How to Draft an Investment Committee Memo With AI

An IC memo records the decision the committee is being asked to make. Good writing helps, although the real work is assembling the evidence and showing where the deal team’s judgment begins.

AI can shorten sections, organise diligence findings and produce an early draft. It can also turn several uncertain inputs into a confident paragraph. The second outcome is dangerous because the prose often looks better than the underlying work.

I start with the sources and the decision points. Drafting comes later.

Fix the memo structure

Formats differ by firm, though most cover familiar ground:

Transaction overview and recommendation
Business and market
Investment thesis
Historical performance and forecast
Valuation and returns
Capital structure
Value-creation plan
Risks and mitigants
Diligence findings
Management and governance
Exit routes
Open items and conditions to approval

Lock the headings, expected exhibits and approval points before generating text. Otherwise, the output tends to resemble a consulting report: smooth, broad and detached from the actual decision.

The committee needs the facts that change the answer. It does not need every fact the team learned.

Put the memo in a controlled workspace

The best setup depends on where the evidence already lives. Cowork can work with a selected local folder through Claude Desktop. ChatGPT Work can use a local folder in the desktop app where local access is enabled, and a ChatGPT Project can hold the relevant files and instructions. Claude Code and Codex suit a deal directory with an evidence CSV, source extracts and a memo file because they can work across the folder and run checks on citation tags.

A simple directory is enough:

/IC Memo
  /sources
  evidence_matrix.csv
  memo_outline.md
  draft.md
  citation_audit.csv

Limit the agent to this directory. Keep the source set frozen during a drafting pass, and date the evidence matrix. If a new diligence report arrives, add it deliberately and rerun the affected sections.

Write from evidence tags

Give every evidence row an ID such as [FIN-014], [CUST-022] or [MKT-008]. The ID survives edits more reliably than a bare footnote number. Here are three example rows:

[CUST-022] Gross revenue retention | 94% | FY2025 | CDD report p.38 | Verified
[CUST-023] Top-10 customer concentration | 31% of revenue | FY2025 | CIM p.54 | Verified
[FIN-014] Price contribution to growth | 3.2 percentage points | FY2025 | QoE databook, Price tab | Verified

Now draft a single section:

Write the customer-quality section in draft.md using only evidence rows tagged CUST or FIN and marked Verified. Maximum 250 words. Put the evidence ID immediately after every factual claim. Keep reported facts separate from the deal team’s interpretation. State the period and definition for every retention or concentration metric. If two sources disagree, show the disagreement in a bracketed editor note. Do not average, reconcile or choose between them. End with two implications for underwriting, each labelled Deal team view.

A useful output would look like this:

Gross revenue retention was 94% in FY2025 [CUST-022]. The top 10 customers represented 31% of FY2025 revenue [CUST-023]. The base therefore combines strong account retention with moderate concentration. Deal team view: the retention figure supports the downside case only if the definition excludes price increases and acquisitions. Deal team view: the three largest contract renewals should be reviewed individually before final approval.

The first two sentences are traceable. The next sentence is interpretation. The final two lines say what the team still needs to decide. That separation makes the paragraph easier to challenge in a meeting.

Run the citation audit

Do not ask the same agent whether the memo is good. Give it a narrower audit task:

Read draft.md sentence by sentence. Create citation_audit.csv with sentence_number, sentence_text, claim_type (fact/calculation/judgment/recommendation), evidence_ids, source_opened, supports_claim, period_matches, definition_matches and issue. Open every cited source. Use TRUE, FALSE or UNCLEAR. Flag any factual sentence with no evidence ID, any citation that supports only part of a sentence, and any number whose period or definition differs from the source. Do not edit the memo.

Then review three groups manually: every key underwriting assumption, every sentence with two or more citations, and every claim the IC is likely to repeat in its approval conditions. The common failure is partial support. A source may prove 94% retention while saying nothing about why retention is high.

Calculations need their own treatment. If the memo says EBITDA reaches £24 million after £6 million of initiatives, the audit should point to the base case, each initiative, timing and any implementation cost. A correct sum can still be a weak investment case.

Before circulation, strip unused evidence rows, refresh the section owners and rerun the audit. The final citation_audit.csv becomes part of the deal record rather than another document somebody has to rebuild later.

Build the evidence matrix

The evidence matrix sits between the deal folder and the memo. Give each claim its own row with:

Metric or conclusion
Source
Page, workbook tab or meeting reference
Date and period
Owner
Confidence or status
Memo section

“Strong growth, recurring revenue and low churn” contains three claims. Treat it that way.

Numbers also need units and definitions. Revenue may be reported, pro forma, organic or constant currency. EBITDA may be reported, adjusted or adjusted again by the deal team. A label lost in the matrix will usually be lost in the memo as well.

The staged CIM review should feed this matrix. Extraction and citation work from the start of diligence then becomes useful again at IC.

Draft one section at a time

Give the tool an approved set of evidence rows and a defined section. For example:

Draft the market section using only rows tagged “Market” in the evidence matrix. Separate sourced facts from the deal team’s judgment. Preserve every citation tag. Omit growth, competitor and market-share claims that are absent from the matrix. Close with the implications for the investment decision.

The section owner should then review it against the source.

Generating 20 pages in one pass feels faster. It usually creates a long clean-up on the night before IC, when somebody discovers that a polished paragraph contains figures nobody can trace.

Write the thesis as a chain

Standard thesis bullets appear in almost every deal: market leader, attractive growth, recurring revenue, multiple value-creation levers. A stronger thesis links cause and effect:

What is true about the market or asset?
Why is the company positioned to benefit?
What must management and the sponsor do?
How does that change earnings and cash?
What will an exit buyer pay for?

Ask the tool to flag broken links and unsupported assertions. If the memo cites a pricing opportunity without customer research, historical price data or an implementation plan, it is still an idea rather than a value-creation lever.

Build the case against the deal

By memo stage, the file structure reflects months spent advancing the transaction. The risks section can become a list of objections followed by reassuring mitigants.

Run a separate challenge pass using a clean context and the same evidence matrix. Questions might include:

Which three facts could break the base case?
Which conclusions rely mainly on management representations?
Where does the forecast exceed the historical range?
Which “mitigated” risk has moved into a covenant, indemnity or earn-out?
What would a competing bidder say about the asset?
Which downside is absent from the model?

The exercise should expose weak points in the case. Drama is unnecessary. The committee needs to know where the underwriting is doing real work.

Rehearse the committee questions

Committee members approach the paper from different angles. The operating partner may focus on execution capacity. A credit-minded member will examine cash conversion, covenant headroom and downside liquidity. The sector specialist will test market structure. Someone who has lived through a similar loss will concentrate on the risk others consider remote.

Generate questions by role, then rank them by how much the answer could change the decision. “What are the key risks?” adds little. A better question is: “If the top two customers renew at current volume but at 5% lower price, what happens to year-two covenant headroom?”

Answers should tie back to the model and matrix. Unanswered questions become open diligence items.

Leave source tags in the working draft

Citation tags are ugly and useful. A sentence might end with [QoE p.34], [Model: Debt Schedule!F42] or [Management meeting, 12 Sep]. Keep the tags through senior review. Convert or remove them only when the final format requires it, and retain a sourced working copy.

Verbal answers need the same treatment. Record who said it, when and in what forum. “Management says” is too vague to audit later.

Reconcile the numbers separately

A smooth draft can hide numerical errors. Run a standalone check:

Sources and uses agree with the model
Entry and exit enterprise values use the stated definitions
Net debt and debt-like items match the transaction bridge
Reported and adjusted EBITDA are clearly labelled
Revenue, margin and cash flow use consistent periods
Leverage and covenant headroom tie to the debt schedule
IRR and MOIC match the approved model
Sensitivities use the current model version

The tool should compare the memo with approved outputs and flag differences. It should never calculate a missing return and quietly add it.

Version control matters. A memo linked to model v27 may become wrong when v31 is circulated after the debt terms change. Put the model version and timestamp in the evidence matrix.

Separate fact, judgment and decision

The clearest papers show the distinction:

Fact: Customer A represented 12% of FY2025 revenue.

Judgment: The concentration is manageable because the relationship spans several sites and products.

Decision: Proceed subject to confirming renewal terms before signing.

AI often blends these categories into one paragraph. Keep them visible. The recommendation becomes easier to assess and more credible.

Cut the investment-speak

Once the analysis works, use AI to remove repetition, identify undefined jargon and shorten long sections. Give it a word limit and ask which sentences can go without losing a decision-relevant point.

Watch for “compelling opportunities”, “strong tailwinds” and “significant potential”. These phrases occupy space without committing to a fact.

Specific language is better:

“Price increased 6% with no measurable change in churn” instead of “demonstrated pricing power”
“Three of the top ten customers have contracts expiring within 12 months” instead of “some renewal risk”
“The base case requires two acquisitions each year” instead of “M&A provides upside”

Plain IC writing usually reflects precise thinking.

Final review

Someone who did not draft the section should ask:

Does every material claim have a source?
Do the figures tie to the current model?
Are reported and adjusted metrics separate?
Do the thesis and value-creation plan connect to evidence?
Does the downside case test the assumptions that matter?
Are open items labelled as open?
Are the mitigants actions, or restatements of the risk?
Is the requested approval clear?

A named person should sign off. Investment decisions need an accountable owner.

In practice

The sequence is straightforward: evidence, draft, challenge, reconcile. Without the evidence matrix, AI produces a cleaner version of the team’s current beliefs. With the matrix, it helps produce a memo that shows the facts, the judgments and the decision the committee is being asked to underwrite.

Tool references

Share0

Tweet0