AI Risk Scoring in M&A: Beyond the Buzzword

If you've spoken with a diligence vendor in the last two years, you've heard some version of the same pitch: "Our AI surfaces risks you'd miss manually." The claim is so ubiquitous it has become meaningless — and that's a problem, because there are things machine intelligence genuinely does well in deal risk assessment, and confusing the hype with the reality means both over-trusting tools that have hard limits and under-using the ones that actually work.

Here's a clear-eyed look at what AI can and can't do in the context of pre-deal intelligence.

What AI Does Well

Aggregation at scale

The most legitimate use of AI in due diligence is volume processing. Pulling regulatory filings, court records, UCC liens, corporate registrations, and news across multiple jurisdictions in hours rather than days is genuinely valuable — not because the analysis is sophisticated, but because the speed and coverage are impossible to match manually.

An experienced investigator working alone can review a meaningful slice of this data in a week. An AI tool running overnight covers the same ground and surfaces items that would require human judgment to flag. The key word is surface — the AI found it, the human needs to assess it.

Pattern flagging in structured data

When you have structured, consistent data — financial statements, court filings, corporate ownership records — machine learning is good at flagging anomalies. Revenue figures that don't align with reported headcount. Ownership structures that suggest circular control. Disclosure inconsistencies across filings from different years.

These are pattern-matching tasks. The data exists, the patterns are defined, and a model trained on enough examples can flag deviations faster than a human analyst scanning rows.

Entity resolution

A common investigation challenge: a person has multiple business identities, uses variations of their name, or is associated with entities through intermediaries. AI-assisted entity resolution — linking records across fragmented data sources — is meaningfully better than manual cross-referencing for this specific problem.

Where AI Consistently Fails

The risk in an M&A transaction is rarely the data you can find. It's the judgment you need to evaluate it.

Unstructured qualitative intelligence

The things that actually kill deals after close — an investor's pattern of behavior with portfolio founders, a CFO's history of creative accounting that never crossed a legal threshold, the reputation that travels by word of mouth in an industry — don't exist in databases. They exist in the minds of people who've worked with this individual.

No AI system can conduct a discreet conversation with a former colleague. No model can read the hesitation in a trusted source's voice when you ask whether they'd work with this person again. No algorithm can assess the credibility of competing accounts from two people with different incentives.

Contextual risk weighting

An AI tool can tell you that the CFO you're about to hire had a company that filed for bankruptcy six years ago. It cannot tell you whether that's a red flag or a learning experience — because the answer depends on context, industry norms, the circumstances of the failure, what she did next, and how the people who went through it with her describe her leadership under pressure.

Risk scoring that doesn't carry context isn't risk scoring — it's noise generation. The volume of false positives from automated systems is the underreported cost of over-indexing on AI diligence.

Novel risk configurations

Machine learning is effective when the patterns it needs to recognize resemble training data. Deal risk often manifests in novel configurations: a regulatory environment that shifted in a specific jurisdiction, a technology IP claim that interacts with a recent court ruling, a personal relationship between the founder and a board member that creates a conflict no template anticipated.

These configurations require reasoning about facts and context, not pattern recall. That's still a human task.

The Practical Implication

The diligence vendors selling AI as a replacement for investigation are selling something that doesn't exist. AI makes good investigators faster. It does not replace the investigator.

The right architecture for pre-deal intelligence uses AI for exactly what it's good at — coverage, speed, pattern flagging — and preserves human judgment for everything that actually determines whether you're making a good decision. That means combining automated data aggregation with direct source inquiry, and synthesizing both into an editorial conclusion by someone who has seen enough deals to know what matters.

The risk of the all-AI pitch isn't that the tool doesn't work. It's that buyers leave the process confident they've done diligence when they've actually only done data retrieval.