How to Write an Investment Memo with AI: A Decision-Record Template

Use AI to build the evidence record behind an investment memo: source inventory, historical facts, peer comparisons, scenario calculations, risk evidence, and a first draft tied to those materials. The analyst owns the variant view, forecast assumptions, valuation method, position recommendation, and final wording. A strong memo records what would falsify the thesis, what the committee decided, and how that decision was reached.

This guide is a workflow and template based on public-source research and control principles. It does not recommend a security or benchmark memo-writing products. We build AllMind, which appears once in the tool section, and that paragraph is our own case.

The memo is a decision record

An investment memo serves three moments:

1. Before the decision: it forces the evidence, assumptions, alternatives, and risks into one reviewable record.
2. At the decision: it captures what the committee accepted, rejected, or changed.
3. After the decision: it lets the team distinguish a bad outcome from a bad process and update the thesis without rewriting history.

That purpose determines the structure. A company overview may be necessary, but a well-written overview is not yet an investment memo.

Assign each claim an owner

AI involvement becomes safer when the memo labels the type of work.

The generated system may propose an inference. It cannot turn that inference into an observed fact by writing it confidently. Keep these labels in the drafting data, even if the final prose reads naturally.

A copyable investment memo template

1. Decision request

• proposed action;
• security, instrument, and account scope;
• suggested size or research status;
• time horizon;
• decision deadline and next catalyst;
• named decision owner.

State the ask first. A committee should not have to infer it from the valuation section.

2. Thesis in two claims

Each claim should have a causal form:

Because [observable mechanism], we expect [operating or financial result] over [time window], while the current market expectation appears to be [contrasting expectation].

Then add one line for the evidence that would make the claim false. If the thesis cannot be expressed as a mechanism and a test, more research is needed.

3. What appears priced in

Show the bridge from market value to operating assumptions. Depending on the business, that may be a reverse DCF, implied margin, volume, take rate, loss ratio, credit cost, or terminal return assumption. Preserve the model and as-of date.

AI can run the arithmetic and sensitivities. The analyst chooses the model and decides which implied assumption is economically meaningful.

4. Evidence ledger

Do not hide contrary evidence in a generic risk list. Put it beside the claim it challenges.

5. Business and financial evidence

Include only the history necessary to evaluate the thesis:

• business model and revenue units;
• segment and geographic economics;
• key operating drivers with definitions;
• historical statements and cash conversion;
• balance-sheet or financing constraint;
• capital allocation;
• peer evidence relevant to the mechanism.

For US public issuers, use the SEC's filing search and preserve the filing accession and period. The SEC's EDGAR API documentation distinguishes submissions metadata and standardized company facts. Neither replaces reading the filing notes for definitions and changes.

6. Scenarios

Build scenarios from observable operating variables.

Probability weights originate with analyst judgment. Show expected-value math if it affects the decision, and keep the unweighted outcomes visible. A single target price can conceal a distribution the committee needs to see.

7. Risks and pre-mortem

Write risks as mechanisms with indicators:

Then run a pre-mortem: assume the position underperformed materially by the review date. List three plausible causes, the earliest evidence each would have produced, and whether the proposed monitoring plan would catch it.

8. Catalysts and monitoring

Every catalyst needs a date or observable condition, the expected evidence, and an owner. A product launch is not automatically a catalyst; specify what disclosure would change the forecast or market expectation.

9. Decision log

Record:

• decision and size/status;
• attendees and decision owner;
• assumptions changed in committee;
• evidence requested but missing;
• conditions for adding, reducing, or exiting;
• next formal review date.

Do not overwrite the original recommendation after the meeting. Store the committee changes as a new record.

Use AI in evidence-first passes

Pass 1: source inventory

Ask for the required documents and missing evidence. The output at this stage is a source table; drafting comes later. Reject stale periods, duplicate documents, and sources outside the permitted set.

Pass 2: extraction and calculation

Populate historical facts, KPI definitions, peer rows, and scenario inputs with source links. Keep calculations visibly separate from reported values. Flag a restatement or definition change instead of automatically bridging it.

Pass 3: opposition research

Ask the system to find evidence against each thesis claim, alternative causal explanations, and base-rate cases. Require sources. Generic “regulatory risk” bullets do not count.

Pass 4: draft

Draft the business and evidence sections from the accepted ledger. The analyst writes or substantially rewrites the thesis, what-is-priced-in section, risk ranking, and recommendation. Remove any sentence whose evidence status is unclear.

Pass 5: verification

Reopen every thesis-critical source, recalculate the valuation bridge, compare the generated model with the prior version, and confirm that the final memo contains only accepted evidence. The reviewer should see what changed after the AI draft.

Select the tool for the memo's evidence universe

A general enterprise assistant can structure the memo, critique arguments, and work over a carefully supplied source pack. A document-analysis platform fits a private-markets memo whose evidence is a bounded data room. A licensed-search platform fits a memo driven by broker and expert content. A specialist financial-data product can populate the historical model.

Our own AllMind already licenses the broker research, expert transcripts, filings, fundamentals, estimates, and market data those specialists sell separately, combines them with a firm's own notes, models, and documents, then generates Word, PowerPoint, or Excel outputs with source traceability. Only the firm's own systems require onboarding and permissions work, so a single public-document memo can run on the licensed corpus alone. Use the same source pack and template across products. Score corrections, missing evidence, and reviewer effort as separate measures.

Review and recordkeeping

The NIST AI Risk Management Framework offers a useful control vocabulary: govern the workflow, map its context and consequences, measure failures and corrections, and manage changes. It is not an investment standard, but it helps turn “human review” into named responsibilities and tests.

Registered investment advisers should have counsel determine how AI-assisted materials fit the firm's recordkeeping policy. 17 CFR 275.204-2 is the primary federal books-and-records rule for registered advisers; this article does not interpret its application to a specific memo, prompt, or intermediate draft.

At a practical minimum, retain the approved source list, evidence ledger, model version, generated draft or change record, analyst edits, final memo, and decision log under the firm's policy.

What this template cannot establish

No memo template proves that a thesis is correct, and no generated draft demonstrates time saved or better performance. We did not test any product, evaluate private data licenses, or assess a firm's legal duties. The structure is deliberately adaptable across public equity, credit, and private markets, so teams must add asset-class-specific sections.

Start by building the evidence ledger for one recently approved idea and compare it with what the committee actually debated. The missing columns will reveal where automation can help and where the process needs judgment and clearer records.

Sources and methodology

• SEC Search Filings and EDGAR APIs, for public issuer evidence architecture.
• NIST AI Risk Management Framework, for the workflow-control vocabulary.
• 17 CFR 275.204-2, cited as primary adviser recordkeeping text. Application requires firm counsel.