AI in M&A: where it helps and where human judgment must remain

Legal, tax or valuation opinion

A model can retrieve a clause or explain a method. It must not present itself as appointed counsel, tax adviser or professional valuer.

Investment decision

Risk summarization is not a recommendation to buy. The acquirer retains the decision and must understand evidence limitations.

Offers and terms

AI may draft. It must not send an offer, change price or accept a counteroffer without explicit action by an authorized party.

Risk management matters as much as the model

The NIST AI Risk Management Framework is voluntary, non-sector-specific and designed to help organizations manage AI risk. Its Generative AI Profile adds considerations specific to generative systems. For M&A, the operational lesson is to govern the use case, know the data, measure quality, document limitations and retain human control where the consequence is material.

Every material answer should be an evidence card

A user should be able to see:

Answer → Source → Page/object → Date/version → Confidence → Limitation / linked issue

An unsupported answer is a lead, not an established fact. When accessible documents are insufficient, the correct response is to say so.

Diligence-specific risks

Hallucination

The model fills a gap with invented content. Control: source requirement and refusal to conclude without evidence.

Permission leakage

A summary reveals information from a restricted document. Control: permission-aware retrieval and leakage tests.

Stale evidence

AI cites a superseded version. Control: versioning, dates, supersession and canonical-source logic.

Automation bias

Users accept a polished answer too readily. Control: confidence, counterevidence and easy verification.

Prompt injection

A document attempts to instruct the assistant. Control: separate trusted instructions from document content and monitor execution boundaries.

Automate, assist or reserve to humans

The long-term moat is not “having AI”

The defensible advantage is connecting AI to clean, permissioned, dated transaction evidence while preserving human responsibility. A powerful assistant operating on poorly governed data mainly accelerates error.

Limitations

Models, standards and threat patterns change quickly. Sensitive AI functions should be re-reviewed when models, providers, retrieval architecture or permission controls change materially.

<!-- ENHANCEMENT_V1 -->

Design a real human-in-the-loop

“Human validation” should not mean clicking OK after an opaque answer. For sensitive decisions, the reviewer should see the question, sources used, cited passages, documents excluded by permission, uncertainties, deterministic rules applied and proposed action.

The system must also know when not to propose an action because evidence is insufficient.

Anatomy of an acceptable diligence AI answer

1. Short answer, what the permitted evidence supports.
2. Sources, document, version, page/section.
3. Status, verified / management assertion / contradiction / insufficient information.
4. Confidence, evidence quality, not psychological certainty.
5. Linked issue, price, SPA, CP, financing, integration or investigation.
6. Next question, what evidence would reduce uncertainty.

What to test before deployment

Can AI retrieve a revoked file? Does Clean Team segregation survive retrieval? Do citations point to the right version? Can malicious document content inject instructions? Can users distinguish extraction from interpretation? Are sensitive prompts/outputs logged appropriately? Is there a non-AI route for critical work?

Vendor due diligence questions

Ask about data architecture, tenant isolation, customer-data training, retention, localisation, encryption, logs, subprocessors, incident response, models, evaluation, deletion, permission propagation, export and reversibility.

A compelling demo on public documents does not prove security in a live VDR.

The moat is not the model

The same foundation models increasingly become available to many platforms. Durable differentiation comes from permission quality, evidence normalisation, issue taxonomy, version history and the ability to connect an answer to transaction decisions.

AI becomes more useful when the evidence architecture exists before the AI layer.

DealPME responsibility matrix

That boundary protects both the user and the credibility of the product.

<!-- ENHANCEMENT_V2 -->

Govern model and vendor change

An acceptable answer today can change after a model update. Sensitive AI functions should therefore preserve model version, date, prompt policy, evaluation sets and release criteria. A model or provider change should trigger permission, citation, hallucination and prompt-injection tests before production use resumes.

Measure quality at task level

A single “AI satisfaction” score hides important failure modes. Measure extraction accuracy, classification, citation correctness, contradiction detection, refusal when evidence is insufficient and permission compliance separately. For tasks influencing price, SPA, CPs or financing, use test cases with expected answers and adversarial counterexamples.

A practical governance loop

MAP the transaction task and decision consequence. MEASURE evidence quality and model behaviour. MANAGE permissions, escalation and monitoring. REVIEW incidents, false positives and user overrides. This is consistent with the risk-management logic in the NIST AI RMF family without treating the framework as a transaction-specific rulebook.

The key governance question is not whether AI can produce a plausible answer. It is whether the system can show when that answer should not be trusted or acted upon.

Procurement should include exit planning

Before adopting an AI provider, define how prompts, embeddings, indexes, logs and derived artefacts can be deleted or exported; what happens when a user revokes a document; whether data are used for model training; and how service termination affects access to prior results. Vendor lock-in is not only a commercial issue when the tool sits inside a confidential transaction workspace.

Human judgement should be concentrated where consequences are high

Not every extraction needs a senior reviewer. The governance model can scale review intensity with consequence. Low-risk classification may be sampled; a clause comparison feeding an SPA issue should be reviewed by the relevant professional; a certification or admission decision remains with the named human authority. That is a more useful design principle than simply declaring that “a human is in the loop.”

This resource explains; it is neither legal, tax or financial advice nor a formal valuation. Have your situation validated by a professional.