AI Deal Screening: Draft a Memo from a CIM

A CIM comes in. The adjusted EBITDA in the executive summary doesn't match the bridge in the back, the customer concentration chart is an image, and your team needs to decide whether the deal deserves diligence budget. Claude or ChatGPT will happily write a screening memo from it. You can only trust that memo if every figure points to a page and a quote that someone on your team can check.

This is the method we use when we set up screening workflows for investment teams. The model extracts and drafts. Your team checks the numbers and makes the call. Anthropic builds its own financial-services agents on the same terms: they draft work "for review by a qualified professional" and stage every output for human sign-off (Anthropic).

What a screening memo contains

No industry standard exists for a screening memo. Most published templates cover the full IC memo, and the screen gets described loosely. Its job is narrow: decide whether a deal is worth diligence time. One private credit template from F2, a vendor in this space, calls it a one-pager with five parts: company overview (including customer concentration), financial overview, investment merits, risks and mitigants, and the process and transaction overview (F2).

Here is a field list to start from. Adapt it to your own template.

Section	Fields	Where it usually sits in the CIM
Deal snapshot	Company, sector, sponsor, banker and process, transaction type, size, instrument, use of proceeds	Cover, executive summary, proposed terms (debt CIMs)
Business overview	What it sells, recurring vs. transactional revenue, customers, geography, management	Executive summary, products and management sections
Financial summary	Two or three historical years, LTM, one or two projected years: revenue, growth, gross margin, reported EBITDA, adjusted EBITDA, margin, capex	Financial section near the back, with headline figures up front
EBITDA bridge	Reported to adjusted, each add-back with its amount and page	Financial section or appendix
Customers	Top 1, 5 and 10 as a percent of revenue, contract length, retention	Customer pages and the financial section
Credit terms (debt deals)	Proposed structure, leverage, pricing, covenants, collateral	Proposed terms section
Merits and risks	A handful of each, with every risk paired to a mitigant	Your team's synthesis
Open questions	What the CIM doesn't answer, and where it contradicts itself	Your team's synthesis
Recommendation	Pass, further diligence, or hard pass	Your team

The fields follow F2's template, the CIM layout described by Mergers & Inquisitions, and the credit additions (facility type, leverage, pricing, collateral, covenants) listed by Hebbia. On length, published sources disagree widely. Match your firm's precedent memos and give the model that target.

Know where the numbers live

CIMs follow a familiar order. Headline revenue and EBITDA show up early in the executive summary, the detailed financials sit near the end, and customer data appears in both places (Mergers & Inquisitions). One sell-side advisor's guide puts the financial section at roughly 20 to 40 pages, with an adjusted EBITDA bridge that itemizes each add-back and a top-10 customer table, often with names redacted (CT Acquisitions).

So the same EBITDA figure can appear in the summary, the financial section, an appendix and a chart. The same guide lists conflicting EBITDA bridges across sections as a diligence problem. When two versions disagree, that belongs in the memo as an open question.

Set up before you upload

Start with policy. CIMs are confidential, and analysts need to follow their firm's AI policy before uploading one (Financial Edge). Use only the workspace your firm has approved. We cover the data questions in Is it safe to put a CIM into ChatGPT?

Then set up a Project with your memo template, your screening criteria and two or three past memos as reference. Claude Projects are available on every plan, and sharing one across your organization requires Team or Enterprise (Claude Help Center).

Check page limits before you trust what the model read. In the Claude app, PDFs can run to 1,000 pages, but Claude reads charts and graphics only in PDFs of 100 pages or fewer. Longer files get text only (Claude Help Center). The CT Acquisitions guide puts middle-market CIMs at about 80 to 120 pages, so plenty will cross that line. If the bridge or the customer table is an image, upload the financial section as its own PDF. On ChatGPT, OpenAI's help center has described visual reading of PDF images and charts as an Enterprise feature, with other plans extracting text only (OpenAI). Plans change, so confirm what yours does.

One more setup point. When a Claude Project's files approach the context limit, paid plans switch to a retrieval mode that pulls relevant passages instead of reading everything (Claude Help Center). Retrieval can surface a customer figure and miss the footnote that qualifies it. For a screen, attach the CIM to the conversation itself and keep the Project for the template and examples.

Step 1: Extract into a ledger before writing anything

Don't ask for the memo first. Ask for a table of facts, with a page and a word-for-word quote for each, and give the model explicit permission to say a figure isn't there. Anthropic recommends exactly this for long documents: pull exact quotes first, then work from them, and let the model say it doesn't know (Anthropic). Put the document before the question. Anthropic reports that queries placed at the end improved response quality by up to 30% in its tests (Anthropic). OpenAI's GPT-5.2 guide gives similar advice for uncertain inputs: never invent exact figures (OpenAI).

A prompt we'd start from:

Extract facts from the attached CIM for a screening memo. Do not write the memo yet.

Build a table with one row per figure. Columns:
ID | Field | Value | Unit and scale | Period | CIM page | Verbatim quote | Status

Extract:
- Revenue, gross profit, reported EBITDA and adjusted EBITDA for every
  historical year, LTM and every projected year shown
- Capex, and any free cash flow or cash conversion figure
- Top customer, top 5 and top 10 as a percent of revenue, with the period
- Recurring vs. non-recurring revenue split
- Transaction size, instrument, use of proceeds, and proposed leverage
  and pricing if stated

Rules:
- Use the page number printed on the CIM page. Note it if it differs from
  the PDF page number.
- Copy the quote exactly. For a table, quote the row label and column header.
- State the unit and scale ($000s, $m, %) for every value.
- If a figure appears on more than one page, give each its own row.
- If you cannot find a figure, write NOT FOUND. Do not estimate.
- If a table is an image you cannot read clearly, write UNREADABLE and the page.

The output is your ledger. Everything after this step works from it.

Step 2: Build the EBITDA bridge

A private company's adjusted EBITDA means whatever its add-back schedule says it means, and Mergers & Inquisitions flags the figure as discretionary and worth scrutiny. A model can copy the right number and still mislabel what it represents. One credit memo guide recommends showing reported and adjusted EBITDA side by side, listing every add-back, and computing leverage on both bases. In its example, rejecting a single synergy add-back moves leverage from 3.1x to 3.4x (uwriter).

Using only the CIM, build the bridge from reported to adjusted EBITDA
for each year shown.

For each add-back give: ID, description, amount, year, CIM page, verbatim quote.
Tag each add-back as one of: one-time, owner or related party, pro forma,
run-rate, synergy, other.
Flag any add-back that appears in more than one year.
Show the starting figure the CIM uses (net income, operating income or
reported EBITDA) and its page.
Do not net, combine or reclassify items. If the CIM's bridge does not sum,
say so.

A cost labeled one-time that recurs year after year belongs on the list of questions for the sponsor, which is why the prompt asks for repeats.

Step 3: Check the arithmetic with code

Both tools can run Python in a sandbox. Claude's code execution can work on uploaded files and recompute figures (Claude Help Center), and OpenAI's code interpreter tool does the same (OpenAI). Ask the model to load the ledger and bridge and to:

re-add every bridge and compare it to the CIM's stated adjusted EBITDA
recompute margins, growth rates and concentration percentages
compare each executive summary figure to the same figure in the financial section
compute leverage on reported and adjusted EBITDA if the CIM states the debt

Ask for a list of mismatches and leave the corrections to your team. A useful line reads something like: "LTM adjusted EBITDA: the executive summary (p. 4) and the bridge (p. 58) differ, and the bridge items do not sum to either figure." Your team decides what it means.

Formatting and citations make an answer look checked without checking it. OpenAI says Structured Outputs guarantee the shape of a response, not the correctness of its values (OpenAI). Anthropic's citations feature guarantees valid pointers into the document, which isn't the same as a correct reading of the page (Anthropic). OpenAI's citation guide tells builders to confirm that each citation actually supports its claim (OpenAI).

Step 4: Draft the memo from the ledger only

Now ask for the memo, and tie it to the checked tables.

Draft the screening memo in the attached template. Use only figures from
the ledger and bridge above.

- After every figure, add the ledger ID and CIM page, like [L12, p. 58].
- Label each statement: FROM CIM, OUR CALCULATION (show the formula),
  or OUR VIEW.
- Put every mismatch and NOT FOUND item under Open Questions.
- Leave the Recommendation section blank.
- When done, list any sentence you cannot support with a ledger row,
  and remove it from the draft.

The last instruction follows Anthropic's advice to have the model find support for each claim and retract what it can't back up. Anthropic also notes that these techniques reduce hallucinations without eliminating them (Anthropic).

If you build this through the Claude API instead of the app, citations return page ranges for PDFs, but they can't be combined with structured outputs in the same request, and scanned PDFs without extractable text can't be cited at all (Anthropic).

Step 5: Review it like a junior analyst's draft

The reviewer's checklist:

Open the cited page for every number in the financial summary, every add-back and every concentration figure.
Confirm unit and scale on each. Thousands read as millions is an easy miss.
Read the mismatch list and decide which items become questions for the banker or sponsor.
Look for what the CIM leaves out. Risk factors are sometimes omitted entirely (Mergers & Inquisitions).
Write the merits, risks and recommendation yourself.
Save the ledger, bridge and prompts alongside the memo as the record of where each number came from.

For more on which judgments to keep away from the model, see where AI isn't reliable in underwriting.

Where this goes wrong

Scanned CIMs are the first problem. Anthropic notes that PDF reading is subject to the same limits as image reading, and scans without a text layer can't be cited through its API (Anthropic). The same page warns that dense PDFs with small fonts and complex tables can fill the context window before the page limit. Split the file.

Models also fill gaps. In the FailSafeQA benchmark (2025), the model that best resisted bad inputs still "fabricated information in 41%" of cases where context was degraded or missing (arXiv). A missing page or a garbled table invites a plausible number. The NOT FOUND and UNREADABLE rules give the model a way out, and the page check catches what slips through.

Answers vary between runs, too. Anthropic suggests generating more than one response and comparing them to spot inconsistencies (Anthropic). Running the extraction twice and diffing the ledgers is cheap insurance on the financial summary.

Finally, a CIM is a seller's document. Anthropic warns that external files can carry instructions that steer the model (Claude Help Center). Treat the CIM as source material and keep your instructions in the Project.

Making it the team's process

A good prompt helps one analyst. The bigger gain comes when everyone on the team screens deals from the same template and runs the same checks. A private credit manager we worked with had us build their deal screening and diligence request workflow into the firm's own AI account. The team was testing the first working version within three days and sent structured feedback. Associates proposed their own additions, which we folded into the shared workflow. Every output is still a draft that a person reviews.

If you're weighing tools, our comparison of Claude, ChatGPT and Copilot for investment firms covers the differences, and why AI seats don't turn into workflows covers what usually stalls adoption. When you want this set up for your team, see our AI investment memo workflow implementation, including the pilot scope, handoff, and review checks.

How to draft a deal screening memo from a CIM with AI