Bankers should treat every AI-assisted CIM statement as a claim that needs a named source, a version, an accountable reviewer and a status, rather than as text that becomes reliable because a model produced it. Build the matrix during drafting, link each factual sentence, metric and exhibit to the financial record, management confirmation or other approved source, and flag unsupported or stale items. Keep permissions and confidentiality decisions beside the source record. A human deal team member should resolve conflicts and approve wording before circulation. This drafting matrix improves traceability, but it does not replace the separate final release verification gate or any applicable retention, supervision or confidentiality obligations [1] [2].
Start with the claim, not the paragraph
Break the draft into reviewable claims. A claim can be a sentence about revenue, a margin movement, customer concentration, a market position, a management statement, or an assertion embedded in an exhibit title. Do not treat a whole page as one evidence unit. A reviewer should be able to ask what exactly is being asserted and what record supports it without rereading the entire data room. Give each claim a stable identifier such as CIM-REV-014 and record the claim text or a concise normalized version. Add claim type, section, page or exhibit location, materiality, and whether the wording is reported, calculated, management-provided, or interpretive. This distinction matters because a reported figure may link to a ledger or statement, while an interpretation may require both underlying figures and reviewer reasoning.
Use a matrix that preserves evidence lineage
The core matrix should include claim ID, draft wording, source-of-truth record, source locator, evidence version, period, units, preparer, reviewer, review status, and notes on conflicts. For a financial metric, capture the statement or workbook name, reporting period, currency, whether the number is reported or adjusted, and the page, tab, cell, or paragraph that supports it. For a management confirmation, record the date, channel, sender role, question asked, response, and approved record location. For an exhibit, link both underlying data and the calculation or transformation used. Add fields for permissions and confidentiality handling. Say whether the source may be used in the CIM, is restricted to the deal team, requires redaction, and who can authorize a change. Do not assume that a source can be copied into an external model or data room simply because the team can view it.
Version the source and claim together
A link is not enough if the underlying record can change. Store evidence version or file date, retrieval date, and relevant reporting period. When a revised trial balance, forecast, customer schedule, or management response arrives, create a new evidence row or version and identify every affected claim. Preserve prior review status rather than silently overwriting it. The goal is to show what the team reviewed and why wording changed. Separate source freshness from source authority. A recent management email may explain a change but may not replace a formally approved financial record. A prior-period filing may be authoritative for its period but not support a current-period statement. Record the basis explicitly and mark a claim unresolved when period, units, definition, or version cannot be reconciled. Never fill gaps with a model’s plausible completion.
Make AI output a draft with visible uncertainty
Use AI for bounded tasks such as extracting candidate claims, matching draft language to permitted records, identifying missing periods, or proposing questions for a reviewer. Ask it to return the claim ID, proposed source, locator, confidence rationale, and unresolved issue instead of silently rewriting the CIM. Require a citation or source locator for every candidate claim. If no approved source is found, the output should say unsupported rather than invent a citation or convert a general description into a fact. A model citation is not evidence by itself. The human reviewer should open the source, check the passage or figure, verify units and period, and compare proposed wording with the record. NIST emphasizes documented roles, provenance, human oversight, and recording overrides, which maps well to this review design [1]. Record whether the reviewer accepted, edited, rejected, or escalated the suggestion.
Assign review ownership by claim type
Name a role, not an unowned queue. The finance owner can verify statement-derived metrics and calculations. The deal lead can assess whether wording accurately reflects the approved narrative. The management liaison can confirm that a management statement is current and authorized for use. A senior reviewer or compliance contact can resolve questions about external communication, restricted information, or firm procedure. The same person may hold several roles in a small team, but responsibility should still be explicit. Use statuses such as proposed, source found, reviewer checking, accepted, edited, rejected, needs management confirmation, restricted, stale, or unsupported. Require a reason for rejection or escalation. A green status should mean that the assigned reviewer checked source and wording, not that an AI system assigned a high probability.
Build checks for numbers, definitions, and exhibits
Automated checks can reduce missed links, but they should produce exceptions for people to resolve. Check that every factual sentence and exhibit has a claim ID, every claim ID has a source locator, and every source has a version and period. Compare repeated metrics across the CIM for consistent definitions, units, signs, and time periods. Recalculate derived figures from approved inputs and record the formula or method. Flag claims using words such as leading, recurring, committed, or approximately unless the team has defined the basis and source. Check an exhibit as a separate evidence object. A chart can be numerically correct while its title, axis, selected period, or footnote overstates what the data show. Link the chart to the input range, transformations, and reviewer. If a claim depends on a forecast, label it as a forecast and identify approved assumptions. Do not present an AI summary as a management projection or a calculated metric as a reported result.
Set permissions and pilot stopping criteria
Before using an AI tool with deal materials, define what may be submitted, what must remain in approved systems, whether retention or training settings are acceptable under firm policy, and who can approve the configuration. Minimize sensitive content in prompts. Use synthetic or redacted material for the first test when possible. FINRA says existing obligations continue to apply when member firms use GenAI and highlights accuracy, privacy, data integrity, reliability, and technology governance as evaluation concerns [2]. That is a reason to test the workflow, not a promise that any tool is suitable. Stop the pilot if the system invents sources, drops qualifiers, mixes periods, exposes restricted content, cannot preserve the review record, or produces exceptions reviewers cannot resolve. Also stop if no human owner can be identified for a claim. Record the failure and return to a manual process until the issue is understood.
Keep the matrix distinct from final release verification
The claim-to-source matrix is a drafting-stage evidence assembly tool. It helps discover missing support, coordinate checks, and preserve reasoning behind edits. It is not final sign-off. Before a CIM is shared, perform a separate release review against then-current approved financials, management confirmations, confidentiality instructions, engagement scope, and distribution list. Recheck changed pages and exhibits, not only rows that AI touched. For firms subject to specific broker-dealer recordkeeping requirements, relevant rules and firm procedures determine what communications and records must be preserved. SEC materials describe retention and preservation requirements for covered broker-dealers, including communications relating to broker-dealer business, but those requirements should not be generalized to every adviser, transaction, or jurisdiction [3] [4]. Ask qualified counsel or the firm’s compliance function when the applicable regime or retention instruction is unclear.
What the implementation outcome should look like
At the end of drafting, the team should have a claim-to-source matrix that a new reviewer can use without relying on an analyst’s memory. Each row should answer five questions: what is being claimed, which approved record supports it, which version and period were checked, who owns the review, and what happens if evidence is missing or restricted. The matrix should expose unsupported statements rather than hide them in polished copy. 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. The workflow described here is a framework for user-operated review. It does not mean software approves claims, represents clients, provides legal advice, or guarantees accuracy, confidentiality, compliance, or a transaction outcome.
Key Points
- Assign a stable ID to every factual statement, metric, management confirmation, and exhibit claim.
- Record approved source, locator, evidence version, period, units, permissions, reviewer, and status.
- Use AI to propose links and exceptions, never to supply missing evidence or approve disclosure.
- Require human checks for source wording, calculations, definitions, confidentiality, and changed versions.
- Set stopping criteria for invented sources, period mixing, restricted-content exposure, and unowned exceptions.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
- [1] Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile · Accessed
- [2] FINRA Reminds Members of Regulatory Obligations When Using Generative Artificial Intelligence and Large Language Models · Accessed
- [3] Electronic Storage of Broker-Dealer Records · Accessed
- [4] Final Rule: Books and Records Requirements for Brokers and Dealers · Accessed
Choose your next step
Acquivo is user-operated software. Owners and their professionals remain responsible for decisions, communications, and professional review.
Educational information, not individual legal, tax, or investment advice. Consult qualified professionals for your circumstances. Acquivo provides software, not representation or negotiation on your behalf.