A CIM arrives at 8:37 p.m. The company is growing, the margin chart moves up and to the right, and adjusted EBITDA appears in six places. You ask an AI tool for an investment memo.
Thirty seconds later, you have five polished pages.
You do not yet have an underwrite.
The draft may contain accurate facts, seller claims presented as facts, rounded numbers from the wrong period, a management adjustment that sounds validated, and a recommendation no accountable investor has made. Fluency has compressed five different levels of evidence into one voice.
The better workflow is not CIM → prompt → memo.
It is:
Authorized sources → as-reported record → evidence ledger → checks and review → model → IC memo → decision.
The evidence ledger is the control point. It lets AI accelerate extraction and drafting without allowing it to silently decide which evidence controls, how an adjustment should be treated, or whether the deal is investable.
This is a practitioner workflow for organizing evidence. It is not an audit, quality-of-earnings report, legal review, cybersecurity assessment, valuation opinion, or investment recommendation.
Why a direct CIM-to-memo prompt fails
The failure is not simply that a model may invent a number. A more common problem is that it may use a real number in the wrong role.
Consider five statements:
- The CIM reports $11.2mm of adjusted EBITDA.
- Management says a $0.8mm cost will not recur.
- The analyst has not yet found support for the adjustment.
- The LBO uses $10.4mm of entry EBITDA.
- The deal should proceed only if the adjustment clears.
All five can be written in confident prose. They do not have the same evidentiary status.
The 2026 FINRA GenAI guidance says summarization and information extraction are the leading observed GenAI use case among its member firms. It also emphasizes approval, testing, provenance, logs, model versions, monitoring, and human review. FINRA directly governs member firms, not every PE fund, but the control lesson is useful: extraction is a workflow, not an assurance opinion.
Tamara Sakovska’s The Private Equity Toolkit describes a transaction process in which seller materials are followed by deeper data-room access, successive investment-committee submissions, phased diligence, and updated views of the company, valuation, structure, returns, and risks. The sequence matters. Later evidence should change the underwrite.
The evidence ledger makes that change visible.
The Claim Ladder
Fluent prose collapses categories. A defensible memo labels them.
- 01Source factThe monthly balance sheet reports $4.8mm of net working capital.
- 02Management representationManagement says a $0.8mm cost add-back will not recur.
- 03Analyst interpretationCash conversion may be weaker than the CIM implies.
- 04Model outputThe base case produces 2.2x MOIC at an 8.0x exit.
- 05Investment conclusionProceed only if the add-back and working-capital bridge clear.
The UpLevered evidence architecture
The architecture has three connected layers.
Layer 1: RAW+
RAW+ means reported → normalized → adjusted → modeled → decided.
The source is never destroyed:
- Reported: What the document, table, or management response says.
- Normalized: Units, signs, periods, and definitions converted into a consistent format.
- Adjusted: A reviewer-approved change with a stated rationale.
- Modeled: The value or assumption actually used in the underwrite.
- Decided: The investment implication and accountable owner.
This separation is more important than it looks. If extraction treats costs as positive, the model expects outflows as negative, and the export silently flips signs, an intelligent system can deliver a perfectly formatted wrong answer. The sign convention is a data contract, not a cosmetic choice.
Layer 2: The Evidence Ledger
For each material claim or line item, retain:
- underwriting question or claim;
- source document and exact page, table, or cell;
- exact source text or table label;
- as-reported value, sign, units, period, and definition;
- controlling source and any conflicting evidence;
- normalization or transformation;
- AI extraction and confidence flag;
- reviewer correction or override;
- final accepted value;
- deterministic check status;
- affected model line;
- thesis impact and memo destination;
- owner, reviewer, status, and required follow-up.
A senior reviewer does not need every field in the first view. The decision view should be concise. The audit view should remain one click below it.
The Evidence Ledger
| Claim | Source | As reported | Transformation | Accepted value | Status |
|---|---|---|---|---|---|
| FY25 revenue | CIM p. 42 | $52.0mm | None | $52.0mm | Tied |
| Adjusted EBITDA | CIM p. 45 | $11.2mm | Remove unverified $0.8mm add-back | $10.4mm | Review |
| Maintenance capex | Mgmt. claim | $0.6mm | Not yet supported by fixed-asset detail | Open | Question |
| NWC requirement | Monthly BS | $4.8mm | Normalize signs; use LTM average | $5.2mm | Tied |
Layer 3: The memo and model
The memo is a downstream artifact, not a separate truth system.
Every material number in the memo should resolve to:
- an accepted ledger entry;
- the current model output; or
- a clearly labeled analyst judgment.
If the memo says EBITDA is $11.2mm while the model uses $10.4mm, the system should not choose the more recent-looking number. It should fail the consistency check and send the conflict to the owner.
The controlled CIM-to-IC workflow
Step 0: Authorize the data path
Before uploading anything, clear the gates in the CIM confidentiality decision tree:
- document classification;
- NDA, VDR, process-letter, and engagement restrictions;
- firm approval for the exact product, plan, workspace, connected app, and feature;
- training, retention, subprocessor, access, deletion, and residency terms;
- data types inside the document; and
- the required logging and review process.
If the environment is not approved, use public, synthetic, redacted, or aggregated material. Do not let convenience decide the disclosure question.
Step 1: Define the decision before the extraction
The prompt should begin with what the team is trying to decide.
For an initial screen:
- Does the opportunity fit the mandate?
- Which claim drives price, leverage, or the investment thesis?
- What could make the deal non-investable?
- Which missing evidence must be resolved before the next spend decision?
For a later IC submission:
- What has changed since the prior view?
- Which assumptions cleared, weakened, or remain open?
- How did diligence change EBITDA, working capital, capex, debt capacity, price, structure, or downside?
- What decision is being requested now?
Extraction without a decision scope creates a pristine table of facts that may not matter. The useful endpoint is a defensible view of cash generation, debt capacity, risk, and open questions.
Step 2: Declare the source hierarchy
Write the hierarchy down before the AI run.
An illustrative hierarchy could be:
- executed documents and definitive terms for legal or financing questions;
- audited or reviewed historical financial statements;
- detailed, scope-matched schedules and raw operating data;
- third-party diligence work with accessible underlying support;
- management responses and forecasts;
- the CIM and teaser;
- external public research, with date and source;
- analyst inference.
The hierarchy changes by question. A debt commitment controls financing terms. A customer-level export may control concentration. A QoE schedule may control an EBITDA bridge only to the extent the team accepts its definitions and support.
When rounded narrative and a detailed table describe the same quantity, use the more granular, scope-matched evidence and record why. Do not quietly pick whichever number helps the case.
Step 3: Preserve the source package and version
Record:
- document name and version;
- received date;
- page count or worksheet set;
- hash or stable document identifier where the workflow supports it;
- permitted user group;
- model or product version used;
- prompt or workflow version; and
- output timestamp.
This is not bureaucratic overhead. A memo generated from CIM v3 and a model updated from data-room v5 can disagree while every individual sentence looks plausible.
Step 4: Run one bounded first pass
Do not ask for a recommendation. Ask for structured entries.
A useful instruction is:
For each requested item:
1. return the source document, page, table label, and exact supporting text;
2. preserve the reported value, sign, units, period, and definition;
3. distinguish source fact, management representation, and model inference;
4. mark missing or conflicting evidence as unresolved;
5. do not normalize, adjust, or choose a controlling source unless the rule is supplied;
6. return one ledger row per claim.The first pass should optimize for recoverability. A claimed 99% extraction rate is less useful than a reliable path from every output back to the source.
Step 5: Run deterministic checks
Use code, spreadsheet formulas, or explicit rules for questions that have objective answers:
- totals equal components;
- percentages use the stated denominator;
- year-over-year bridges reconcile;
- signs match the contract;
- periods and definitions are consistent;
- repeated facts agree across documents;
- EBITDA adjustments sum to the reported total;
- customer concentration adds to the expected population;
- debt and cash values use the same measurement date; and
- every memo number appears in the accepted ledger or current model.
Do not ask three models and average their confidence. Run one controlled pass, test it, and escalate failed checks and ambiguity.
Step 6: Review exceptions, not everything equally
Human review should concentrate on:
- a failed tie-out;
- conflicting sources;
- an unsupported adjustment;
- a definition change;
- an unclear sign or period;
- a blank the model tried to fill;
- an item that changes price, leverage, liquidity, or returns; and
- a conclusion that moves beyond the cited evidence.
This is where AI can save real time. The work product is still human-reviewed, but the reviewer starts with the exception population instead of rereading every page at the same depth.
Step 7: Convert uncertainty into source-linked questions
If maintenance capex is not disclosed or derivable, the answer is not a confident estimate. The answer is:
Not answerable from the available evidence. Request the fixed-asset register, historical repair and replacement spend, and management’s capex policy. Link the question to the affected free-cash-flow line.
The question should include:
- the exact source and number that created it;
- why the current evidence is insufficient;
- the model line and decision it affects;
- the evidence needed to close it; and
- the owner and due date.
Uncertainty is useful when it is routed. It is dangerous when it is polished away.
Step 8: Map accepted evidence into the LBO
The evidence ledger should feed:
- entry EBITDA and accepted adjustments;
- revenue and margin assumptions;
- working-capital requirements;
- maintenance and growth capex;
- debt terms and cash restrictions;
- tax, fees, and one-time items;
- value-creation initiatives;
- downside assumptions; and
- the exit case.
Preserve the extracted value, reviewer override, and modeled value separately. The AI LBO model-control guide explains how to freeze model invariants and build a variance bridge after any AI-assisted change.
Step 9: Draft the memo from accepted evidence only
The drafting system may use:
- accepted ledger entries;
- current model outputs;
- approved public sources;
- explicitly labeled analyst interpretations; and
- the current list of open questions.
It may not silently convert:
- a management representation into a verified fact;
- an unresolved adjustment into EBITDA;
- a market observation into a company forecast;
- a base-case output into a recommendation; or
- missing evidence into an estimate.
The memo should distinguish:
- what is known;
- what management says;
- what the team believes;
- what the model produces;
- what can break; and
- what decision is requested.
Step 10: Reconcile the package
Before circulation, confirm:
- memo numbers equal the current model;
- the model uses the accepted ledger values;
- all cited pages remain valid in the current document versions;
- open questions are visible rather than buried;
- the downside addresses the thesis-breaking risks;
- changes from the prior IC version are bridged; and
- one accountable human signs off on the recommendation.
Worked example: one add-back changes price, leverage, and the memo
Northwind Managed Services: From seller EBITDA to accepted EBITDA
Assume the synthetic Northwind case reports:
| Ledger item | Seller / source view | Reviewed treatment |
|---|---|---|
| FY25 revenue | $52.0mm | $52.0mm |
| Adjusted EBITDA | $11.2mm | $10.4mm |
| Included cost add-back | $0.8mm | Excluded pending support |
| Entry multiple | 8.0x | 8.0x |
| Opening leverage | 4.0x | 4.0x |
If the model accepts the seller adjustment:
- enterprise value is $89.6mm;
- opening debt is $44.8mm; and
- sponsor equity, including $4.0mm of combined transaction fees and funded minimum cash, is $48.8mm.
If the evidence ledger excludes the $0.8mm pending support:
- enterprise value is $83.2mm;
- opening debt is $41.6mm; and
- sponsor equity is $45.6mm.
The adjustment changes enterprise value and opening debt by $6.4mm and $3.2mm, respectively. The question is not whether AI extracted $11.2mm correctly. It did.
The decision questions are:
- What evidence supports the adjustment?
- Does the buyer accept it for price?
- Will lenders accept it for leverage?
- Does the downside assume it recurs?
- How should the recommendation change if it does not clear?
The evidence ledger keeps the source fact, management representation, reviewer treatment, model line, and memo conclusion separate. Until the item clears, the memo should say:
The seller presents $11.2mm of adjusted EBITDA, including a $0.8mm cost add-back that remains unverified. The current underwrite uses $10.4mm and treats the adjustment as upside.
The review ladder still owns the decision
Paul Gompers and Steven Kaplan’s Advanced Introduction to Private Equity describes a practical organizational ladder: junior professionals assemble data, spreadsheets, and presentations; associates oversee diligence and modeling; VPs supervise the work and investment thesis; senior professionals approve the investment.
AI can move work within the assembly layer. It does not erase the ladder.
A useful ownership map is:
| Work product | AI role | Human owner |
|---|---|---|
| Source index | Extract and classify | Analyst |
| Evidence ledger | Populate first pass and flag conflicts | Analyst / associate |
| Diligence questions | Draft from gaps and exceptions | Workstream owner |
| LBO updates | Perform defined changes and checks | Model owner |
| IC memo | Draft from accepted evidence | Deal team |
| Investment thesis | Challenge and structure evidence | VP / deal lead |
| Invest or pass | No delegated authority | Investment committee |
The faster the assembly becomes, the more important it is to make ownership explicit.
Common failure modes
Where the CIM-to-IC workflow breaks
| Trap | What goes wrong | How to catch it | Fix |
|---|---|---|---|
| Prompt straight to memo | The draft blends source facts, seller claims, assumptions, and recommendations into one confident narrative. | Choose three material sentences and demand the exact controlling source, page, classification, and reviewer. | Populate and review the evidence ledger before generating memo prose. |
| Silent normalization | Signs, units, periods, or definitions change without preserving the reported value. | Compare the raw source, extracted field, transformed field, and modeled value side by side. | Use RAW+ and retain each layer separately. |
| Missing becomes estimated | The output invents maintenance capex, working capital, or a market size because the source is incomplete. | Require every unsupported field to return unresolved rather than a plausible number. | Create a source-linked question with an owner, evidence request, and model impact. |
| Multiple-model voting | Three fluent answers are averaged even though they may share the same source or failure mode. | Ask what deterministic check proves the value instead of comparing confidence. | Use one controlled pass, objective tests, then targeted exception review. |
| Memo-model drift | Diligence changes the model, but the memo still carries the prior EBITDA, leverage, return, or recommendation. | Run a cross-artifact check on every material number and the decision statement. | Draft from the current accepted ledger and model, then publish a short variance log. |
Pre-IC evidence checklist
Before the package goes to investment committee
- Authorization: The exact documents, product, workspace, connected features, and users are approved.
- Versions: Every source document, model, prompt or workflow, and output has a stable version and date.
- RAW+: As-reported, normalized, adjusted, modeled, and decided values remain separate.
- Traceability: Every material claim has a source, page or cell, definition, period, and controlling-source decision.
- Checks: Totals, signs, periods, repeated facts, bridges, and cross-artifact values have been tested.
- Exceptions: Conflicts, unsupported adjustments, missing evidence, and failed checks have named owners.
- Model: Accepted ledger values flow into the current model and the variance bridge is understood.
- Memo: Source facts, management representations, analyst interpretations, model outputs, and conclusions are labeled.
- Downside: The case tests what can break rather than applying a cosmetic haircut to the base case.
- Decision: The requested action, open conditions, and accountable recommendation owner are explicit.
Sources and methodology
The evidence-ledger framework and RAW+ architecture are original UpLevered practitioner frameworks. They were generalized from hands-on work on document-to-model finance workflows, source-grounded financial-reasoning tasks, synthetic LBO cases, and investment-committee work products. No target, client, employer, or confidential deal data appears in this guide.
The workflow is also grounded in the following sources:
- Tamara Sakovska, The Private Equity Toolkit (Wiley, 2022), especially chapters 7, 10, and 11 on business-plan analysis, seller materials, data rooms, phased diligence, adviser work, transaction documents, and successive IC views.
- Paul Gompers and Steven Kaplan, Advanced Introduction to Private Equity (Edward Elgar, 2022), chapters 4 and 10 on qualitative and quantitative investment analysis and the professional review hierarchy.
- FINRA, 2026 Annual Regulatory Oversight Report: GenAI, for securities-industry guidance on approval, testing, provenance, logging, monitoring, and human review. Applicability depends on the firm.
- NIST AI Risk Management Framework, a voluntary cross-sector framework for governing, mapping, measuring, and managing AI risk.
- BankerToolBench, an April 2026 preprint built with input from 502 investment bankers. Across its 100 end-to-end banking tasks, the best tested model failed nearly half of the expert rubric criteria, and bankers rated none of its outputs client-ready. Results are model, task, and scaffold specific.
Revision History
- : Original publication. Added RAW+, the Evidence Ledger, the Claim Ladder, the controlled ten-step workflow, the Northwind worked example, and the pre-IC evidence checklist.
Frequently asked questions
Can AI write a private equity investment committee memo?
AI can help draft and reconcile defined sections, but the memo should be generated only from reviewed evidence and the current model. The deal team must still own source selection, assumption acceptance, downside design, investment judgment, and the recommendation.
What is an evidence ledger in private equity?
An evidence ledger is a source-tracked register of material claims and numbers. It preserves the source document, page or cell, as-reported value, definition, transformation, reviewer override, accepted value, check status, model impact, and memo destination.
Should a CIM be treated as the controlling source?
A CIM is seller-prepared marketing material and is the starting point for underwriting, not the automatic controlling source for every claim. Material items should be reconciled to more detailed, scope-matched evidence as diligence progresses.
How do you prevent AI hallucinations in an IC memo?
Do not draft directly from an unreviewed document set. Require page-level citations, preserve an as-reported layer, run deterministic checks, mark unsupported items as unresolved, restrict drafting to accepted ledger entries, and reconcile every memo number to the current model.
Does an AI evidence ledger replace financial, legal, or commercial diligence?
No. It is a control layer for organizing evidence and review. It is not an audit, quality-of-earnings report, legal opinion, cybersecurity review, valuation, or investment recommendation.
Continue through the control system
- Clear the data gate: Use the CIM confidentiality decision tree before any deal document enters an AI workflow.
- Control the model: Use the AI LBO audit standard after accepted evidence moves into Excel.
- Apply the reviewer lens: Use What a VP Marks Up in Your First LBO to reconcile model mechanics with the investment recommendation.
- See the full map: Return to the AI in Private Equity field guide.
The goal is not an IC memo that appears faster. It is a decision package where every important sentence can survive the question, “What controls this?”
Stay sharp. Subscribe to Deal Flow Bullet.
PE frameworks, AI-in-PE workflows, and deal analysis for middle-market practitioners. Free, every two weeks.