An AI deal memo is not a decision. It is a first draft of the facts. An investment committee memo has two layers: the assembled record of what the deal is, and the judgment about whether to do it. The first layer is retrieval, formatting, and reconciliation, work that consumes most of an analyst's memo hours and produces none of the insight. The second layer is the recommendation, the risk framing, and the conviction, and no model should write it. The argument of this piece is that AI belongs on the first layer entirely and the second layer not at all, and that firms confusing the two either underuse the tool or trust it past its competence.
The distinction matters because the memo is where a deal gets decided, and a memo that took four days to assemble leaves less time to argue about the parts that count.
Key Takeaways
An AI deal memo drafts the factual layer of an IC memo: market context, financial summary, rent roll reconciliation, and diligence status. It does not draft the recommendation.
Development teams that automate the first draft report IC prep dropping from three to four days to six to eight hours, per Build.inc, freeing the reconciliation hours, not the judgment hours.
A 2024 study found LLMs fabricate information in up to 41 percent of finance-domain queries, which is why every drafted number must trace to a cited source document, not to the model.
61 percent of institutional investors reported using AI for market analysis in 2025, up from 22 percent in 2023, per industry survey data, but adoption is not the same as trusting the output unverified.
The recommendation, the risk narrative, and the conviction are the parts of the memo AI cannot write, and they are the parts that were always the point.
What sections of an IC memo can AI actually draft?
AI can draft the sections of an IC memo that are assembly rather than argument: the market context pulled from supply, absorption, and pipeline data; the financial summary formatted from pro forma inputs; the rent roll and T-12 reconciliation; and the diligence status log. These are retrieval and formatting tasks, and they consume the majority of memo hours while producing none of the judgment.
The pattern across the sections is the same. Each is a compression of source documents into a structured narrative. The market context section restates third-party market data. The financial summary restates the model. The tenant summary restates the rent roll. None of these requires an opinion, and all of them are slow to produce by hand because the analyst is transcribing and cross-referencing rather than thinking.
Memo section | What it requires | AI first draft? |
Market context | Retrieve and summarize supply, absorption, pipeline | Yes |
Financial summary | Format pro forma inputs, restate returns | Yes |
Rent roll and T-12 reconciliation | Extract, structure, cross-check line items | Yes |
Diligence status | Track outstanding items and dates | Yes |
Risk framing | Judge which risks bind the return | No |
Recommendation | Argue the decision and own it | No |
Build.inc reports that development teams automating this first-draft workflow cut IC prep time from three to four days to six to eight hours. The hours saved are the reconciliation hours. The rent roll still has to tie to the T-12, the T-12 still has to tie to the model, and a model that does this cross-check in seconds returns the analyst's afternoon. For the mechanics of those inputs, see the rent roll and pro forma glossary entries, and how extraction feeds them in our piece on how AI document extraction compresses the diligence cycle.
Why should AI never write the investment recommendation?
AI should never write the recommendation because the recommendation is a judgment about risk and conviction, and a model asked to produce one will manufacture confidence it does not have. A 2024 study found large language models fabricate information in up to 41 percent of finance-domain queries. A recommendation is exactly the kind of plausible, unverifiable claim they invent most fluently.
The failure is specific and it is dangerous because it reads well. Ask a model to write an investment memo for a property and it will not say it lacks the data to recommend. It will produce a confident paragraph, invent a supporting margin or comp if one is missing, and format the whole thing to look like an analyst wrote it. The 41 percent finance-domain fabrication rate is not a rounding error on a low-stakes task. It is the base rate on the exact class of numeric, hard-to-verify claim a recommendation depends on.
This is why the drafting boundary is not a preference but a control. The factual layer is verifiable: every number the model writes should point to a source document a human can open. The recommendation layer is not verifiable in the same way, because it is a forward opinion, and a forward opinion the model fabricates looks identical to one it reasoned to. The expert-voice line worth keeping: a model can assemble the case for a deal, but the moment it is asked to render the verdict, it will invent the evidence it is missing rather than admit the gap.
The institutional trend confirms the boundary rather than erasing it. Industry survey data shows 61 percent of institutional investors used AI for market analysis in 2025, up from 22 percent in 2023. Analysis and drafting are the factual layer. The committee still votes.
How does an operator keep an AI-drafted memo trustworthy?
An operator keeps an AI-drafted memo trustworthy by enforcing source citation on every drafted figure, keeping a human on the recommendation, and treating the draft as a starting point to be verified rather than a finished document. The control is not the model's confidence score. It is whether each claim traces to a document the analyst can open and check.
The workflow that holds up has three rules. First, the model drafts only from provided source documents, never from its own memory, which removes the invented-comp failure at the root. Second, every figure in the draft carries a citation back to the page and field it came from, so verification is a click rather than a re-underwrite. Third, a person writes the recommendation and owns it in front of the committee. This is the same human-in-the-loop logic that governs extraction: the machine proposes structured facts, the human disposes of the decision. We argue the general case in why source citations matter more than accuracy in AI extraction.
The payoff of the discipline is asymmetric. Done right, the analyst spends the reclaimed hours on the parts of the memo that decide the vote: which assumption is the return most sensitive to, which risk is underpriced, whether the exit is credible. Done wrong, the firm ships a fluent memo built on numbers no one verified, and the committee votes on fiction that reads like analysis.
Frequently Asked Questions
Can AI write an entire investment committee memo?
No. AI can draft the factual sections of an IC memo, such as market context, financial summary, and rent roll reconciliation, but it should not write the investment recommendation. The recommendation is a judgment about risk and conviction, and a 2024 study found LLMs fabricate information in up to 41 percent of finance-domain queries, so a machine-written recommendation invents confidence it does not have.
How much time does AI save on IC memo preparation?
Build.inc reports that development teams automating the first-draft workflow cut IC prep time from three to four days to six to eight hours. The savings come from the reconciliation and formatting work, not the analysis, so the reclaimed hours go to judging the deal rather than assembling the facts.
How do you stop an AI deal memo from including made-up numbers?
Require every drafted figure to cite the source document it came from, and let the model draft only from provided documents rather than its own memory. When a number traces to a specific page and field, verification is a click, and the invented-comp failure that produces fabricated figures is removed at the source.
Conclusion
The investment committee memo has always had two layers, and the industry mislabeled the slow one as the important one. Assembling the facts, reconciling the rent roll to the T-12 to the model, formatting the market context, was never the analysis. It was the tax paid before the analysis could start. An AI deal memo can pay that tax in hours instead of days, provided every number it writes points back to a document a person can open. What it cannot do, and should never be asked to do, is render the verdict. The 41 percent finance-domain fabrication rate is the reason the recommendation stays human. For the operator, the shift is not that AI decides deals. It is that the analyst finally has the hours to argue about the parts that decide them.