The most common objection I hear from M&A partners considering AI tools is some version of: “What about hallucinations?” It is a legitimate concern, and most people asking it have a worse mental model of where AI errors actually come from than they realize. They are thinking about the kind of mistakes ChatGPT made in 2023: completely fabricated quotes, made-up case law, plausible-sounding nonsense. That is not the failure mode that matters in M&A documents in 2026.
The failure modes that matter are more subtle, harder to detect, and more dangerous. They are also entirely preventable with the right architecture. I want to walk through what actually goes wrong, because partners deserve a more accurate model of the risk before they decide what to do about it.
Failure mode 1: The misread Excel cell
The single most common AI failure I have seen in financial document work is not a hallucination at all. It is an extraction error. The model reads a financial spreadsheet, encounters a merged cell or an oddly-formatted header, and pulls a figure from the wrong row.
The number it returns is real. It exists in the document. It is just the wrong one.
This is dangerous because the standard mitigation for hallucinations, fact-checking against the source, does not catch it. The fact is checked, the source is real, the citation is accurate. The error is that the AI attributed the wrong row’s number to the wrong line item. A €4.2M figure that belongs to gross margin gets reported as EBITDA. The IM goes out with the wrong profitability profile. A buyer notices in week three of diligence and the deal repositions.
The prevention here is not better prompting. It is better extraction architecture: page-level, row-level traceability that lets a human reviewer click on any figure in the output and immediately see exactly where in exactly which document it came from, including which cell of which sheet.
Failure mode 2: Source confusion across similarly-named documents
A typical mid-market data room contains 200-400 documents. Many have similar names: “Acme Holdings - Audited Financials 2023.pdf” and “Acme Operating Ltd - Audited Financials 2023.pdf” sit next to each other in the folder. The legal entity structure is intentionally similar across the group.
When an AI tool retrieves a figure from “the financial statements,” which one did it use? If the answer requires a forensic review of the AI’s reasoning trace, the tool is not ready for M&A work.
The right architecture for this is to never let the AI reference a source generically. Every figure has to be tagged with the specific document, page, and where possible, the specific cell or paragraph. The user reading the IM should be able to hover over any number and see “Source: Acme Holdings Audited Financials 2023, page 14, Note 3.” Not “Source: Audited Financials.”
Failure mode 3: Currency and unit confusion
This is a quiet killer in European mid-market deals. The financial statements report in €k. The management accounts report in €. The pitch deck the founder gave the advisor reports in €M. The AI tool aggregates them and reports a figure in €.
If the conversion logic is not explicit and verifiable, you get a company that looks 1,000x bigger or 1,000x smaller than it is. This sounds like a comic problem until you see it happen on a real document.
The same pattern affects USD vs. EUR, GBP, sometimes USD vs. CAD when a Canadian operating subsidiary’s numbers get mixed with US parent numbers. Prevention requires explicit unit handling at every extraction point: not “extract revenue” but “extract revenue, source unit, conversion factor, target unit, with all four traceable.”
Failure mode 4: Cross-section inconsistency
The IM has twenty sections. The same fact, say, customer concentration, gets referenced in the executive summary, the business overview, the risk factors, and the financial section. If the AI generates these sections independently, the customer concentration figure might be stated as “top 5 customers represent 35% of revenue” in one section and “top 5 customers represent 42% of revenue” in another. Both numbers are from the data room. Both are real. They use different definitions or different time periods.
A human reviewer reads the document linearly and may not catch the inconsistency until a buyer flags it. By then the firm has lost credibility on a document that was supposed to demonstrate diligence.
The prevention is to architect the document generation so that core facts are extracted once, defined once, and then referenced from a single source of truth across every section. This is harder than it sounds because the right level of granularity for “core facts” varies by deal. But it is achievable, and any AI tool a partner is evaluating should be able to demonstrate it.
Failure mode 5: Semantic role attribution
This is the most M&A-specific failure I have seen. The data room contains documents that reference multiple people in management roles: a former CEO from 2019, the current CEO since 2021, an interim CEO during a six-month transition. An AI tool that does not carefully track temporal attribution will mix attributes, applying the current CEO’s track record to the prior CEO’s tenure, or vice versa.
This shows up in the management section of the IM, which is one of the sections buyers read most carefully. An error here is immediately credibility-destroying.
Prevention requires semantic tracking that goes beyond name-matching. The AI has to understand that “John Smith” in document A is “the CEO who led the 2020 reorganization” while “John Smith” in document B might be “the chairman as of 2024.” If the tool cannot make this distinction reliably, it should refuse to populate the management section autonomously.
What partners should require from an AI vendor
When you are evaluating any AI tool for M&A work, the questions to ask are not about model quality or context window size. They are about validation architecture:
Can every figure in the output be traced to the exact source document, page, and cell? Are similar-named documents handled with explicit disambiguation, or just retrieved by similarity? Is unit and currency handling explicit and auditable, or implicit and prone to silent conversion errors? Are cross-section consistency checks built in, or does the tool generate each section independently? Does the tool refuse to populate sections where it does not have high confidence, or does it fill in gaps with plausible-sounding content?
Any vendor that cannot answer all five questions specifically is not ready to handle a real deal. Any tool that can answer them well is a different category of product than what most people picture when they hear “AI for M&A.” The difference between “the AI said it” and “the AI cited page 47 of the audited financials, Note 3, line 8” is not a feature. It is the whole product.