AI for due diligence: what it reads, what it flags, what it misses

Due diligence is fundamentally a document problem. A typical M&A or investment process generates hundreds of documents — financial statements, contracts, corporate filings, board minutes, IP registrations, employment agreements, regulatory correspondence — spread across a data room that analysts need to read, extract from, cross-reference, and flag within a compressed timeline. The document volume is the constraint, not the analysis itself.

AI addresses the document problem directly. What it doesn't do is replace the judgement layer that sits on top of it — the interpretation of what the extracted data means for deal valuation, negotiation leverage, and investment risk.

The three tracks where AI changes the workflow

Due diligence typically runs on parallel tracks, and the impact of AI differs across them.

Financial track. The high-volume work here is extraction and normalisation: pulling figures from multiple years of financial statements, identifying inconsistencies between reported numbers and supporting schedules, flagging unusual patterns in revenue recognition, working capital movements, or related-party transactions. AI models trained for financial document analysis can process a full set of audited accounts — balance sheet, P&L, cash flow, notes — in minutes and produce a structured extract with anomaly flags. For Italian companies specifically, statutory accounts filed with the Camera di Commercio in XBRL format are directly machine-readable, which makes extraction more reliable than from scanned PDFs. The Visura camerale and CCIAA filings provide additional public data on corporate structure, share capital history, and registered charges that can be automatically pulled and incorporated into the analysis.

Legal track. Contract review is where AI delivers the most consistent time savings. A data room with 200 contracts — customer agreements, supplier contracts, leases, IP licences — would take a legal team days to review for standard risk clauses: change of control provisions, termination rights, IP ownership, non-compete obligations, material adverse change clauses. AI contract review tools can flag the presence or absence of these clauses across the full document set in hours, with the relevant excerpts surfaced for legal review. The analyst's time shifts from reading every contract to reviewing the flagged items — a much smaller task.

Commercial track. This is the weakest fit for AI automation, because the relevant data is less structured. Customer concentration analysis, market positioning assessment, and management quality evaluation depend partly on documents (customer lists, pipeline data, management CVs) and partly on information that doesn't exist in the data room — market context, competitive dynamics, reference checks. AI can assist with synthesis and structuring of what's available, but the commercial judgement call relies on inputs that aren't documentable.

What AI flags that humans often miss

The most practically useful capability is cross-document consistency checking — comparing claims made in one document against data in another at a scale and speed that's impractical manually. A company's representation that no material contracts contain change-of-control provisions can be checked against the actual contract set. Revenue figures in the management presentation can be reconciled against the statutory accounts. Related-party disclosures in the notes can be cross-referenced against the corporate structure chart.

These consistency checks catch discrepancies that might be innocent (different accounting treatments, presentation choices) or material (misrepresentations, undisclosed liabilities). Either way, surfacing them early is valuable — it focuses human review time on the items that matter rather than spreading it evenly across all documents.

What AI doesn't do

Three things that remain human, and where substituting AI judgement creates risk rather than reducing it.

Materiality assessment. AI can flag that a contract contains a termination right exercisable on change of control. It can't assess whether the counterparty is likely to exercise it, what the commercial relationship is worth, or whether it's a standard clause that's never been triggered in practice. That assessment requires market knowledge and relationship context.

Management evaluation. The quality and integrity of the management team is one of the most important variables in any investment, and it's almost entirely inaccessible through documents. Reference checks, structured interviews, background verification — these are human processes. AI can help structure the questions and synthesise publicly available information, but the evaluation itself isn't automatable.

Deal-specific risk framing. A flag is only useful if it's evaluated in the context of the specific deal — the price, the structure, the buyer's risk appetite, the strategic rationale. AI produces findings. Analysts decide which ones are deal-breakers, which require price adjustment, and which can be managed through warranties and indemnities. That translation from flag to recommendation is the analyst's job.

A practical setup for Italian deal teams

For advisory firms and corporate M&A teams operating in the Italian market, a useful AI-assisted workflow combines automated extraction from XBRL filings and CCIAA data (for public company information that doesn't need to wait for a data room), AI contract review configured around standard Italian law risk clauses, and a document Q&A interface that lets analysts query the full data room in natural language — "show me all references to pending litigation across all documents" — rather than navigating folder structures manually.

The setup reduces the time from data room access to first-pass findings significantly, which matters when timeline pressure is a deal variable. It doesn't change what experienced analysts need to do with those findings.