and Raymond, found a 14% productivity increase on average and 34% for novice and low-skilled workers, with minimal impact on the most experienced. Support work is not banking, but the shape carries over. An analyst with an AI assistant gets closer to a senior associate's first draft faster. A managing director gets little from the same tool, which matches the Intralinks finding that senior bankers use AI far less for valuation.

We take a firm position on what follows from those two studies together. The people who gain most from AI are also the least equipped to spot when it is wrong. A firm that hands juniors AI tools without a review rule is optimizing for speed on exactly the work most likely to go sideways.

Polished output can cost more than it saves. Research from BetterUp Labs and the Stanford Social Media Lab, first written up in Harvard Business Review, names the problem "workslop": AI-generated work that looks polished but lacks substance and leaves colleagues to do the real thinking. It found 40% of the 1,150 U.S. desk workers surveyed had received workslop in the past month, and that each instance took close to two hours to sort out. We agree with the diagnosis and would sharpen it for M&A. On a deal team, workslop does not just waste a VP's afternoon. A teaser paragraph with an unverified margin figure or a buyer email that misstates a process date can cost the firm credibility with a counterparty it will see on the next ten deals.

Put those three findings against the deal work itself and three areas stand out. Sourcing is left out on purpose; the lifecycle table and the sourcing guide linked above cover it.

Outreach

AI drafts first touches and follow-ups from what the CRM knows about each contact, and this is where most firms feel time savings first. The risk is volume without judgment: a cadence that reads like software burns a small market fast, because owners in a niche talk to each other. The craft side is covered in running high-volume outreach without losing the human touch.

Relationship memory

The hardest operational problem in deal flow is remembering context across dozens of slow-moving relationships: who said what, what you promised and when you last spoke. AI that reads email, calendar and call transcripts can keep that record current without anyone typing it in. Calls and meetings feed it most. For the calling side, see what a power dialer built for investment banking needs; for turning that history into a brief before a management meeting, see AI deal briefs for investment banking.

This is where we think boutiques underinvest. The failure described in why good deal relationships go cold is not a decision. Nobody chooses to drop an owner who said "call me after the busy season." The call just never happens, and eighteen months later a competitor runs the sale.

Follow-up and pipeline management

AI can flag quiet deals, propose a stage change when an email suggests movement, and surface open commitments. This ambient layer is worth more than it sounds, because stale pipeline data quietly corrupts every forecast built on it. If pipeline hygiene is the real problem, start with what deal tracking software for investment banking should track.

What can AI not do in M&A?

AI cannot build the trust that gets a founder to sign an engagement letter, judge management quality, read a seller's real motivation, or decide what a business is worth to a specific buyer. It also cannot guarantee accuracy. Every AI output that informs a decision or reaches a client needs a human check against the source.

Relationships close deals, not software. The moment that wins a mandate or saves a wobbling LOI is a call or a meeting. AI prepares the banker for it and follows up after. It does not have the conversation. Salt Creek Advisory, the M&A advisory firm run by HelmIQ's founder, draws the same line from the sell side in how AI is actually changing business: it uses AI for research, sourcing and workflow, while negotiation, the read on a buyer or seller, and accountability for the outcome stay with its principals.

Qualitative risk stays human. Management depth, customer concentration the data room understates, the real reason a founder wants out: these calls come from watching enough deals go sideways.

Models state wrong things confidently, and bad data makes it worse. If contacts are stale and stages are logged inconsistently, AI will surface bad information with a straight face.

Four bankers, six mandates: where the hours come back and where they leak

Picture a small sell-side shop with more live processes than people, and run the math on what AI actually returns once review is counted. Every number here is an assumption chosen to show the arithmetic, not a measurement from any firm, so swap in your own before drawing a conclusion.

The firm. One managing director, one director, one associate and one analyst. Six live sell-side mandates, plus an origination list of about 300 owners.

Where the hours go today (assumed):

On these assumptions the firm gets back about 13 hours a week. That is a meaningful slice of one banker's week, but across a four-person team working 50 hours each it is only about 6% of total hours. The headline savings in the surveys describe single phases like diligence, not a whole team's week, which is why the two numbers look so different. Our table is arithmetic, not evidence.

Now the cost side, which most rollout plans skip. Suppose two AI drafts a week reach a VP or the MD with an error that has to be traced and fixed. At the rework cost from the workslop research (close to two hours per instance), that is about four hours a week, close to a third of the gain. Suppose one of those errors instead reaches a buyer. The cost is no longer measured in hours.

What the week looks like when it works:

• Monday. The analyst spends the morning on the dialer. Before each call a brief shows talking points and likely objections. Two owners agree to follow-up calls, and the transcripts land in the CRM without anyone typing a note.
• Tuesday. A buyer emails its indication of interest on one mandate. The CRM proposes moving the deal from Engaged to IOI Received. The director accepts with one click, or dismisses it if the email was ambiguous.
• Wednesday. An hour before a management meeting, the MD reads a brief on each attendee already in the CRM: recent emails and calls, open items and questions to raise, each point traceable to the record it came from.
• Thursday. On Monday's call, someone said "I will send the one-pager by Wednesday." Nothing went out. The system raises a task quoting the transcript line and asks the analyst to confirm whether it was their promise or the owner's.
• Friday. A mandate has not moved in three weeks. The MD calls the client before the client calls them.

The conclusion we draw: the gain is real but modest, and sloppy review eats most of it. The firms that come out ahead treat review time as part of the workflow, not as overhead to cut next quarter.

Regulators already have a view

Most AI policy discussion at boutiques starts with "what could go wrong." It is faster to start with what regulators have already said.

FINRA. If your firm is a FINRA member, Regulatory Notice 24-09 is the document to read. It creates no new rules. It says FINRA intends its rules to be "technologically neutral," that they apply to generative AI "just as they apply when member firms use any other technology or tool," and that the content standards of Rule 2210 on communications with the public apply "whether member firms' communications are generated by a human or technology tool." In plain terms, an AI-drafted teaser is your teaser.

The SEC. In March 2024 the SEC charged two investment advisers, Delphia and Global Predictions, with making false and misleading statements about their use of AI, with $400,000 in combined penalties. Neither was an M&A firm, but the principle travels. If your pitch book says "AI-driven buyer matching," make sure it describes something you actually do.

Recording law. Transcription starts with a recording. Federal law, 18 U.S.C. 2511, permits recording with one party's consent, but some states are stricter. California Penal Code 632 requires the consent of all parties to record a confidential communication. A firm calling owners nationwide needs a rule that works in the strictest state it dials into. This is general information, not legal advice.

Frameworks. For a firm that wants a structure rather than a blank page, the NIST AI Risk Management Framework is voluntary and publicly available, and NIST released a Generative AI Profile for it in July 2024. Its full process assumes risk and governance staff most boutiques do not have, so a small firm should borrow its categories rather than its process.

Is it safe to put deal data into ChatGPT?

It depends on the version and the terms. A consumer chat account is the wrong place for client names, financials or anything under NDA unless your firm has confirmed how inputs are stored and used. Business or enterprise tiers, or tools under a signed agreement that rules out training on your data, are the defensible route.

Confidentiality is the product in M&A: a leaked process name can reach a client's employees or competitors before the deal is ready. The access-control failures in the Intralinks data (AI agents given permission over too much information, or making unwanted changes) are the modern version of the misdirected email.

Template: a one-page firm AI policy

Copy this, fill in the brackets, and have every banker sign it.

1. Approved tools. Only [list tools] may process confidential deal data. Each is covered by a business agreement reviewed by [name] on [date].
2. No confidential data in consumer tools. Client names, codenames, financials, CIMs and buyer lists never go into a personal or free AI account.
3. Training and retention on file. For each approved tool we record, in writing from the vendor, whether inputs train models and how long they are kept.
4. Human review before anything leaves the firm. No AI-drafted email, teaser, memo or buyer communication is sent without a named banker reading it in full.
5. Sources for factual claims. Any AI summary that feeds a decision names the record behind it: an email, a call, a filing or a document. Unsourced claims are treated as unverified.
6. Numbers checked against the document. Figures extracted by AI are verified against the CIM, model or financial statements before they appear anywhere else.
7. Recording consent. Calls are recorded only under a notice rule that satisfies the strictest state we call into, and we can show which calls carried notice.
8. Marketing claims match reality. We describe our use of AI to clients only in terms we could demonstrate.
9. Access is scoped. An AI tool or agent gets access only to the deals and mailboxes it needs, and we review that list quarterly.
10. One named owner. [Name] reviews this policy whenever tools or vendor terms change, and at least twice a year.

How do you use AI for deal flow management?

Start with the bottleneck, not the tool. Pick the one place time disappears each week, apply AI there with a review step, and measure it for a quarter. Firms that bolt AI onto every process at once usually get more noise, not more deals.

1. Write the policy first. Use the template above and name its owner before anyone uploads a CIM.
2. Fix the pipeline definitions. Agree what moves a deal from Engaged to IOI Received before letting AI propose it. Inconsistent stage history produces inconsistent suggestions.
3. Automate capture before generation. Calls, emails and meetings have to land in the CRM on their own, because drafting and briefing run on that history.
4. Add briefs and follow-up tracking. They carry the best value for the risk, since they summarize your own records and a banker reads them before acting.
5. Then scale outreach. Wait until the record is clean and the team has edited enough drafts to trust them.

Checklist: measure it before you believe it

Record these for four weeks before rollout, then for a quarter after.

• Hours per banker per week spent logging activity
• Minutes of prep per external meeting
• Follow-ups that slipped past their promised date, per month
• Deals with no logged activity in 21 days
• AI drafts that needed a factual correction, per week
• AI errors that reached a client, buyer or owner (target: zero)

If the first four improve and the last two stay low, keep going. If the first four do not move, the cause is usually unclean pipeline data or a tool that sits outside the daily workflow, not the model.

What AI tools do M&A advisors use?

Most advisors combine a general assistant (ChatGPT or Claude) for drafting and analysis, a data source for company screening, a notetaker or conversation-intelligence tool for calls, and a CRM that increasingly embeds its own AI. In the Intralinks study, the most common stack was three to five AI tools at once (52 percent of respondents).

General assistants. ChatGPT and Claude handle ad hoc drafting and summarizing public information. They sit outside the workflow, so they know nothing about your pipeline unless someone pastes it in: a context problem and a confidentiality problem.

Research and data. AlphaSense, Grata and PitchBook feed screening and sector work; Clay enriches prospect lists before they reach a CRM.

Conversation intelligence and notetakers. Gong and Chorus were built for sales teams. Bankers use them for recording and summaries, though their scoring and templates assume a SaaS sales motion rather than an M&A process. Lighter notetakers such as Granola and Fireflies are popular for meeting capture.

Deal CRMs with embedded AI. Competition is sharpest here, and every serious vendor now has AI. Intapp markets zero-entry activity capture, conversational AI and agentic playbooks for DealCloud. Affinity markets Affinity Ascend, whose agents prep meetings, capture conversations and write updates back to the pipeline. Automatic capture is table stakes. The real differences are whether the AI is built for M&A execution, and whether the execution layer (dialer, data room, CIM screening, structured outreach) lives in the same system or in three other tools.

What does AI inside an M&A CRM look like?

In a CRM built for deal work, AI runs on the firm's own records: it preps you before calls, suggests a stage change when an email linked to the deal shows movement, flags promises nobody kept, and screens documents. Good systems propose; the banker decides.

HelmIQ is the platform we build, so weigh this section accordingly. Four examples of what its AI does today, mapped to the lifecycle table above:

• Calls: a pre-call brief in the power dialer. Recent emails, call transcripts, notes and deal context become talking points, objection prep and questions to ask. Meeting briefs work the same way, and each takeaway cites the record it came from; AI deal briefs explains the mechanics.
• Follow-up: commitments pulled from call transcripts. A daily pass looks for promises such as "I'll get you the customer list after the board meeting." If the date passes with no matching email or call, it raises a task that quotes the transcript line. Because the commitment check does not yet use the speaker labels on dialer transcripts, the task asks the banker to confirm whose promise it was rather than asserting it.
• Pipeline: the deal's stage is only ever suggested. An email linked to a deal can produce a suggested single-step forward move (for example Engaged to IOI Received). The deal's stage does not change until a banker accepts. Individual buyer stages inside a sell-side process work differently: a hard event such as a signed NDA can move a buyer automatically, as can a buyer-stage move the firm has accepted 30 times in a row without undoing it, and every automatic move can be undone.
• Recording: the safe default. HelmIQ's dialer announces recording on every call by default. A firm can change that only by recording who accepted the risk, and each call log stores which notice rule applied.

HelmIQ records calls made from its own dialer and transcribes them when the firm's AI features are on. It does not record Zoom or Google Meet video meetings; meeting notes come in through Granola or Fireflies imports.

For how that stacks up against DealCloud, Affinity and the rest, feature by feature, see the best AI CRMs for investment banking, compared.

Who should not choose HelmIQ. If your main pain is finding new companies at scale, a sourcing database will do more for you than any CRM, and HelmIQ's AI owner and company discovery is still in development. If you prefer the best-of-breed approach nearly half of Intralinks respondents favor, and have someone to manage the integrations, point tools may suit you better. If your clients or compliance team require a SOC 2 report today, we do not have one yet; it is planned. And if you have a dedicated CRM administrator and value deep configurability over speed to value, an enterprise platform may still fit better. The best CRM for investment banking buyer's guide weighs these trade-offs.

What adoption surveys and lab experiments can and cannot tell a boutique

None of the research above watched a six-person advisory firm run a sale process, so treat it as a compass, not a map. The adoption numbers come from big-company samples: Bain's 300 M&A executives, and Intralinks' 400 senior deal professionals in fixed quotas, surveyed by a vendor that sells AI-enabled deal software. The BCG and NBER results were measured on consultants doing set tasks and on customer support agents, and both predate today's models. The workslop figures come from 1,150 U.S. desk workers surveyed in September 2025. The FINRA, SEC and statute text is primary, but how it applies to your firm is a question for your counsel.

The rest is judgment, formed building HelmIQ and from sell-side work at Salt Creek, not measured in any study: the ranking of memory over target-finding, the case for fewer tools, the review rule for juniors, and the capture-first rollout order. The four-banker week uses assumed numbers throughout.

Frequently Asked Questions

What is the first AI workflow a boutique investment bank should adopt? Capture first, then briefs; see the rollout order above for why.

Should AI send outreach emails to business owners automatically? Not without a banker's review. AI can draft first touches and full cadences quickly, but an owner who receives a clumsy or factually wrong email rarely offers a second chance. The workable model is that AI drafts, a banker edits and approves, and the system handles timing.

How do you stop AI from inventing facts in deal materials? Require sources. Use tools that link each claim to its source record, and treat anything unlinked as unverified. Re-check financial figures against the original CIM or statements before they appear in a teaser, memo or buyer communication.

Can AI help coach junior bankers on cold calls? Yes, if calls are recorded and transcribed. AI can grade a call against a firm's rubric, such as whether the banker gave a reason for calling, asked about the business before pitching and asked for a meeting. A senior banker should still review the grades and do the coaching.

Does AI change what analysts do on a deal team? It moves analyst time away from logging activity and first drafts, toward checking outputs, building models and preparing for owner and buyer conversations. The analyst becomes more of an editor, which still requires knowing what a good memo looks like.

Should a firm tell clients it uses AI on their deal? We think yes, in plain terms and only for what you actually do. FINRA members already answer for AI-drafted communications under Rule 2210, and the SEC's 2024 cases show the risk runs the other way too: overstating AI is itself the violation. A short line in the engagement letter naming the approved tools and the human review step is easier to defend than silence or a pitch-book slogan.

Next step

Pick one bottleneck from the checklist above, measure it for four weeks, and write the one-page policy before you add a tool. If the bottleneck turns out to be capture, briefs or follow-through, those now ship inside the deal CRM, and our investment banking CRM guide compares the platforms that do it side by side. HelmIQ is $249 per banker per month with everything included. Twilio telephony usage is billed separately, and because each dialer call bridges through the banker's own phone, a call bills two outbound legs. Sign-up is currently by access request, and once you are in, you import your existing contacts and deals yourself.

Related articles