AI Agents vs Human Expertise in M&A: Where Does the Human Premium Sit Now?

    AI Agents vs Human Expertise in M&A: Where Does the Human Premium Sit Now?
    Simone Vascotto··7 min read

    Key Takeaways

    • 1.When everyone has the same AI, the advantage shifts to the people who know what to ask it. Shared access widens the gap between practitioners rather than closing it.
    • 2.The human premium in M&A now sits in framing: starting from the investment decision and the assumptions behind the price, not from the diligence checklist.
    • 3.Agents do best on volume and verification work across the deal lifecycle. Humans stay essential for framing, valuation judgment, negotiation, reading people and owning the decision.
    • 4.Set agent autonomy per task, from human-led to run within bounds, and put the human checkpoint before the agent runs as well as after.
    • 5.Framing skill has to be built and captured deliberately. Otherwise the firm's AI investment mostly makes individual practitioners more productive, and that capability can leave with them.

    Short answer

    As AI agents become standard across deal teams, execution stops being the differentiator. Agents can screen targets, read the data room, extract clauses and draft the plan. The human premium moves upstream, to framing: what has to be true for the price to make sense, which assumptions the buyer is taking on trust, and which risks nobody has put in scope. In a working human-in-the-loop model, people frame and scope the question, the agent executes, and a named person checks and signs.

    Where does the human premium sit when every deal team has AI?

    James O'Dowd, Founder and CEO of Patrick Morgan, a talent and advisory firm for the professional services sector, recently put the question sharply in a LinkedIn post. Professional judgment has long been the reason clients pay a premium for human expertise. But, he writes, "once a model can raise a question an experienced practitioner might have missed, I think we need to be more precise about where that premium sits."

    Deal teams face the same question, and faster. Bain's 2026 M&A Report says AI adoption among M&A practitioners "more than doubled in 2025," with 45% of more than 300 executives surveyed now relying on it (Bain & Company). Deloitte's 2025 GenAI in M&A Survey of 1,000 US corporate and private equity leaders found 86% had integrated generative AI into their M&A workflows, with due diligence among the top uses at 35% (Deloitte).

    And the AI is becoming more autonomous. Anthropic describes agents as systems "where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" (Anthropic, Building effective agents). It can work through a data room and draft the memo without a person prompting each step.

    You might expect this to level the field. James argues the opposite: "Widespread access to the same technology is set to widen the gap between practitioners." When every team can produce a diligence pack in an afternoon, the question that shaped it becomes the advantage.

    What does first-principles framing look like on a deal?

    James draws the line between two kinds of practitioner. One understands the diligence brief well and executes it fast. The other starts with the investment decision itself:

    "What has to be true about this business for the price to make sense? Which of those assumptions has the buyer taken on trust?"

    From there, he writes, they can use the model to investigate a risk that neither the client nor the engagement partner thought to put in scope.

    That maps directly onto how a deal team should use agents:

    1. Write the "what must be true" list first: revenue retention after change of control, the synergy case, the teams the plan depends on.
    2. Mark what is being taken on trust: management's growth story, a customer's stated intent to renew.
    3. Scope the agent against those assumptions, not just the checklist. Point it at the evidence that would confirm or break each one, including areas outside the standard request list.

    Agents execute a scope. They don't decide whether it is right. As I argued in Diligence AI Has a Ceiling, the next advantage is in the questions nobody asks.

    AI agents vs human expertise: a stage-by-stage comparison

    Read the human column for a pattern. Almost all of it is framing, interpretation and ownership.

    Stage / taskWhere AI agents do wellWhere human expertise is essentialSuggested autonomy
    Sourcing and screeningScanning large target universes and first-pass screens against thesis criteriaThe thesis, strategic fit, founder relationshipsExecute with checkpoints
    Valuation and modelingComps, model population, sensitivities, peer benchmarkingWhich assumptions matter, whether a benchmark applies, what the asset is worth to youDraft only
    Diligence: document and data room reviewIndexing, summarizing and flagging missing itemsDeciding what is material for this deal and what to ask nextRun within bounds
    Diligence: contract clause extractionFinding change-of-control, assignment and termination clauses at scaleInterpreting intent, weighing exposure against priceExecute with checkpoints
    Diligence: financial anomaly spottingReconciling ledgers and tying numbers across documentsWhy the anomaly exists and whether it changes quality of earningsExecute with checkpoints
    Diligence: management interviews and reading the roomQuestion lists, transcripts, summariesRunning the conversation, noticing hesitation, testing rehearsed answersHuman-led
    Diligence: cultural and behavioral assessmentStructuring evidence and summarizing patterns at team levelInterpreting how teams work together and what it means for the plan; never judgments about named individualsHuman-led
    Negotiation and deal structuringTerm sheet options, SPA markup comparison, earnout scenariosReading the counterparty, trading concessions, choosing the structureDraft only
    Integration planning and executionDraft 100-day plans, milestone tracking, status reportsSequencing around people, leading change, adjusting the planExecute with checkpoints
    Accountability and liabilityLogging actions, sources and uncertaintySigning the memo and the decision, and answering for itHuman-owned
    Hallucination and verification riskCiting sources, labeling evidence versus inferenceChecking citations against the sourceHuman check on every output used

    How does the split play out across the deal lifecycle?

    Sourcing, screening and valuation. Agents can watch a far larger target universe than an analyst and handle comps, model population and sensitivities. Benchmarking is not judgment, though. Whether a peer multiple applies, and what the asset is worth to your firm specifically, are human calls.

    Diligence. Data room review, clause extraction and anomaly spotting suit agents, provided every finding links back to its source. The risk running through all of it is hallucination. Stanford RegLab and HAI researchers found that leading legal research tools built to reduce hallucinations still produced incorrect information more than 17% of the time (Lexis+ AI and Ask Practical Law AI) and more than 34% (Westlaw AI-Assisted Research), including "misgrounded" answers that cite a real source that doesn't support the claim (Stanford HAI). Management interviews stay human. Cultural and behavioral assessment stays at team level, never pointed at individuals. In Europe that is also a legal line, since the EU AI Act prohibits AI systems that infer emotions of a natural person in the workplace, except for medical or safety reasons (EU AI Act, Article 5).

    Under the Hood of M&A

    Get articles like this one weekly.

    Subscribe on Substack

    Negotiation and structuring. Agents can draft term sheet options, compare SPA markups and model earnout scenarios. They can't read the counterparty. Agents draft. People negotiate.

    Integration. Agents turn findings into a first-draft 100-day plan and track milestones. Sequencing around people, leading change and deciding when to adjust the plan stay with the integration lead.

    What does a human-in-the-loop operating model look like?

    Most human-in-the-loop models put the human at the end. James's argument puts the most valuable human step at the start, so the model needs a checkpoint on both sides of the agent.

    Before the agent runs: frame and scope.

    A named deal team member writes the "what must be true" list, marks the assumptions taken on trust and sets the agent's scope, including at least one question outside the standard checklist.

    While it runs: set autonomy per task.

    • Human-led. The agent prepares material; a person does the work. Interviews, negotiation and anything about people.
    • Draft only. The agent drafts; a person rewrites and decides. Valuation assumptions, term sheets, IC memo sections.
    • Execute with checkpoints. The agent runs multi-step work and pauses for review. Screening, clause extraction, anomaly spotting, integration tracking.
    • Run within bounds. The agent runs unattended on narrow, verifiable tasks, and a person samples the output. Indexing, classification, status consolidation.

    After it runs: verify and sign.

    • Calibrate before you trust. The American Bar Association's Formal Opinion 512 gives a practical model: a lawyer relying on AI to summarize lengthy contracts would not necessarily have to review every document if they had first tested the tool on a smaller subset (ABA Formal Opinion 512). Test each agent on a closed deal first.
    • Require sources for every finding, and separate evidence from inference.
    • Name an owner per output. Opinion 512 says AI tools "cannot replace the judgment and experience necessary" for lawyers to advise clients. The same holds on a deal: the person who signs owns the answer.

    See also What Claude Cannot Do in M&A and the division of labor between AI and humans.

    Who owns the advantage: the firm or the deal professional?

    James raises an ownership question for any firm investing in AI platforms. A firm that can take what one practitioner discovers on a deal and change how hundreds of colleagues approach similar decisions has built something of its own. But where the edge comes from one person's market knowledge and first-principles reasoning, the platform may simply have made that person more productive. In his words: "A fair amount of that capability can leave with them."

    He connects this to partner economics: a partner who delivers more with a smaller team and technology available anywhere has a stronger claim on the economics of their work.

    For deal teams, the answer is to capture framing, not just execution: turn "what must be true" lists, out-of-scope questions that paid off and post-mortem lessons into shared workflows. That is the part of the advantage a firm can own.

    How do deal teams build framing capability?

    James points to a gap in how professional services firms develop people. Many processes reward recognizing a familiar problem, applying an established framework and producing a convincing answer in the time allowed. As he puts it: "Someone can score exceptionally well on all three without ever having to decide whether the question was worth asking."

    For deal teams, this is a capability-building question, not a judgment on anyone. Practical steps at team level:

    • Practice framing on closed deals. Before reviewing what diligence found, have the team write what would have had to be true for the price to make sense.
    • Review the question, not just the answer. In deal reviews, ask what was scoped out and why.
    • Pair junior and senior people on scoping, since the tools to act on a better question are now in everyone's hands.

    The deal teams that get the most from agents won't be the ones that automate the most. They'll be the ones that know exactly which question to hand the agent, and where the human sits.

    About the Author: Simone Vascotto is the Founder of HumanAQ, the AI-powered platform for Cultural Due Diligence and leadership alignment in M&A.

    This piece is part of the Under the Hood of M&A newsletter.

    Frequently Asked Questions

    Where does human expertise add the most value when deal teams use AI agents?
    In framing: deciding what has to be true for the price to make sense, which assumptions are being taken on trust and which risks are out of scope, then checking and owning what the agent produces.
    What is first-principles thinking in due diligence?
    Starting from the investment decision rather than the request list. You identify the assumptions the price depends on, then point diligence, including AI agents, at the evidence that would confirm or break them.
    What is human-in-the-loop in M&A?
    An operating model where a person frames and scopes each AI task, each task has a set autonomy level and review checkpoints, findings require sources, and a named person owns the output before it reaches an IC memo, term sheet or plan.
    How do you manage hallucination risk when using AI agents in diligence?
    Test agents on a closed deal, require a source for every finding and check citations against the original, because a real citation can still fail to support the claim.