How to Review a CIM with AI

To review a CIM with AI, have it read every page, pull each figure with its page number, rebuild the banker’s adjusted EBITDA on your own definition, and score the business against your firm’s screening criteria before an associate spends a day on the deal. Kepler, the enterprise intelligence platform we build for investment firms, runs that first pass as a saved playbook on your firm’s definitions, computes every figure in code and cites it to the CIM page, so the deal team starts from a screen it can check in minutes.

A confidential information memorandum is the sell-side banker’s case for the business, sent to bidders after they sign an NDA. It’s written to sell. Every adjusted figure has an argument behind it, and the job of the first review is to find the arguments that won’t survive the data room. This guide covers what to extract from each section, how to rebuild the EBITDA bridge, the revenue and cash tests, red flags, and how to carry CIM claims into second-round diligence. For the full deal cycle, see how private equity firms use AI in due diligence.

What a CIM contains, and what to pull from each section

Section	What it says	What to extract	What to test later
Investment highlights	The banker’s thesis	The five to seven claims the price rests on	Each claim, against data-room evidence
Business and products	What the company sells and how	Revenue by product, pricing model, contract type	Product revenue in the management accounts
Customers and go-to-market	Who buys and why	Top customer share, top-ten share, contract terms	Signed customer contracts
Historical financials	P&L, often three years plus LTM	Revenue, gross margin, reported and adjusted EBITDA	Audited statements and the QoE
EBITDA bridge	The add-backs	Every adjustment, by year, with its stated reason	General ledger, invoices, settlements
KPIs and cohorts	Retention, net revenue retention, cohort charts	The tables behind each headline figure	Customer-level revenue file
Capex and working capital	Investment needs	Maintenance vs growth capex, working capital days	Cash flow statements, fixed-asset register
Management	Team and ownership	Tenure, rollover, key-person dependencies	References and the management presentation
Projections	The management case	Growth, margin path, which add-backs carry forward	History, pipeline, cost plans

Ask the AI for each item with the page it came from. Charts deserve extra care: cohort curves and revenue bridges often appear as images, and a tool that reads only text will skip them or guess.

Step 1: Write down your screening criteria first

The fastest screen compares the CIM against rules the firm already holds. Write them down once so every CIM gets the same test:

• Size: an EBITDA band, such as $10 million to $50 million.
• Fit: the sectors, business models and geographies the fund targets.
• Revenue quality: a minimum recurring share and a ceiling on any single customer.
• Margins and cash: a gross margin floor, a capex ceiling as a share of revenue, a cash conversion target.
• Deal-breakers: for example, more than a quarter of revenue from one customer, heavy reliance on government reimbursement, or an unresolved regulatory inquiry.
• History: whether the firm has seen the company, its competitors or its management before, and why it passed.

Kept in a shared document, these criteria get applied unevenly. Saved as a playbook in Kepler, they run identically on every CIM, and the screen shows which criteria passed, which failed and the page behind each result.

Step 2: Rebuild the adjusted EBITDA bridge

The asking price is a multiple of adjusted EBITDA, so the bridge from reported to adjusted EBITDA is where most of the value argument lives. Sort every add-back into a type:

• Normalizations: owner compensation reset to market, related-party rent.
• Non-recurring items: a litigation settlement, transaction fees, a one-off system migration.
• Recurring items labeled one-time: severance, recruiting or “restructuring” that shows up every year.
• Pro forma items: cost savings not yet achieved, a full year of an acquisition closed mid-year.

Pro forma add-backs carry the most risk. In a 2020 review of M&A and LBO deals originated from 2015 to 2019, S&P Global Ratings found that add-backs inflated marketing EBITDA by an average of more than 28%, and that reported EBITDA in 2018 came in 27% below the marketing figure.

Here is a worked example with made-up numbers for a target’s latest fiscal year:

Line	Amount ($M)	Type	Your firm’s view
Reported EBITDA	18.4		Accept
Owner compensation normalization	0.6	Normalization	Accept
Litigation settlement	1.1	Non-recurring	Accept, after checking prior years
Severance	0.9	Labeled one-time, recurred in two prior years	Reject
Pro forma cost savings	2.4	Not yet achieved	Reject
Acquisition, full-year contribution	1.6	Pro forma, audited	Accept
Adjusted EBITDA	25.0 (banker)		21.7 (your definition)

The gap is $3.3 million, or 13% of the banker’s figure. At an 11.0x multiple, that’s $36.3 million of purchase price. An AI review should produce both columns, cite each add-back to its CIM page, and check whether each “one-time” item appears in earlier years of the same CIM. On Kepler, both versions are computed in code from cited inputs, and your firm’s definition of adjusted EBITDA lives in the ontology, so it is the same definition the deal model and the IC memo use.

Step 3: Test revenue quality

Concentration. Compute the largest customer’s share and the top-ten share for each year shown. A top customer at 14% is manageable; at 14% with a contract renewing next year, it’s a diligence priority.

Retention. Headline net revenue retention is the banker’s chosen number. Rebuild it from the cohort table if the CIM has one: for each annual cohort, compare revenue in year two with year one. In a made-up example, a CIM states 112% net revenue retention, but the cohort table shows that one large customer’s expansion accounts for most of it; excluding that customer, the remaining cohorts compute to 104%. Both figures are true. The second one tells you more about the business, which is why Kepler computes retention from the cohort table itself and shows the largest customer’s contribution as its own line.

Pricing and volume. Split growth into price, volume and new customers where the CIM allows. Growth that comes mostly from price increases behaves differently in a downturn.

Recurring vs repeat. “Recurring” sometimes means contracted and sometimes means customers who usually come back. Check which one the CIM means, page by page.

Step 4: Capex, working capital and cash conversion

EBITDA doesn’t pay the debt; cash does. Pull maintenance and growth capex separately, and watch for capitalized software or development costs rising faster than revenue, which moves expense out of EBITDA. Compute cash conversion as (EBITDA minus capex) divided by EBITDA for each year. Look at working capital by month if the CIM shows it, since seasonal businesses need a peg that reflects the trough as well as the average. Deferred revenue and customer deposits flatter cash in a growing business and reverse when growth slows.

Step 5: Management and the projections

List who runs each function, how long they’ve been there, and who is rolling equity or leaving at close. Then read the projections as a set of assumptions. Ask the AI to list every one with its page: revenue growth by year, margin expansion, which add-backs carry into the forecast, and capex. Projections built on the pro forma savings you rejected in Step 2 should be rebuilt on your number.

Red flags checklist

• “One-time” costs that appear in more than one year
• Add-backs for savings the company hasn’t achieved
• A top customer whose contract renews within the hold period
• Cohort charts that start at a convenient year, or skip one
• KPI definitions that change between the summary and the detail pages
• Summary figures that don’t tie to the financial pages
• A recent change in fiscal year, revenue recognition or segment reporting
• Capitalized costs growing faster than revenue
• Margin expansion in the projections that depends on the add-backs
• LTM figures without the quarters that build them
• A founder or key seller leaving at close

Where generic AI goes wrong on a CIM

A chat assistant gives a quick, readable summary of a CIM, and that’s useful. The failures show up in the numbers. A language model writes figures as text, so it can round a margin, mix fiscal and calendar years, or report the banker’s adjusted EBITDA as if it were the company’s. It reads what it retrieves, which can mean skipping appendix tables and image-based charts. It doesn’t know your screening criteria or your firm’s history with the company unless someone pastes them in every time. And its first-round work doesn’t carry into the data room, so second-round diligence starts again from the PDF.

Several products target this step. V7 Go’s private equity agents extract, validate and structure data across CIMs, data rooms and portfolio reports with citations to the source page, and a September 2026 Affinity integration turns an inbound CIM into a CRM deal record. Hebbia’s investing page offers a first pass on CIMs and VDRs. Datasite acquired Blueflame AI in July 2025 and runs it inside Datasite rooms, and Keye turns raw deal files into structured outputs such as cohort trends, with Excel exports. Claude and ChatGPT will summarize an uploaded CIM and answer questions about it. For a ranked comparison, see the best AI tools for private equity due diligence.

Testing CIM claims against the data room

The CIM’s claims become second-round diligence tests. Map each one to the evidence that would confirm it:

CIM claim	Evidence in the data room	Test
92% of revenue is recurring	Customer-level billing data, contracts	Recompute recurring share from billing
Top customer under contract through 2029	The signed master agreement and amendments	Read the term, renewal and termination clauses
Litigation add-back of $1.1M	Settlement agreement, general ledger	Confirm the amount, period and that it’s closed
Net revenue retention of 112%	Customer revenue file by month	Rebuild cohorts with and without the largest customer
Maintenance capex of 2% of revenue	Fixed-asset register, capex by project	Re-split capex by project

Mark each claim confirmed, revised or unsupported. Revised figures flow into the model and the memo with their new sources. Our guide on how to check AI-generated numbers covers the citation checks in more detail.

How Kepler does this

A CIM rarely arrives cold. Most firms have met the company, a competitor or the management team before, and that history sits in old memos and CRM notes nobody has time to search. Kepler starts there: our engineers build your firm’s ontology from your past CIMs, IC memos, pass notes, models and CRM entries, so every company, executive, adviser and deal resolves to one record, and a new CIM lands next to everything the firm already knows about it.

Stage	What Kepler does	What the deal team gets
The CIM lands	Extraction models we train read every page at ingest, appendix tables and image-based charts included, and file each figure with its page, period and units	Every figure in the CIM, searchable, before anyone opens the PDF
First screen	Runs your criteria as a saved playbook and computes both EBITDA bridges, concentration, cohort retention and cash conversion in code	A pass or fail on each criterion in your screening memo format, every figure opening to its formula and CIM page
Conflicts	Compares the summary pages with the financial pages	Both values side by side, with the conflict flagged
Round two	Links each CIM claim to the data-room documents that test it and re-reads every upload overnight	A claims tracker marked confirmed, revised or unsupported, with the exact quote
Into IC	Carries revised figures into the model in Excel and the IC memo in Word	One definition of adjusted EBITDA from first screen to committee

A Kepler agent working a data room has exactly the access of the associate who started it, and a sentence inside a seller’s document is read as content, never followed as an instruction. Deal walls apply to people and agents alike, and every query is logged (security). Kepler delivers the same answers in PowerPoint, Claude, ChatGPT and Slack. The engineers who set this up with your deal team built this layer before at Palantir, and public company data for comps and precedents is ready on day one.

Three questions to run on your next CIM:

• “Screen this CIM against our criteria and show which ones fail, with the page for each.”
• “Rebuild adjusted EBITDA on our definition and list every add-back we’d reject, with its page.”
• “Have we looked at this company or its two closest competitors before? Show why we passed.”

Start at app.kepler.ai, or book a demo and watch it screen a CIM from your own pipeline against your firm’s criteria. For how the pieces fit together, see the Kepler platform.

Frequently asked questions

Can AI review a CIM?

Yes, for the first pass. Kepler, which we build, reads every page of the CIM, rebuilds the banker's adjusted EBITDA on your firm's definition, measures customer concentration and retention, and scores the business against your screening criteria as a saved playbook, with every figure computed in code and cited to its CIM page. Every CIM gets the same review, and the deal team keeps the judgment.

What should you look at first in a CIM?

Start with the adjusted EBITDA bridge, because the asking price is a multiple of it. Then check revenue quality (recurring share, customer concentration, retention by cohort), the margin trend, capex and working capital needs, and who in management stays after the deal. Read the investment highlights last, as a list of claims to test, since the banker wrote them to sell the business.

What are common red flags in a CIM?

One-time costs that recur every year, add-backs for savings not yet achieved, a large customer whose contract renews soon, cohort charts that start at a convenient year, KPI definitions that shift between sections, summary figures that don't tie to the financial pages, and projections whose margin gains depend on the add-backs. Any one is a question for the banker. Several together usually change the price.

Is it safe to put a CIM into an AI tool?

Check the NDA first: many restrict who can see the CIM and require the information to stay within the firm and its advisers. Use an enterprise deployment where your data isn't used to train shared models, access mirrors your deal team, and every query is logged. On Kepler, each firm's data is isolated, never trains models other firms use, and deal walls apply to people and agents alike.

How do you test CIM claims in the data room?

Turn each claim into a test with a named source. Recurring revenue share gets checked against customer-level billing data, contract terms against the signed agreements, add-backs against the general ledger and settlement documents, and retention against the customer revenue file. Track each claim as confirmed, revised or unsupported, and carry the revised figures into the model and the IC memo.

What is the difference between a teaser, a CIM and a management presentation?

The teaser is a short anonymous synopsis sent to gauge interest. The CIM follows the NDA and lays out the business, its financials and the banker's investment case in detail. The management presentation comes later, for bidders that make the next round, and is where the deal team questions management directly. Each should be reviewed against the one before it.