CIM analysis
AI CIM Analysis: What M&A Teams Should Look For
What an AI first pass on a CIM can genuinely do, the ways it fails, the tests to run before trusting one, and where human judgement stays in the process.
Key takeaways
- An AI first pass earns its place by doing the transcription and reconciliation — extraction, structuring, scoring against stated criteria and listing the gaps — not by writing a fluent summary.
- The failures that matter are invisible in a summary: the wrong year, an adjusted figure presented as reported, a projection as history, an invented number, a valuation the model made up.
- Six tests separate a tool you can act on from one you have to re-read: provenance, "not stated" behaviour, a deterministic valuation, mandate scoring that does not penalise unknowns, the confidentiality terms, and what happens next.
- Judgement on the reason for sale, the management team, the price worth paying and whether to spend the firm's time stays with people.
What an AI first pass can genuinely do
Most of the hours in a manual first pass on a memorandum are spent finding, transcribing and reconciling rather than thinking: locating the financial tables, spreading three to five years into a screening sheet, tracing each add-back, and writing down what the document does not say. How to analyse a CIM describes the method; the point of automating its first step is to spend the judgement on every opportunity rather than on the ones that arrived in a quiet week.
Four things a language model, properly harnessed, does well on a memorandum:
- Extraction. Reading a long, inconsistently formatted document and pulling each figure — revenue, gross profit, EBITDA, each add-back, recurring revenue share, customer concentration, headcount, the asking price — into a structured record, year by year.
- Structuring. Separating what the document reports from what it adjusts and what it projects, so that the three never sit in one column.
- Scoring against stated criteria. Measuring the extracted facts against a mandate you have written down as numbers — sector, size, earnings, geography, margin — and saying which criteria the document answers and which it does not.
- Listing the gaps. Producing the "not stated" list: the years, metrics and evidence a business of this kind would normally disclose and this memorandum does not.
Each of these is a clerical task done at scale and without fatigue, which is exactly the kind of work that gets done badly by hand at five o'clock on the day a third memorandum arrives. None of them is a decision.
The failure modes
A fluent summary with a wrong number in it is worse than no summary, because it gets read and the memorandum does not. The failures to design against share one property: none of them is visible in the output unless the tool shows where each figure came from.
| Failure | What it looks like in the output | What catches it |
|---|---|---|
| A figure from the wrong year or table | A plausible revenue figure that belongs to the prior year, or to one segment rather than the group | A link from the figure to the page and table it was read from |
| Adjusted presented as reported | One EBITDA line, with the add-backs absorbed into it | Reported and adjusted extracted separately, with the bridge listed |
| A projection presented as history | A five-year table with the same formatting in every column | Each period labelled as historical, management-stated or forecast |
| An invented figure | A customer concentration percentage in a document that never states one | "Not stated" wherever the document is silent, and never a plausible default |
| A model-generated valuation | An "estimated value" with no method, no inputs and no assumptions | A valuation computed by a stated method from the extracted inputs, which the model cannot alter |
The last two are the ones a general-purpose assistant is built to commit, because a model asked for a figure will produce one. A tool built for this work has to be built to refuse.
The tests to run
Run these on a memorandum you have already analysed by hand, so that you know the answers. Compare the tool's first pass with yours figure by figure, and classify every difference: your error, its error, or a genuine ambiguity in the document. Best CIM analysis software sets the same tests out as an evaluation framework; here they are in the order to run them.
1. Provenance
Pick any extracted figure and ask to see where it came from. A tool that can show you the page and the table can be verified in seconds; one that cannot has to be re-read, which is the cost it was meant to remove. Provenance should extend to the findings: a red flag should name the passage it was read from.
2. "Not stated" behaviour
Find something the memorandum genuinely does not state — the top customer's share of revenue is the usual candidate — and look at what the tool reports. The right answer is "not stated", listed as missing information. A percentage, a range or an "industry-typical" figure is an invention, however reasonable it looks, and it tells you the tool will fill other gaps you did not check.
3. A deterministic valuation
If the tool produces a valuation, ask which method computed it, on which inputs, with which assumptions, and whether the same inputs produce the same answer twice. A valuation is only useful in a first pass if it reconciles with the full valuation you will run later, which means it has to come from the same engine. A figure the model wrote is a guess with a currency symbol.
4. Scoring against your mandate, with unknowns not penalised
The score should be against your criteria — the acquisition mandate you hold as numbers — not a generic quality grade. Then test the unknowns: a memorandum that omits EBITDA should be flagged for the gap, not scored as though the figure were bad. A tool that treats silence as a miss will rank a thin document on a good business below a thick one on a poor business.
5. The confidentiality terms
A memorandum is a confidential document received under a non-disclosure agreement. Before uploading one anywhere, read where it will be stored, who can see it, whether it is used to train anything, and whether the processing terms for the AI extraction are written down in a data processing agreement you can hold the vendor to. A consumer assistant's terms may not be an appropriate place for a counterparty's memorandum at all. This is a question for your own compliance and legal advisers as much as for the vendor.
6. What happens next
The output of a first pass should be a deal with the figures and findings on it, not a PDF to be retyped. Ask what the tool does with a pursue decision: do the extracted figures travel into the valuation and the returns model, do the findings become diligence questions, and can a second analyst see the first analyst's pass? A first pass that ends in a document has moved the transcription, not removed it.
Where human judgement stays
The decisions the document cannot settle are the ones the tool cannot make either. Whether the reason for sale is credible. Whether the management team can run the business without the founder, which is a judgement about people you have not met yet. Whether the price the numbers support is a price worth paying, given what else is on the desk. Whether the risks the first pass surfaced are ones diligence can bound or ones that are structural. And whether to spend the firm's time on this opportunity rather than the next one.
A good tool makes those judgements better informed and cheaper to reach. It does not make them, and a tool that presents a pursue-or-pass verdict as its own output has confused a first pass with a decision. What it should present is a position with the evidence attached — the three multiples the guide price implies, the mandate criteria met and unknown, the red flags with their sources, the list of what is missing — and leave the verdict to the person who carries it.
How TrueValue answers the tests
The TrueValue CIM Analyzer is built to the six tests above, and it is worth being precise about what that means in practice.
- Provenance. Every figure extracted from a memorandum is stored with the document it came from, and every figure the Agent later writes into a memo or a proposal is checked against the figures actually read — a number with no source in the record is refused rather than repaired.
- Not stated. Extraction records each financial year's revenue, gross profit, EBITDA, add-backs and SDE, plus recurring revenue share, customer concentration, owner involvement and growth rate, along with the company details, the asking price and the reason for sale. A figure the document does not state is recorded as not stated, never estimated, and the missing information is listed as a finding.
- A deterministic valuation. The indicative valuation runs the memorandum's stated figures through the same five-method engine — net asset value, DCF, seller's discretionary earnings, EBITDA multiple and comparable transactions — as a full valuation, and the AI cannot alter a computed figure. Only stated values travel into it; the missing ones are named.
- Mandate scoring. Your acquisition mandate is held as numbers and every memorandum is scored against it. An unknown is never scored as a miss. The screening scorecard is a versioned rule set, so the same document produces the same score twice.
- Confidentiality. Memoranda are stored in your workspace's own private storage, isolated per workspace, with roles controlling who sees a deal and an audit log recording who did what. The processing terms for AI extraction are in the data processing agreement.
- What happens next. Promoting the analysis creates the deal with the company, the figures, the findings and the mandate score already on it. Strengths, red flags and missing information become diligence items; the extracted figures seed the full valuation and the deal economics.
The TrueValue Agent runs the same analysis on any memorandum that lands on a deal, without being asked, and files a proposal with the evidence attached. Every finding names what it was read from; every action it takes leaves a receipt with an undo; and anything that would leave the workspace — an indication of interest, a request list, a message to a seller — waits for a person to click.
What it does not do is as important. It does not read a scanned page that cannot be recognised as text; it reports the page as unread. It does not fill a gap in a memorandum with an industry estimate. And it does not decide whether to pursue.
How to read an AI first pass
Treat the output as the extraction step of the method, done: the record is built, the gaps are listed, the three implied multiples are in front of you. Then do what you would have done with your own screening sheet. Challenge the story against the numbers, using the CIM analysis checklist for the items that matter. Price the business on your sceptical figure, not the seller's. Decide — pursue, pass, or pursue subject to questions — and write the reasons on the deal.
The time you save is the transcription. Spend it on the judgement, and on the memoranda you would otherwise not have opened.
Frequently asked questions
Can I use a general-purpose AI assistant to analyse a CIM?
- It can summarise the document and answer questions about it. It will not score against your mandate, compute a valuation by a stated method, refuse to invent a figure, or turn findings into a deal with diligence items — and its terms may not be an appropriate place to upload a confidential memorandum at all. Check them before you do.
How accurate is AI extraction from a CIM?
- Accurate enough to remove the transcription from a first pass; not accurate enough to skip the check. That is why provenance is the first test: a figure that links to its page is verified in seconds, and one that does not has to be re-read.
Does AI CIM analysis replace the analyst?
- It replaces the hours the analyst spent transcribing and reconciling. Challenging the story, pricing the business and deciding whether to pursue remain the analyst's, and a good tool is designed to make those steps faster to reach rather than to perform them.
What does a deterministic valuation mean?
- That the valuation is computed by a stated method from stated inputs, so the same inputs always give the same answer and the assumptions can be inspected. A figure produced by a language model is neither reproducible nor inspectable, and should not be called a valuation.
Should the AI tell me whether to pursue or pass?
- It should present a position with the evidence — mandate fit, the implied multiples, the red flags and what is missing — and a recommendation you can accept, edit or dismiss. The decision, and the reasons written against it, are yours.
How should I evaluate an AI CIM tool before buying?
- With a memorandum you have already analysed by hand. Run the six tests in this guide, compare the first pass with yours figure by figure, and classify every difference. A tool that fails the provenance or the "not stated" test is a summariser, whatever else it does well.
Put this to work in TrueValue
Related guides
- How to Analyse a CIM: A Complete Guide for M&A Buyers
The first-pass method: extract, challenge, price, and decide — in the order that makes each step useful.
- CIM Analysis Checklist: Every Item to Extract, Verify and Question
The exhaustive list: what to pull out of a memorandum, what to test it against, and what to ask when the document is silent.
- Best M&A Software for Deal Teams: How to Evaluate It Honestly
Not a ranking. The category, the capabilities that matter, a scoring sheet, the questions to ask on a demo, and how to run a one-week evaluation on a real deal.