How Should M&A Attorneys Validate AI Extraction From Scanned Deal Documents?

Assign extraction, verification and legal decisions to different owners

Define the task narrowly: turn scanned deal documents into searchable text and structured fields while preserving a route back to the images. Optical character recognition, or OCR, recognizes text in images. A further AI step might map that text into fields or summarize it. Treat those as separate stages so a reviewer can identify whether an error arose in reading the scan or interpreting the extracted text.

A document coordinator should own the inventory and scan-quality triage. A trained paralegal or attorney should own source comparison and corrections. Supervising matter counsel should determine the review required for the intended use and retain responsibility for legal conclusions. The firm's designated technology or security reviewer should assess the proposed processing environment. These are recommended assignments, not roles a tool can assume.

For generative AI, ABA Formal Opinion 512, issued July 29, 2024, describes task-dependent independent verification and says lawyers cannot substitute AI for professional judgment. It interprets ABA Model Rules, not a universal rule governing every attorney. Counsel should check applicable jurisdictional rules and current local guidance before adopting the workflow. [1]

Approve the processing environment before uploading deal scans

Before a trial, identify exactly which files the team is authorized to process and who may see the inputs, outputs and exception queue. Review client instructions, data-room restrictions and relevant agreements. Receiving a document does not establish permission to submit it to every outside processing service. Matter counsel should resolve uncertain permissions rather than allowing a document coordinator to infer them.

For generative AI tools, the ABA opinion recommends understanding terms of use, privacy policies and related contractual terms. It calls for a fact-specific confidentiality assessment and addresses circumstances in which informed consent is required. The applicable professional rules and circumstances determine the legal obligation. Do not reduce this review to a generic engagement-letter clause or a vendor's description of a product as secure. [1]

Ask the technology reviewer to establish retention and deletion arrangements, training use, provider and internal access, and available access restrictions. If those facts or required permissions remain unresolved, do not upload the affected files. Use public sample documents or approved synthetic test material instead. Neither local processing nor redaction by itself establishes confidentiality, privilege or compliance.

Reconcile the source inventory and missing pages first

Create a source manifest before extraction. Record a document identifier, original filename, version description, receipt date and page count. Preserve the received file unchanged and associate any working copies with that identifier. A reviewer should be able to distinguish the original scan from a rotated, enhanced or corrected derivative without relying on memory.

Compare the PDF page sequence with printed page numbers, the contents page, exhibit references and the document coordinator's inventory. Inspect apparently blank pages before removing them. Look for duplicate pages, missing reverse sides, cropped margins and schedules referenced but not supplied. A matching extracted-page count proves only that processing covered the supplied file, not that the supplied file contains the complete agreement.

NARA's October 2023 digitization guide emphasizes accounting for source records, identifying missing pages and checking whether source information was captured accurately. Its federal records requirements do not prescribe a private M&A review procedure; those completeness principles are useful design references. [2] Where completeness cannot be established, record the gap and have the coordinator request the missing source through authorized channels. Do not let AI reconstruct it.

Triage scan quality without altering the source evidence

Classify pages as readable, readable with defects, or unsuitable for dependable extraction. Inspect rotation, skew, blur, shadows, glare, clipped text, faint printing and markings over words. Check small footnotes and handwritten additions at a usable magnification. NARA cautions that spatial resolution alone does not guarantee information quality, so a scanner setting or file-size label should not decide acceptance. [2]

Use tool-generated quality indicators as triage aids where available. Google Cloud's Document AI documentation, for example, describes optional page readability scores and detected defects such as document cutoff, text cutoff and glare. Those are documented outputs of that service, not evidence that Acquivo uses it or that a particular extracted field is correct. [3]

Rotate or adjust contrast only on a clearly labeled working copy. Compare any improvement with the original to ensure that punctuation, decimal points or handwriting have not disappeared. If a page remains unreadable, seek a better scan or another authorized source copy. Do not treat an AI-generated reconstruction of obscured characters as recovered source text.

Compare extracted text with the exact page image

Review text alongside the image, using page identifiers and, where available, highlighted source regions. Google documents OCR output with text anchors, page references and bounding polygons that connect recognized text to its location. Its raw text field does not preserve full layout. Require equivalent traceability from the chosen workflow rather than accepting an isolated text export as the transaction source of truth. [3]

For the initial trial, inspect every page for omissions and reading-order problems. Compare all fields designated critical for the intended use, including party names, amounts, dates, clause references and material qualifications. Check whether text from a header, footnote or neighboring column has been inserted into a clause. Searchable text that reads naturally may still omit a handwritten insertion or a short qualification.

Preserve the initial extracted value and the reviewer-corrected value separately. Record the source page, correction reason and reviewer role. A second AI pass can flag possible differences, but it should not resolve them without a human opening the image. If source alignment cannot be reproduced, place the item on hold rather than accepting its apparent plausibility.

Validate tables as relationships, not just recognized numbers

For each table used in diligence, compare its heading, row labels, column labels, units and footnotes with the source. Verify that each amount belongs to the correct entity, period and category. Inspect continuation pages for repeated headers or rows that were dropped or duplicated. A number can be transcribed correctly yet become misleading when attached to the wrong column.

Pay particular attention to merged cells, multi-line labels and nested tables. Google's documentation states that its Form Parser table extraction recognizes conventional tables without cells spanning rows or columns. That specific limitation illustrates why reviewers should check the selected processor's documented layout support rather than assume every AI extractor handles every schedule. [3]

Keep blanks distinct from zeros and unreadable entries. Confirm currency, scaling labels and parentheses before exporting values into a diligence worksheet. Where the source presents a total, a reconciliation can flag a possible transcription problem, but it cannot establish that the source accounting is correct. Route source inconsistencies separately from extraction errors so the coordinator does not silently change the document's figures.

Separate dates and signature observations from legal conclusions

Keep literal date text alongside any standardized date. Google documents that specialized processors can return normalized dates and that interpretation depends on location information. [3] If a numeric date could reasonably be read in different orders, mark it ambiguous unless the document context and a human reviewer resolve it. Do not replace the original characters with an unsupported assumption about the parties' convention.

Check whether the date appears in the opening clause, an amendment, a signature block or a handwritten annotation. Extraction can identify where a date appears; deciding its legal effect is a separate counsel task. Similarly, verify visible signature marks, typed names, titles and adjacent dates against the images without labeling a signature authentic or a document enforceable.

A cropped signature block, unreadable handwriting or apparently blank counterpart belongs in the exception queue. Counsel may need a better copy or separately obtained execution records through approved channels. The required evidence depends on the transaction and applicable law. A visible mark is not a substitute for that assessment, and an extraction tool should not declare execution complete.

Build an exception queue with explicit resolution evidence

For each uncertain item, record the document identifier, version, page and field or region. Include the raw extraction, the observed problem, any permitted source-image reference, intended downstream use, assigned owner and status. Useful statuses include awaiting better source, awaiting transcription review, awaiting legal interpretation and resolved. Restrict queue access consistently with the underlying material.

Distinguish unreadable source text from missing material, ambiguous text and a tool error. For an illustrative example, a faint handwritten date should enter the queue as unreadable, not be completed from another agreement's date. A clear date misread by OCR can be corrected from the image. A clear clause whose effect is disputed belongs with counsel, not with the transcription reviewer.

Close an item only when the owner records the supporting source and resolution. Retain earlier values so downstream users can identify what changed. If a replacement scan arrives, preserve its version relationship and recheck affected fields and page continuity. An unresolved item should remain visibly excluded from any verified dataset, even if a business deadline makes the incomplete result inconvenient.

Run a difficult-scan trial before releasing verified output

Choose an authorized trial set containing clean pages and the difficult layouts actually expected in the matter. Create a human-checked reference for critical fields. Assess omissions and structural errors as well as character recognition. ABA Opinion 512 discusses testing on a smaller manually reviewed document set when determining appropriate reliance on generative AI, while leaving the needed review dependent on the tool and task. [1]

Set concrete stopping conditions before processing: unauthorized access or upload, an untraceable field, an unexplained missing page, an unresolved critical value, or a repeated table-alignment failure. Stop the affected use, correct the cause and rerun the relevant checks before widening the trial. A high average score should not override a failure in a field needed for a legal decision.

Use a release checklist covering inventory completeness, scan triage, page comparison, critical fields, tables, dates, signature observations and open exceptions. Name the human who accepts the extraction for its specified use, and keep disclosure approval separate. Revalidate affected checks when the tool version, settings or document type changes. No extraction result should trigger autonomous disclosure, negotiation or approval.

Acquivo is a software platform that delivers the entire sell-side M&A process, valuation, marketing materials, buyer matching, and deal management, on one platform for business owners, M&A attorneys, and boutique investment banks.

It is user-operated software with no success fee, not legal representation. This checklist is a recommended implementation framework, not a claim that the platform supplies every control described. Attorneys and authorized users retain responsibility for source verification, permissions and legal decisions.