When I started building NaS_OS, I had a working assumption that the hard problem was generating good prose: writing the equity story, structuring the investment thesis, getting the tone of the IM right. After running the platform against real M&A datasets across multiple sectors, I was wrong. The hard problem is extraction. And the failure modes are different for every document type in the data room.
This is the post I would have wanted when I started. A practitioner-level map of where AI extraction actually breaks, by document type, with the failure modes I have personally seen and what they look like in production.
Excel financial models
Excel is the worst extraction surface in a typical data room, by a wide margin. The problems compound.
Truncated column widths. A revenue figure of €1,247,392 displays as ”€####” if the column is too narrow. AI tools that read the displayed value rather than the underlying cell value miss the figure entirely or extract nonsense. Prevention: always read the cell value, never the displayed value.
Merged header cells. Multi-row headers with merged cells produce ambiguous column attribution. The AI extracts “Revenue” from a merged cell that actually means “Revenue by region: North” and applies it to the wrong subtotal. This is the source of most “the number is real but in the wrong place” errors I have seen.
Hidden rows and columns. Sellers sometimes hide rows that contain detail they do not want highlighted. AI tools that ignore hidden cells will miss material information. Tools that include them indiscriminately may surface information the seller intentionally suppressed. There is no universally correct behavior here; the tool has to flag the existence of hidden cells and let a human decide.
Cross-sheet references with broken links. A financial model may reference values from a sheet that was removed before the file was shared. The AI sees #REF! and either ignores the cell (losing real information) or treats the error as a value (introducing nonsense).
Scanned PDFs
Scanned PDFs (usually older legal documents, signed contracts, regulatory filings) are the second-hardest surface. The failure modes are different from Excel.
Table structure loss. A clean PDF of a tax filing might have a table with clear columns: “Year,” “Revenue,” “Taxable Income,” “Tax Paid.” A scanned version of the same table renders as flat OCR’d text with the column structure lost. The AI has to reconstruct the table from spatial heuristics, and the success rate on this is meaningfully below 100% even with the best 2026 OCR.
Stamp and signature interference. Notary stamps, signatures, and watermarks overlap with text in ways that confuse OCR. The figure on the bottom-right of a contract page may be partly obscured by a stamp. The AI extracts a number that is missing a digit. The €4.2M cap on a liability appears as €4.2k.
Multi-column page layouts. Legal documents often use two-column layouts. OCR systems that read left-to-right top-to-bottom rather than column-by-column produce text that mixes the two columns and is nearly impossible to parse correctly.
Long-form text documents (PDFs, Word files)
Text-native PDFs and Word documents (strategy decks, board materials, sales force presentations) have different failure modes.
Document version confusion. The data room often contains multiple versions of similar documents: “Strategy Deck v3,” “Strategy Deck v4 final,” “Strategy Deck final final.pptx.” Without explicit version tracking, the AI may pull a number from v3 that was revised in v4. The number is real but obsolete.
Source attribution when filenames are similar. “Acme Holdings - Financials.pdf” and “Acme Operating Ltd - Financials.pdf” sit next to each other in the folder. The AI extracts a figure from one and cites the other. This is a real failure mode I have seen and the only prevention is explicit document fingerprinting beyond filename matching.
Charts and graphs. A PowerPoint slide with a chart showing revenue growth contains the chart image, sometimes the underlying data table, and a text caption. The AI may extract the figure from the caption (which can be outdated), from the chart label (which is what the slide visualizes), or from the underlying data (which is most accurate). Without explicit preference logic, results are inconsistent.
Management accounts and accounting exports
This is its own category of failure mode because the documents come from accounting software exports.
Currency unit confusion. Management accounts are often reported in €k while audited financials are reported in €. A naive AI extracts both and reports them in the same units, producing a company that looks 1,000x its actual size in some sections and the correct size in others. This is the most common arithmetic error I have personally caught in NaS_OS’s output during testing, and it is exactly the kind of error that takes minutes to make and hours to find.
Fiscal year vs. calendar year. A company on an April fiscal year reports “2023 revenue” that means April 2023 - March 2024. The AI may treat this as calendar 2023 revenue, producing a comparison error when matched against calendar-year industry benchmarks.
Restated periods. Companies sometimes restate prior periods to reflect changes in accounting policy or corrected errors. The AI that pulls “2022 revenue” from the original 2022 statements and then pulls a different “2022 revenue” from the restated 2023 comparative column produces internal inconsistency in the IM.
Cross-document semantic failures
The hardest failures are not specific to any one document type. They emerge from the interaction between documents.
Management role rotation. A CEO from 2019 left in 2020. A new CEO took over in 2021. Documents from 2018-2024 reference different people in the same role. The AI that does not carefully track temporal attribution will mix attributes, applying the current CEO’s track record to the prior CEO’s tenure. The management section of the IM ends up factually wrong in a way that is immediately credibility-destroying.
Entity structure confusion. A group structure includes a holding company, an operating company, and a finance subsidiary. Revenue numbers are different for each. Documents reference them inconsistently: sometimes “the Group,” sometimes “the Company,” sometimes by specific entity name. The AI that aggregates without disambiguating produces a P&L that is internally inconsistent.
Subsidiary disposals. A subsidiary was sold in 2023. Pre-2023 financials include it; post-2023 financials don’t. An AI that extracts revenue across years without flagging the disposal will produce a misleading growth trend.
What this means for the buyer of AI tools
If you are evaluating an AI tool for M&A document work, do not be impressed by the prose. The prose is easy in 2026. The extraction is hard. Ask the vendor specifically how their tool handles each of these failure modes. If they cannot answer in detail, they have not built for production use.
Every one of these failure modes is solvable with the right architecture. Most are solved by combining strong extraction with explicit traceability: every figure tied to a specific document, page, cell, and version. None are solved by just throwing a bigger language model at the problem. The model is rarely the bottleneck. The pipeline around it is.