How Should Bankers Check AI-Extracted IOI and LOI Terms Against the Original Offers?

Start with a controlled source set

Before reviewing an AI extraction, define which offer files are in scope and which version is authoritative for the comparison. Create one record for each indication of interest (IOI) or letter of intent (LOI) with the file name, received date, version or revision label, document date, page count and confidentiality classification. If a revised offer replaces an earlier one, preserve both records and identify the version used for the current comparison. Do not let a convenient upload or a renamed file silently become the source of truth.

The reviewer should also record whether the document is an email attachment, scanned PDF, word-processing file or image. Format affects what must be checked. A scanned page may need visual confirmation of numbers, symbols and footnotes. An email may contain a term in the body that is not repeated in the attachment. An exhibit can include schedules or signature pages whose qualifiers matter. The extraction task is transcription from a defined source set, not a judgment about which offer is economically best.

Define fields before asking for extraction

Use a fixed field list so the system is not rewarded for producing a polished paragraph while omitting a material qualifier. A practical list begins with bidder identity, target or assets covered, proposed consideration, cash or equity mix, assumed debt or other adjustments, rollover or retained ownership, and any stated valuation basis. Add structure fields such as transaction form, proposed closing timing, funding language, financing condition, approval condition, due diligence scope, and required definitive agreements.

Keep process and restriction fields separate from economics. Capture exclusivity, confidentiality, announcement rights, expense allocation, access to information, contact restrictions, expiration, acceptance deadline and governing-law language when present. Add a status field for each item: stated, qualified, absent, unclear or not applicable. This prevents a blank cell from being mistaken for a negative answer. A field marked absent means the reviewer checked the relevant source pages and did not find it. A field marked unclear means the wording or scan requires escalation.

Require page-level evidence for every populated field

Every extracted value should point back to a page, section, paragraph or exhibit locator in the original offer. For a multi-page clause, cite the page where the operative language starts and note the additional pages that complete the condition. For a schedule, identify the schedule name and page rather than citing only the cover page. If the source is an email, preserve the sender-provided date and subject in the internal record and cite the relevant message portion under the firm’s document-control rules.

A useful row has five connected parts: field name, extracted text or normalized value, source locator, qualifier, and reviewer disposition. Keep the extracted wording alongside any normalized value. For example, a reviewer may normalize a consideration description into separate cash and equity columns while retaining the original sentence. That makes it possible to compare bidders consistently without erasing whether the source said approximately, up to, subject to, contingent on, or another limiting phrase.

NIST’s AI Risk Management Framework is voluntary guidance, but its emphasis on documenting AI use and managing trustworthiness risks supports this traceable design. Its Generative AI Profile specifically discusses data provenance, human oversight roles, tracking, documentation and monitoring overrides. Those concepts are useful controls here, not proof that an extraction is correct. [1]

Check numbers, units and qualifiers against the page image

Review numeric fields twice: first as a transcription check and then as a comparison check. Confirm currency, units, decimal placement, percentages, share counts, dates and whether a number applies to the target, a subsidiary, a class of securities or the combined company. Check whether a value is stated as a range, estimate, cap, floor, percentage, multiple, per-share amount or total amount. Do not convert an approximation into a precise figure in the comparison table.

Read the surrounding sentence and any footnote before accepting a value. A proposed price may be subject to a closing adjustment, assumed debt, working-capital mechanism, financing, approval or a definitive agreement. A percentage may refer to post-closing ownership rather than the purchase price. A deadline may be an expiration for acceptance rather than a closing date. Preserve the condition in the qualifier column and keep the normalized field conditional when the source is conditional.

The SEC-filed LOI used as an example in this article describes preliminary principal terms and conditions, says that other terms remain to be addressed, and separately lists due diligence, closing conditions, confidentiality, exclusivity, announcements and expiration. It also distinguishes provisions described as binding from the broader transaction terms. That document demonstrates why a reviewer must capture both the term and its surrounding qualification. It does not establish a universal rule for every IOI or LOI. [2]

Treat missing fields as findings, not blanks

A verified table should make uncertainty visible. Use explicit labels such as not stated in reviewed pages, source unreadable, conflicting versions, terminology unclear, and counsel review required. Do not infer that a missing financing condition means the bidder is fully funded. Do not infer that no exclusivity clause exists because it was not found on the first page. Record the search scope and the pages checked, then return the field for human review if the scope is incomplete.

When two offer versions differ, do not overwrite the earlier value. Keep the competing values, identify their dates and status, and ask the deal-team owner which document controls the current comparison. If the source uses a defined term that is not explained in the offer, preserve the defined term and flag it rather than expanding it from memory. Questions about enforceability, tax treatment, regulatory effect or whether an offer should be accepted belong with qualified counsel and the authorized decision-makers.

Assign review to the person accountable for the comparison

The extraction reviewer should be named by role, not treated as an invisible quality check. A banker who owns the bidder comparison can confirm whether the field list fits the process, while counsel can address legal meaning and the client or authorized deal lead can resolve which version or interpretation governs the work product. Separate transcription review from approval to advance an offer. An AI system should not contact bidders, negotiate language, accept an offer, approve disclosure or decide that a condition has been satisfied.

For firms subject to FINRA rules, Rule 3110 requires a supervisory system reasonably designed to achieve compliance and written procedures, and it includes written evidence of review for investment banking and securities transactions. Applicability depends on the firm’s status, activities and jurisdiction. The operational lesson for this workflow is narrower: define the responsible reviewer, use written procedures where required, and retain evidence of what was checked. This article is not a compliance determination for a particular firm. [3]

Pilot the workflow with stopping conditions

Start with a small, representative set of offers that includes clean text, scanned pages, tables, footnotes and conditional language. Before broader use, have the accountable reviewer compare every populated cell with the original page and log corrections by field type. The point of a pilot is to learn where the workflow fails, not to produce a success percentage that has no defined sample or date.

Stop the pilot or narrow the permitted use if page references are missing, qualifiers are routinely dropped, numbers change during normalization, version conflicts are hidden, or reviewers cannot reproduce the extraction from the cited source. Stop using the output for bidder comparison when a material field is unclear until a human resolves it. Record the error, source page, correction and disposition. NIST’s Generative AI Profile recommends documented measurement, domain-expert input, monitoring and attention to human overrides. Those practices support a controlled trial, but they do not guarantee accuracy, confidentiality, security or compliance. [4]

Use a final exception pass after sign-off. Filter the table for every qualified, unclear, absent, conflicting or counsel-review-required status, then confirm that each exception has an owner and next action. This keeps unresolved work visible when the comparison is circulated and avoids presenting a clean-looking summary that hides open questions.

Set confidentiality choices before uploading any offer. Follow the firm’s access controls, retention rules, approved tools and client instructions. Minimize copied text and restrict access to the source set and extraction table to people who need it. If the selected tool cannot meet the firm’s requirements for sensitive transaction documents, do not upload the files. Use a manual review or an approved environment instead.

Deliver a comparison table that can be challenged

The final deliverable should let another reviewer move from a comparison cell to the exact source passage without guessing. Include the offer identifier, field, extracted value, original wording or short quotation where permitted, page reference, qualifier, version, reviewer status and unresolved question. Keep a separate change log for corrected fields. If a value is normalized, show the normalization rule and retain the original unit. If a field is not stated, show the checked scope rather than leaving an unexplained blank.

Before circulation, run a field-by-field sign-off: identity, consideration, structure, conditions, timing, process restrictions, confidentiality, expiration and unresolved items. Confirm that no source was silently substituted, no qualifier was removed, and no legal conclusion was added to a transcription. The output is ready for comparison only when the accountable banker can explain what is supported by the offer, what remains uncertain and what requires counsel or client direction.

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, so the banker remains responsible for selecting sources, setting permissions, reviewing extracted terms and deciding what may be shared.