AI can help you move faster on a diligence memo. It can organize source multiai.pro material, surface questions, compare scenarios, and draft clear prose. It can also confidently repeat a typo, confuse revenue with bookings, or produce a calculation that looks plausible but is wrong.
The answer is not to ask one model to “research the deal and write the memo.” A safer approach is to build a workflow in which evidence, assumptions, calculations, and judgments remain distinguishable—and where people verify the numbers that matter.
This guide uses the Skybridge $42M acquisition as an illustrative example. Skybridge and all supporting operating figures below are fictional; the example is not a claim about a real company or transaction. The point is to show how to use AI without letting invented or unverified figures enter an investment decision.
Start With a Number-Control System, Not a Draft
Before drafting, create a simple evidence ledger. Give every material figure a source, a date, a unit, a definition, and a status. That status should make clear whether a number is a reported fact, a calculated figure, an assumption, or an unresolved item.
This is where the discipline to separate facts assumptions becomes practical. A seller-provided figure may be a fact about what the seller reported, but it is not automatically a verified fact about the business. Label it accordingly.
Item Value Classification Source or basis Verification status Proposed enterprise value $42M Transaction term Illustrative indication of interest Confirm whether cash-free/debt-free and subject to adjustments Last-twelve-month EBITDA $6M Seller-reported figure in this example Illustrative management presentation Reconcile to financial statements and quality-of-earnings work Implied EV / EBITDA 7.0x Calculated figure $42M ÷ $6M Recalculate independently; confirm numerator and denominator definitions Next-year EBITDA growth 10% Assumption Illustrative scenario input, not a verified forecast Test downside and support with operating evidenceKeep the ledger close to the memo. If a claim cannot be traced to a source or an explicitly labeled assumption, it should not appear as an established fact. This small control prevents a common failure: a draft that turns “management expects” into “the company will.”
Use AI to Organize Evidence, Not Manufacture It
Give models a bounded task
Use AI first for tasks that can be checked against supplied material: extracting figures from a data room index, summarizing a customer contract, listing inconsistencies between two documents, or turning interview notes into follow-up questions. Tell the model to quote or identify the source for every material claim and to say “not provided” when evidence is missing.
A useful instruction might be: “Use only the documents in this thread. For each numerical claim, provide the document name, page or section, period, unit, and exact wording where possible. Do not infer missing values. Label seller statements as seller-reported.” This is more effective than asking for an impressive-sounding market overview.
AI can help locate potential benchmarks, but it should not be treated as a benchmark database. Ask it to identify what benchmark would be relevant, what definitions need to match, and what primary source to retrieve. Then cite benchmarks in the memo with enough detail for a reviewer to inspect the source, including publication or dataset, date, sample, and metric definition. If you cannot find and validate the source, describe the benchmark as unavailable rather than inserting a model-generated number.
Maintain a source hierarchy
Not all evidence deserves equal weight. For financial diligence, audited statements, transaction documents, bank records, and reconciled accounting exports may be stronger than a presentation or an interview recollection. For market sizing, a transparent calculation from documented inputs is generally more reviewable than an unsupported industry estimate.
Have AI flag contradictions, not decide which source wins. For example, it might notice that a presentation reports $6M of EBITDA while a spreadsheet shows $5.4M. The appropriate conclusion is not “use the higher number” or “average them.” It is to investigate the periods, adjustments, and definitions behind each figure.
Use Multi-Model AI Chat in One Shared Thread
When several models can work in one shared conversation, keep the evidence packet, instructions, and current ledger in that thread. This helps each model see the same context and makes it easier for a human reviewer to compare their reasoning. It also reduces the risk that one model is working from a stale version of the deal assumptions while another uses updated terms.
Shared context is not the same as independent validation. Models may repeat the same source error, inherit a bad premise, or agree because they were given the same misleading prompt. Treat each answer as analysis to inspect, not as a vote that establishes truth.
A practical division of labor is:
- Model A: extraction. Pull reported figures and source references from provided documents. Model B: challenge. Identify missing definitions, inconsistencies, and questions that would change the conclusion. Model C: calculation review. Recompute specified formulas and show the arithmetic, without supplying new source data. Human reviewer: adjudication. Check source documents, resolve definitions, approve assumptions, and decide what belongs in the memo.
Ask models to distinguish “I found this in the source” from “I infer this from the source.” That wording makes it harder for polished prose to conceal uncertainty.
Choose Parallel or Sequential Workflows Deliberately
Parallel workflows: broaden the search
Parallel work is useful when you want multiple independent perspectives on a bounded question. For example, ask separate models to review the same Skybridge materials for revenue-quality risks, customer concentration, and transaction-structure questions. Each should return evidence, not just conclusions.
Parallel review can reveal issues one model misses, but it can also create a pile of overlapping claims. Do not merge answers automatically. Compare them against the source packet and ledger. If three models repeat an unsupported market-size figure, you have three repetitions—not three sources.
Sequential workflows: preserve dependencies
Sequential work is preferable when later tasks depend on validated earlier outputs. First extract the historical financials. Next reconcile definitions and periods. Then calculate valuation metrics. Only after those steps should AI draft the valuation discussion.
For the Skybridge example, the sequence matters: confirm what the $42M price represents; verify the $6M EBITDA figure; then calculate the implied multiple. If the price is enterprise value and EBITDA is a comparable last-twelve-month figure, $42M divided by $6M equals 7.0x. If one number is equity value, or the EBITDA period includes unsupported adjustments, the comparison may be invalid even though the division is correct.
Use parallel workflows to discover and challenge. Use sequential workflows to calculate and write from approved inputs.
Pick an Orchestration Mode for the Work
Orchestration modes are simply ways to assign, order, and review tasks. The right mode depends on the risk and complexity of the work, not on how sophisticated the tool sounds.
- Single-model, human-led: Best for a small packet or a narrow task. The human provides sources, asks for a structured output, and checks each claim. Parallel review: Best for issue spotting across several diligence areas. Models independently inspect the same approved packet; a human reconciles the results. Sequential pipeline: Best when outputs depend on prior validation: extract, reconcile, calculate, analyze, draft. Human approval gates: Best for high-impact numbers. Require sign-off before a figure moves from source extraction into analysis, and again before it appears in the final memo.
For a transaction memo, a hybrid is often strongest: parallel reviews to find questions, followed by a sequential, human-approved path for the figures and final conclusions. Keep an audit trail of source versions, prompts, model outputs, and human changes. That makes later review faster and clarifies who accepted an assumption.
Verify Calculations Outside the Model
Asking AI to “double-check the math” is useful, but it is not enough. Enter material calculations into a spreadsheet or another controlled calculation environment, preserve the formula, and have a person inspect the inputs. In the memo, show the calculation or its basis where it helps the reader reproduce the result.
In the illustrative Skybridge case:
Proposed enterprise value: $42M. Illustrative seller-reported LTM EBITDA: $6M. Implied EV / EBITDA: $42M ÷ $6M = 7.0x.That result is only as reliable as its definitions and inputs. Before relying on it, verify calculations and confirm that enterprise value includes the correct treatment of debt, cash, and other purchase-price adjustments. Confirm that EBITDA covers the intended twelve-month period, uses a consistent accounting basis, and does not include adjustments that have not been supported.
Then test sensitivity. If verified EBITDA were $5M, the same $42M enterprise value would imply 8.4x. If it were $7M, it would imply 6.0x. These are arithmetic scenarios, not claims that either EBITDA figure is true. Label them as sensitivities so they cannot be mistaken for diligence findings.
Do not let a model silently fill gaps in a spreadsheet. If working capital, net debt, customer churn, or another input is missing, mark it as missing. A blank cell is safer than a fabricated estimate presented with false precision.
Treat Disagreement as a Signal for Better Decisions
Models may disagree about whether $42M is an attractive price, which risks deserve the most attention, or what evidence is missing. Do not resolve that disagreement by counting answers or asking the models to reach consensus. Disagreement is a prompt to inspect assumptions, definitions, and evidence.
For example, one model may call 7.0x “reasonable” while another calls it “expensive.” Ask each to expose the comparison set, metric definition, time period, and assumptions behind its judgment. If one relied on a benchmark with no source or an incomparable company set, discard that rationale. If the disagreement remains after checking the inputs, describe the uncertainty and its decision impact in the memo.
A useful follow-up prompt is: “List the factual premise, assumption, and source behind your conclusion. What evidence would change your view?” This turns disagreement into a diligence agenda. It may reveal that the real question is not the multiple, but whether EBITDA is sustainable, whether customer retention is weakening, or whether the deal terms shift risk to the buyer.
Write the Memo So a Reviewer Can Trace Every Claim
AI can draft after the evidence and calculations have been reviewed. Keep the memo concise, but preserve a clear line between findings and uncertainty. A practical structure is:
- Decision and rationale: State the recommendation and the conditions attached to it. Transaction overview: Present the $42M price as a proposed or illustrative term, with its definition and status. Financial analysis: Separate reported results, verified adjustments, and scenarios. Commercial and operating diligence: Tie conclusions to source-backed evidence and identify unanswered questions. Risks and mitigants: Explain what could impair value, what evidence supports the risk, and what action could reduce it. Open items: List unresolved questions, owners, and the evidence needed before approval.
Ask the model to draft only from the approved ledger. Then review every number, qualifier, and causal claim against the source. “Revenue grew 12%” needs a period, definition, and source. “Management expects 12% growth” is a different statement. Keep those distinctions intact during editing.
A Final Review Checklist
- Does every material number have a source, period, unit, definition, and status? Have you clearly separated facts, seller-reported claims, assumptions, and calculated values? Have calculations been reproduced outside the model, with inputs checked by a person? Are cited benchmarks traceable and comparable—or explicitly identified as unavailable? Have you investigated model disagreements instead of treating consensus as proof? Can a reviewer reproduce the memo’s key conclusions from the evidence and formulas?
AI can make diligence work faster and more systematic, but it cannot confer truth on an unverified input. Use shared model context to coordinate review, parallel workflows to surface issues, sequential workflows to protect dependencies, and human approval gates for consequential figures. The goal is not a memo with the most confident numbers. It is a memo where every number has a known origin, every assumption is visible, and every important conclusion can withstand scrutiny.