DEV Community: DEVALAND

Best Due Diligence Software for Search Funds (and the VDR Alternative)

DEVALAND — Thu, 09 Jul 2026 14:00:04 +0000

For a search fund, the best due diligence software is not a virtual data room. It is the tool that turns a CIM and a folder of financials into a brief you can defend, in minutes, where every figure is tied to its source page. A data room stores documents. It does not read them, check them, or tell you when two of them disagree. If you are buying one company without an analyst bench, that difference is the whole job.

Why the obvious answer, a VDR, is the wrong one for a small deal

Search the term and you land on the big virtual data room names: Datasite, Ansarada, iDeals. They are real, capable products. They are also built for the sell-side and for large, multi-party processes, priced and shaped for a corporate development team or a bank running a competitive auction. For a self-funded searcher or an independent sponsor buying one business, three things go wrong:

You pay enterprise pricing, often per page or per project, for what amounts to a secure folder.
The tool organizes documents but leaves the reading, cross-checking, and risk-spotting to you.
The work it does not do, turning ninety pages of CIM into a defensible summary, is exactly the work you do not have an analyst to do.

A VDR answers "where are the documents." A searcher's real question is "what is in them, and which numbers do not add up." Those are different tools.

VDR vs diligence automation vs doing it by hand

	Virtual data room (Datasite, Ansarada)	Diligence automation (Deal OS)	By hand or Claude on its own
What it does	Stores and shares documents securely	Reads the CIM and financials, returns a cited brief, flags contradictions	You read everything and build the memo
Built for	Sell-side, banks, large auctions	Searchers, independent sponsors, micro-PE	Anyone, at the cost of your hours
The reading and checking	You	The tool, with every claim cited	You
Confident wrong number	Yours to catch	Discarded before you see it, or shown with its source	High, an uncited summary hides its errors
Fits a one-deal buyer	Overbuilt and overpriced	Yes, self-serve	Free, but slow and easy to miss things

What a search fund actually needs from diligence software

Strip it back to the job and the checklist is short:

It reads the documents, not just stores them. The first-pass read of a CIM and a P&L is the time sink, and it is automatable.
Every claim is cited. A figure in the brief opens to the exact line in the source, or it does not get used. That is the difference between a tool you can stand behind at the IC and a chatbot that sounds sure of itself.
It catches contradictions. The seller's number and the data room's number quietly disagree, and a search fund without an analyst is the most likely to miss it.
It is self-serve and affordable. No implementation project, no per-page meter, no sales call before your first deal. Less than an analyst-day a month is the right order of magnitude, not enterprise VDR pricing.
Your documents stay isolated and encrypted, not pasted into a consumer chat.

That is the wedge: not feature parity with a data room, but verification. A checklist tells you what to check. It does not tell you whether the seller's answer is true. The number that is wrong is the one that costs you, and it stays invisible until something checks it against the source.

An honest, affordable alternative for small deals

This is the gap Deal OS is built for. You upload a CIM, and you read a cited diligence brief in minutes: pipeline, data room, and AI briefs in one place, where every claim carries a verbatim quote checked against the page it came from, and anything it cannot verify is discarded before you see it. It is self-serve, priced for a one-deal buyer rather than an auction, and the judgment stays with you. It is not a VDR with more storage. It is the reading and checking a search fund cannot hire out.

You can see exactly what that looks like on a synthetic deal, no login, in the sample brief, where verified claims show their source and a contradiction shows both sides.

Frequently asked questions

What is the best due diligence software for a search fund? The one that reads the documents and cites every claim, not the one that only stores them. A searcher without an analyst needs the CIM turned into a defensible brief, which is diligence automation, not a data room.

What is a good Datasite alternative for small deals? For a one-company acquisition, the alternative is usually not another VDR but a diligence tool that actually reads and verifies the documents, self-serve and priced for a single deal rather than an enterprise auction.

VDR vs diligence automation, what is the difference? A VDR is secure storage and sharing. Diligence automation reads the stored documents, returns a cited summary, and flags contradictions. One answers where the documents are, the other answers what is in them.

Is there an affordable due diligence tool for a search fund? Yes. Self-serve diligence automation runs at roughly an analyst-day a month, far below enterprise VDR pricing, and does the first-pass reading a small team cannot staff.

What about Ansarada, DealRoom, or Intralinks for a search fund? Those are virtual data rooms built for sell-side auctions and larger processes, often priced per page or per project. For a single-company acquisition the problem is not storage, it is analysis. A searcher without an analyst needs the documents read, cited, and checked for contradictions, which is a different tool than any of them. Compare on analysis, not on storage. See AI CIM analysis tools compared (2026).

See source-cited diligence on a real-looking deal at Deal OS, and compare the field in best due diligence software (2026).

AI Due Diligence vs Hiring an Analyst: Cost, Speed & Accuracy

DEVALAND — Mon, 06 Jul 2026 14:00:06 +0000

If you're buying a business without a ten-person deal team, the first read of a CIM or data room is the bottleneck. You have three options: read it yourself, hire an analyst, or outsource a Quality of Earnings engagement. AI due diligence is now a fourth. Here is how they actually compare.

Cost (first read of a CIM / small data room):

Hire a junior M&A analyst: $80,000-150,000/year loaded
Outsourced FDD / QoE engagement: $15,000-50,000 per deal, 2-3 weeks
Do it solo: "free," but it's the week you don't have
AI due diligence (Deal OS): $750-4,500/month, unlimited deals

Speed (one CIM):

Analyst read: 3-5 business days
AI cited brief: minutes

Accuracy on the part that matters:

Human read: strong judgment, but fatigue misses contradictions buried across 60+ pages
AI cited brief: every figure traced to its source page, anything unverifiable discarded before you see it

This guide compares all four across cost, speed, accuracy and coverage, so you avoid both the expensive mistake (paying analyst rates for a first-pass read) and the dangerous one (trusting an AI summary you can't check).

The Manual Diligence Landscape

Reading it yourself is what most solo searchers and independent sponsors do, and the cost is hidden because it's your time. A serious first read of a single CIM is a full day; a small data room is a week. The real risk is not the hours, it's what a tired reader misses on page 22 that quietly contradicts page 8. Customer concentration, a loosened "recurring revenue" definition, a working-capital quirk: the landmines hide in the detail, not the headline EBITDA.

Hiring an analyst buys you leverage but at a fixed, ongoing cost. A junior M&A or FDD analyst runs $80,000-150,000/year fully loaded, and they still read one document at a time. For a searcher doing two or three live looks a quarter, that's an expensive way to get a first-pass read, and you carry the cost whether you have a deal in front of you or not.

Outsourcing FDD or a Quality of Earnings report is the right tool late in a deal, on a target you're serious about. Independent QoE providers typically deliver in 2-3 weeks at a fixed fee in the $15,000-50,000 range for deals between $500K and $20M. It's rigorous and worth it before you wire money. But it's far too slow and expensive to run on every teaser and CIM that crosses your desk, so most opportunities never get a rigorous read at all.

The gap in all three: the first-pass read, the screen that decides whether a deal is worth an analyst's week or a QoE fee, is either skipped, rushed, or done at a cost that doesn't scale.

AI Due Diligence: The Modern Alternative

AI due diligence targets exactly that first-pass read. You upload a CIM or a data room, and minutes later you have a screening memo: the figures that matter, the risks, the questions worth putting to management, and, critically, every claim linked back to the exact page it came from.

The reason it's safe to act on is one rule that separates it from pasting a CIM into a chatbot: every claim an agent makes is cited to its source, or it gets cut. A confident summary that invents an EBITDA figure or softens a customer-concentration risk is worse than no summary, because you'll act on it. So anything the model cannot trace to a verbatim source passage is discarded before it reaches you. What's left is checkable.

It also reads the whole data room at once, which a human can't hold in working memory. That's where it earns its keep: catching the contradiction between two documents that no single-document read would surface.

Head-to-Head: The Numbers

Cost on a single deal's first read. An analyst's time on one CIM costs roughly $1,500-3,000 in loaded salary; an outsourced read isn't available at that scope. AI due diligence runs the same read for a few dollars of compute inside a $750-4,500/month subscription that covers unlimited deals. For anyone screening more than one or two opportunities a month, the per-deal economics aren't close.

Speed. A human analyst returns a CIM read in 3-5 business days. An AI cited brief comes back in minutes. The point isn't to replace the analyst's judgment, it's that you can now screen ten teasers in the time it used to take to screen one, and only spend human hours on the deals that survive.

Accuracy and coverage. This is the honest trade. A human brings judgment AI doesn't have. But a human reading 60-plus pages under time pressure misses things, and they read one document at a time. AI brings consistency and full-data-room coverage, and the citation rule means you're never trusting an unverifiable number. The strongest setup uses both: AI does the exhaustive, cited first pass; the human spends their judgment on what it surfaces.

A Real (Synthetic) Example

On a synthetic deal we publish openly, the brief flags a contradiction a fast human read routinely misses: page 8 of the CIM states "no customer represents more than 15% of revenue," while page 9's own figures show one account (Northgate, $1,366,000) at 22.0% of $6.21M revenue. Two pages apart, in the same document. It also catches a "78% recurring revenue" claim on page 7 against a quietly loosened definition on page 22, and logs the claims it discarded because they couldn't be verified. You can read the full cited brief, no login, at https://os.devaland.com/sample-brief.

That single catch, customer concentration, is the kind of thing that re-rates a multiple or kills a deal. Surfacing it in the first-pass read, not after a QoE fee, is the entire value.

When Each Option Makes Sense

Do it yourself: one deal at a time, you have the week, and you trust your own read.
AI due diligence: you're screening real volume, want a cited first-pass read on every CIM in minutes, and you don't have a deal team. This is the searcher / independent sponsor / micro-PE sweet spot.
Hire an analyst: you have steady deal flow and want a dedicated person, ideally pointed at what the AI surfaces rather than the raw read.
Outsourced QoE: late-stage, on a target you're committed to, before money moves.

These aren't mutually exclusive. The point is to stop paying analyst or QoE prices for the first-pass screen, and stop skipping it because it's too slow.

Frequently Asked Questions

Does AI due diligence replace a Quality of Earnings report?
No. QoE is a late-stage, rigorous engagement before you close. AI due diligence replaces the expensive, slow, or skipped first-pass read that decides whether a deal even deserves a QoE. Use AI to screen everything, QoE on the one you're closing.

How is this different from pasting a CIM into ChatGPT?
A general chatbot will summarize confidently whether or not the summary is true. The difference is verification: every claim is traced to its exact source page, and anything that can't be verified is discarded before you see it. The summary is the easy part; the discard-if-unverifiable layer is what makes the output safe to act on.

Is it accurate enough to make decisions on?
You don't take its word for anything, that's the design. Every figure is a clickable citation back to the source document. You verify the ones that matter in seconds instead of re-reading 60 pages to find them.

What does AI due diligence cost?
Deal OS plans run $750-4,500/month for unlimited deals, versus $80,000-150,000/year for an analyst or $15,000-50,000 per outsourced QoE engagement. For anyone screening more than one deal a month, the per-deal cost is a rounding error.

Who is it built for?
Searchers, independent sponsors, small funds and micro-PE buyers, anyone carrying a deal without a ten-person diligence team behind them.

What Due Diligence Automation Actually Catches (And What It Can't)

DEVALAND — Thu, 02 Jul 2026 14:00:04 +0000

The honest version, from someone who builds the tooling.

Every few weeks now, a searcher tells me some version of the same story. They ran a CIM through an AI tool, got back a clean, confident summary, and then, because they're careful, went back to the source documents anyway. And the summary was wrong. Not catastrophically, not obviously. Just wrong in the quiet way that matters: a number smoothed over, a risk softened into a strength, a claim stated with more certainty than the document actually supported.

Here's the uncomfortable part. The person caught it because they didn't trust the tool. Which raises the question nobody selling AI diligence wants you to ask: if you have to re-read the document anyway, what did the tool actually save you?

That question is worth sitting with, because the answer separates due diligence automation that helps from the kind that quietly hurts. So let me be specific about what this category of tooling genuinely does well, where it earns its place in a deal process, and just as importantly, where it does not belong, and where trusting it is actively dangerous.

Quick answer: Automated due diligence uses software (increasingly AI) to run the repeatable, document-heavy first pass of a deal, reading the CIM, the financials, and the data room, reconciling figures, and flagging risks and missing items. The automated due diligence process accelerates the read; it does not replace your judgment, your QoE provider, or your advisors. The one rule that keeps it safe: every claim must trace to its source document, or it should be discarded. (For a like-for-like tool comparison, see the best due diligence software.)

What it's actually good at

The repeatable, document-heavy, attention-draining first pass. That's the real job.

What automation catches vs. what still needs you

Automation does well	Still needs a human
Reading a full data room fast	Judging whether the business is worth buying
Reconciling figures across documents	Negotiating price and structure
Flagging contradictions and missing documents	Reading the owner and the market
A consistent first pass on every deal	The final QoE and legal sign-off
Tracing each claim to its source page	Deciding what risk you can live with

A first-pass review of a data room, reading the CIM, tying it to the financial statements, normalizing how the seller chose to present their numbers, building a list of what to ask, is days of senior time per target. Most of it is spent on deals that die. For a two-person search fund or a solo independent sponsor reviewing the same volume of documents a mid-market PE firm sees, with none of the analyst bench, that first pass is the bottleneck. Not judgment. Reading.

This is where automation earns its keep. Done properly, it compresses that week-long first read into hours: a structured summary against your screening criteria, a ranked list of red flags, a draft question pack for the management call. The deals worth a real look rise to the top faster. The weak ones die earlier, before you've sunk a weekend into them. Your review throughput goes up without adding headcount.

The second thing it does genuinely well is something humans are bad at, not because they're careless but because of how memory works: cross-referencing every figure across every document at once. A person reading a CIM on page 4 does not reliably remember the slightly different version of the same number on page 40 of the tax return, or the third version the seller said on the management call. Machines don't get tired on page 40. Pointed at the CIM, the financials, the tax returns, the contracts, and your call notes together, the tool can surface where they disagree, the add-back that doesn't reconcile, the growth claim the statements don't support, the customer concentration the narrative glossed. These are the things that surface post-LOI and cost money and credibility. Catching them before you commit is the highest-value thing automation does on a deal.

So: first-pass speed, and contradiction-spotting across sources. Those are real. If a tool does only those two things well, it's worth having.

Where it quietly hurts you

Now the part most vendors skip.

The failure mode of AI in diligence isn't that it can't read. It's that it reads confidently and you can't tell when it's wrong. A fluent, well-structured summary is psychologically harder to distrust than a messy one, it feels authoritative, which means a confident summary you can't verify is worse than no summary at all. It doesn't just fail to help; it gives you false comfort on the exact thing that kills deals. You relax precisely where you should have looked harder.

I'll give you the example that made me build the way I did. An AI tool described a target's revenue as "well-diversified across its customer base." It read clean. It wasn't true, one customer was a large share of revenue, and the number was sitting right there in the CIM. The model didn't lie, exactly. It smoothed. It took an ambiguous picture and resolved it in the optimistic direction, because that's what fluent summarization does: it produces the most plausible-sounding sentence, not the most accurate one.

That's the trap. And no "better model" fixes it, because the problem isn't capability, it's that a summary, by definition, throws away the evidence and asks you to trust the conclusion. In a domain where being wrong on one number blows a seven-figure decision, "trust the conclusion" is not an acceptable interface.

The line that actually matters

The right place for tooling in diligence is not "tell me what to think." It's "show me the evidence faster, and let me judge it." An ETA investor I was trading notes with put it better than I usually do: the tool should make it harder to miss the line that matters.

That's the whole design principle, and it's a narrow one. Every claim a tool puts in front of you should carry the verbatim quote it came from, checked against the source page. If a claim can't be tied back to the document, it shouldn't be allowed to appear, it should be discarded before you ever see it. Not flagged with a confidence score. Discarded. Because a claim you can't verify is exactly the claim that smooths a 40% customer into "well-diversified."

This matters even more in European SME deals, where "messy" is not automatically bad. Sometimes a process that looks like a mess is bad bookkeeping. Sometimes it's twenty years of one owner's exceptions, customer intimacy, and supplier trust that actually make the business work. A spreadsheet, and a confident AI summary, treats both the same. The interpretation has to stay human, because the same line in a CIM means different things depending on the owner, the local market, and how the business has actually been run. The tool's job is not to do that interpretation. It's to surface the line, with its source, so the human with the context can do the interpreting properly and not miss it.

What it doesn't replace, and shouldn't pretend to

Worth saying plainly, because the overclaiming in this category is exactly what makes careful buyers distrust all of it:

Automation does not replace your QoE provider, your counsel, or your accountant. It does the preliminary, repeatable collation, organizing the data room, surfacing the contradictions, drafting the question lists, so that when your advisors engage, their expensive hours go to judgment and confirmatory work rather than assembling the picture. Confirmatory diligence stays human. The advisors stay essential. The tool just makes their hours count by arriving prepared.

And it does not make the decision. It can show you that the add-back doesn't reconcile. Whether that's a dealbreaker or a negotiating point is yours.

So, does it save you anything?

Back to the question we started with. If a tool gives you a summary you have to re-verify line by line, it saved you nothing, it added a step.

But if it gives you every claim already tied to its source, so verification is a click instead of a re-read; if it catches the contradiction across page 4 and page 40 that you'd never hold in your head at once; if it kills the weak deals early so your real attention goes to the live ones, then it saved you the week, and it didn't cost you the false comfort. That's the version worth using. The difference between the two isn't the model. It's whether the tool shows its work.

That's the only standard that matters for AI in diligence: not how smart the summary sounds, but whether you can check it. Demand to see the citation. If a tool can't show you the line, don't trust what it says about the line.

You can see what "shows its work" looks like in practice, every claim tied to its source page, unverifiable claims discarded before you see them, in a cited brief on a synthetic deal.

Anthropic Paused Its Most Capable Model, Then Made the Case for Verifying AI

DEVALAND — Wed, 01 Jul 2026 20:38:27 +0000

Anthropic just restored access to Fable 5, its most capable model, after a two-week pause triggered by a reported safeguard bypass. The company retrained its guardrails, added layers of defense so no single failure is fatal, and proposed an industry standard for scoring how dangerous a given bypass really is. Strip out the cybersecurity specifics and the lesson is general, and it lands hardest wherever AI touches a decision with money on it: as models get more capable, the value stops being in what they can generate and moves to whether you can verify what they produced.

What actually happened

Anthropic released Fable 5 and Mythos 5 on June 9. The two share an underlying model, but Fable 5 shipped with the strongest safeguards the company had ever applied, while Mythos 5, with fewer guardrails, went only to a small set of trusted partners for defensive cybersecurity work. On June 12, the US government applied export controls after learning of a report from Amazon researchers who had found a way to prompt Fable 5 into identifying software vulnerabilities, and in one case producing code that showed how a vulnerability could be exploited. Because the order took effect immediately and nationality could not be verified in real time, Anthropic suspended access to both models for everyone. On June 30 the controls were lifted, and Fable 5 returned globally on July 1. The full account is in Anthropic's own post.

Two details matter more than the headline. First, when Anthropic tested the reported technique, less capable models, including its own Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities, and every model tested could reproduce the single exploit demonstration. The bypass was a borderline case, not a unique superpower. Second, the fix was not a smarter model. It was a stronger verification layer: a retrained classifier that now blocks the specific technique in over 99 percent of cases, with blocked requests rerouted to Opus 4.8.

The pattern: capability is getting cheap, verification is the moat

The instinct is to read a story like this as being about cybersecurity, or politics, or one company. It is really about a shift that touches every serious use of AI. When a frontier lab, a weaker open model, and a competitor can all reach the same capability, the capability itself stops being the differentiator. What separates a tool you can rely on from one you cannot is whether its output is checkable. Anthropic did not respond to the incident by making Fable 5 less capable. It responded by making its outputs easier to verify and harder to misuse. That is the whole move.

Defense in depth, and why it is not just an AI-lab idea

Anthropic describes its safety approach as defense in depth: no single mechanism is trusted to be perfect, so several imperfect ones are layered until the system as a whole is very hard to misuse. Classifiers watch for dangerous requests. A deliberate safety margin errs toward caution, blocking some benign requests rather than risk missing a harmful one. Humans set the policy; the system enforces it.

Anyone who has run real due diligence will recognize this, because good diligence has always worked the same way. You do not trust a single number because it appears in a polished CIM. You check it against the financials, then against the tax return. You treat a claim you cannot source as a question, not a fact. The table below maps the lab's safety principles onto the diligence equivalents.

Anthropic's safeguard idea	The diligence equivalent
Defense in depth, no single control trusted alone	Cross-document tie-out: the same figure checked across the CIM, the financials, and the tax return
Classifiers that block unverifiable outputs	Cite or cut: a claim with no source becomes a question, never a stated fact
A safety margin that errs toward caution	A visible discard log: anything the tool cannot stand behind is surfaced, not smoothed over
Humans set policy, the system enforces it	The tool reads and verifies; the acquirer decides

A shared way to score how bad a jailbreak is

The most forward-looking part of Anthropic's announcement is a proposed industry framework, drafted with Amazon, Microsoft, Google, and other partners, for scoring the severity of an AI jailbreak. Today there is no common standard, so every new bypass creates uncertainty about how urgently to act. The proposal scores a jailbreak on four questions.

Criterion	The question it asks
Capability gain	How far beyond existing tools does it take the user?
Breadth of capability gain	For how many different attacks does the same technique work?
Ease of weaponization	How much human effort to turn it into a real attack?
Discoverability	How easy is it for someone to obtain the technique?

A shared vocabulary for severity is the same thing diligence needs and rarely has: a way to communicate, consistently, how serious a finding is. A contradiction on page eight is not the same as a rounding difference in a footnote, and treating them the same wastes time or misses risk. Scoring beats vibes.

What this means if you point AI at a deal

The practical takeaway is not "avoid AI." It is "demand the receipts." If you use a model to read a data room, the output is only as trustworthy as your ability to check it. That means insisting on a few things:

Every figure cited to a source document and page, openable and checkable.
A visible discard log: the claims the tool could not verify, shown as questions rather than quietly dropped.
Cross-document tie-out, so the same number is confirmed across the CIM, the financials, and the return.
Your documents kept isolated and encrypted, not pasted into a general consumer chatbot.
The judgment left to you. The tool reads and verifies; you decide.

If a tool cannot show you why to trust a given line, treat its output the way you would treat an analyst who refuses to show their work.

The takeaway

The right amount of trust to place in an AI is exactly as much as you can verify. Anthropic just spent two weeks and doubled a team to prove that principle at the frontier. It applies in miniature on every acquisition: a confident, uncited summary of a data room is a liability dressed as a shortcut, and a summary where every claim opens to its source is a genuine edge.

You can see the cite-or-cut discipline on a synthetic deal, with no login, in the sample brief, where verified claims show their source, a contradiction shows both sides, and unverifiable claims are discarded in front of you. For the longer argument, read Can You Trust AI for Due Diligence, and for the method, how to use Claude for due diligence.

Frequently asked questions

Why did Anthropic pause Fable 5? After the US government applied export controls on June 12, 2026, in response to a report of a safeguard bypass, Anthropic suspended access to Fable 5 and Mythos 5 because it could not verify user nationality in real time. Access was restored on July 1 after the controls were lifted and stronger safeguards were added.

What is an AI jailbreak? A jailbreak is a way of prompting a model so it bypasses its own safeguards and produces output the system was meant to block. Anthropic notes that most jailbreaks are narrow, unblocking one specific behavior rather than a broad class of harmful ones.

Does the Fable 5 incident mean AI is unsafe for due diligence? No. It is a reminder that AI output should be verified rather than trusted on faith. For reading and first-pass analysis, AI is genuinely useful, provided every claim is cited to a source you can check and unverifiable claims are discarded.

What is defense in depth in AI safety? It is the practice of layering several independent safeguards so that no single failure exposes the system. Anthropic uses trained refusals, classifiers, a cautious safety margin, and after-the-fact analysis together, rather than relying on any one of them.

Can you trust AI for high-stakes financial decisions? Only to the extent you can verify its output. Use AI to compress the slow first-pass reading, then require a citation for every figure and treat anything unsourced as a question. Do not act on an uncited summary.

What is Anthropic's jailbreak severity framework? A proposed industry standard, drafted with Amazon, Microsoft, and Google, that scores a jailbreak on capability gain, breadth, ease of weaponization, and discoverability, so developers and governments can judge how urgently to respond.

See verified, source-cited diligence on a real-looking deal at Deal OS.

Can You Trust AI for Due Diligence? An Honest Answer

DEVALAND — Tue, 30 Jun 2026 07:46:45 +0000

Can you trust AI for due diligence? Yes, for reading and first-pass analysis, but only if it shows its work: every figure traced to the document it came from, and anything it cannot verify discarded rather than guessed. An AI summary you cannot check is the one thing you should not act on in a deal.

AI's limits are in the headlines again, and it is a fair question to ask before you point a model at a seven-figure decision. The honest answer is not a flat yes or no. It is "yes, under one condition," and that condition is verification.

Why trust means something different in an acquisition

In most jobs, a wrong AI answer is an annoyance you catch later. In diligence it is the basis you priced a deal on. If a model invents an EBITDA figure, rounds a churn number in the seller's favor, or softens a customer-concentration risk into something that sounds fine, the error does not announce itself. By the time you find it, you may have signed the LOI and spent a month and credibility you cannot get back.

So the real question is not "is the AI smart." It is "can I tell its right answers from its wrong ones." That is a verification problem, not an intelligence problem.

Where AI is trustworthy, and where it fails quietly

Task	Trust level	Why
Reading a long CIM fast	High	It genuinely compresses hours of reading
Restating dense financials plainly	High	Strong at summarizing what is on the page
Extracting a figure with a citation	High, if cited	You can check each number against its source
Catching a document that contradicts itself	High, if asked	It surfaces both passages when told to look
Filling a gap with a plausible number	Dangerous	It will invent an average contract value or a clean litigation history sourced from nothing
An uncited summary taken at face value	Dangerous	You cannot tell a right number from a wrong one

The pattern: AI is trustworthy when its output is checkable, and dangerous when it is not. The failure mode is not a rough paragraph, it is a confident wrong number.

The standard that makes AI output safe to act on: cite or cut

One discipline separates a tool you can trust from a chatbot that sounds sure of itself: every claim traces back to a line in a document, or it does not get used. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim never reaches you; it becomes a question instead.

That is the whole point. You are not trusting the model, you are trusting the citation, which you can open and read yourself. The verification layer, not the summary, is what makes the output safe.

What to demand from any AI diligence tool

Every figure cited to a source document and page, openable and checkable.
A visible discard log: the claims it could not verify, surfaced as questions rather than silently dropped.
Cross-document tie-out: the same number checked across the CIM, the financials, and the tax return.
Your documents kept isolated and encrypted, not fed into a general consumer chat.
The judgment left to you. The tool reads and verifies; you decide.

If a tool cannot show you why to trust a given line, treat its output the way you would treat an analyst who refuses to show their work.

The bigger picture

The current debate about where AI needs limits is healthy, and it applies in miniature on every deal: the right amount of trust in AI is exactly as much as you can verify. Used with that discipline, AI does the slow first-pass reading in minutes and hands you the real questions faster. Used as an oracle you believe on faith, it is a liability dressed as a shortcut.

You can see the cite-or-cut discipline on a synthetic deal, no login, in the sample brief, where verified claims show their source, a contradiction shows both sides, and unverifiable claims are discarded in front of you. For the practical method, see how to use Claude for due diligence and what diligence automation actually catches.

Frequently asked questions

Can you trust AI for due diligence? For reading and first-pass analysis, yes, provided every claim is cited to a source you can check and unverifiable claims are discarded. Do not act on an uncited AI summary.

Does AI hallucinate financial figures? It can, especially when asked to fill a gap. The defense is not a better model but a discipline: require a citation for every figure and treat anything unsourced as a question, not a fact.

Is AI accurate enough to replace a quality of earnings report? No. AI replaces the slow first-pass read that decides whether a deal deserves a QoE. The formal QoE still happens on the deal you are closing.

How do I know an AI diligence tool is trustworthy? It shows its work: each claim traces to a document and page, it lists what it could not verify, and it ties the same figure across documents. If you cannot check the output, do not trust it.

See verified, source-cited diligence on a real-looking deal at Deal OS.

How to Use Claude for Due Diligence (Without Trusting a Word It Says)

DEVALAND — Tue, 30 Jun 2026 07:46:41 +0000

Claude can read a hundred-page CIM in a couple of minutes and hand you a clean summary. In an acquisition, that summary is the one thing you cannot afford to trust on its own. The failure mode is not a rough paragraph; it is a confident, wrong number you cannot tell apart from a right one, sitting in the basis you priced a deal on.

So the goal is not to get Claude to summarize. It is to get Claude to read, and to make every factual claim trace back to the document it came from, or cut the claim before it reaches you. Used that way, it genuinely compresses the slowest part of diligence. Used as a summarizer you trust, it is a liability.

The one rule: cite or cut

Every figure traces back to a line in a document, or it does not get used. That single discipline is what separates a tool you can act on from a chatbot that sounds sure of itself. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim does not reach you at all; it becomes a question.

A practical method

Step	What to do	Why
1. Give it the real documents	Attach the CIM, financials, tax return, and key contracts, not a paste of one page	It can only cite what it can see
2. Ask for citations, not prose	Require every claim to quote the source and page, verbatim	Makes the output checkable
3. Ask what it cannot support	Tell it to list claims it could not tie to a document	Surfaces the gaps as questions, not silence
4. Cross-check the figures	Ask it to tie the same number across the CIM, financials, and tax return	Contradictions are where deals move
5. Verify the citations yourself	Spot-check the quotes against the documents	The model can still misquote; trust is earned per claim

The order matters. If you ask for a summary first, you anchor on a confident narrative and spend the rest of the time defending it. If you ask for cited claims and a list of what could not be verified, you start from evidence.

Where Claude is strong, and where it fails

It is strong at reading volume fast, restating dense financial language plainly, drafting a question list, and catching a document that contradicts itself when you ask it to look. It fails, quietly, when you let it fill gaps: it will produce a plausible average contract value, a tidy "no outstanding litigation," or a churn number that sounds right and is sourced from nothing. Those are not bugs in the model; they are what a summarizer does when you do not force it to cite.

If you want the discipline pre-built rather than prompting it each time, there is a free, MIT-licensed pack of twenty single-job agents for reading deals this way, described in 20 Claude agents for M&A diligence, each built to show its work and cite or cut.

When a turnkey tool fits better

Prompting Claude by hand works, and for a quick read it is enough. The friction shows up at volume: re-pasting documents, re-writing the prompt, keeping each deal isolated, and checking citations every time. If you are screening deal after deal, a hosted tool that does the cite-or-cut read on your uploaded documents and returns a source-linked brief can fit how you actually work better than a fresh chat each time. That trade-off is laid out in the best due diligence software and in what diligence automation actually catches. You can see the cited output on a synthetic deal, no login, in the sample brief.

Frequently asked questions

Can I use Claude for due diligence? Yes, for reading and first-pass analysis, as long as you force it to cite every claim to a source document and to list what it could not verify. Do not treat an uncited summary as fact.

Is it safe to paste a CIM into a general AI chat? Treat confidentiality and verification separately. On verification, an uncited summary is not safe to act on; on confidentiality, use a tool that keeps each deal isolated rather than a general consumer chat. A purpose-built workspace handles both.

How do I stop Claude from making up numbers? Require citations, ask it explicitly to list claims it cannot tie to a document, and cross-check the same figure across the CIM, financials, and tax return. Anything it cannot source should become a question, not a stated fact.

Claude for due diligence versus a turnkey tool? Hand-prompting is flexible and free but high-friction at volume. A turnkey diligence tool runs the cite-or-cut discipline on your documents automatically and returns a cited brief, which fits better when you are screening many deals.

See the cite-or-cut discipline running on a real-looking deal at Deal OS.

20 Claude agents for M&A diligence, built on one rule: cite the source or cut the claim

DEVALAND — Sat, 20 Jun 2026 08:55:20 +0000

Most people use AI to summarize a document.

In an acquisition, the summary is the one thing you can't afford to trust.

If a model invents an EBITDA figure, or quietly rounds a churn number in the seller's favor, or softens a customer-concentration risk into something that sounds fine, that isn't a typo you catch later. It's the basis you priced a deal on. By the time you find out, you've signed the LOI, burned a month, and spent credibility you can't get back.

So I built a pack of agents the other way round. Not one big agent that reads everything and hands you a confident paragraph. Twenty small ones, each with a single job, each built to show its work, and to cite every factual claim back to the source it came from, or cut the claim before you ever see it.

It's called The Deal Team. It's MIT-licensed, free, and up on GitHub.

Why small acquirers feel this hardest

A two-person search fund reviews the same volume of CIMs, financials, and legal documents as a mid-market PE firm. Same teasers, same data rooms, same seven-figure decisions, minus the analysts, the associates, and the back office that institutional buyers take for granted.

That's where AI looks like an obvious win. And it is, right up until you remember that the failure mode isn't "the summary is a bit rough." The failure mode is a number that's wrong, stated with total confidence, that you can't tell apart from a number that's right.

Reading a CIM, tying it to the financials, and building a question list is days of principal time per target, most of it spent on deals that die. You want to move faster. You just can't move faster by trusting output you can't trace.

The fix isn't a smarter summarizer. It's a discipline: every figure traces back to a line in a document, or it doesn't get used.

What "cite or cut" actually looks like

The clearest demonstration isn't in the prompts, it's in a finished brief. There's a full sample brief on a synthetic target, "Project Sentinel," a fire-and-safety inspection business. It's worth opening because it shows the discipline doing the one thing a plain summary can't: catching a document contradicting itself.

The CIM's page 8 says the customer base is well diversified, with no single customer over 15% of revenue. Page 9 of the same document lists the actual accounts, and the top one runs about 22% of revenue. Both pages can't be true. A model that just summarizes the CIM repeats the page-8 diversification line, and you price the deal on a concentration risk you never saw. The brief instead shows you both passages, verbatim, each cited to its page, and leaves the call to you.

Then there's the part most AI tools never show: the discard log. While drafting, the model produced the claim "the company has no outstanding litigation." Nothing in the documents supported it, so it was cut before it reached the brief and turned into an action instead: request a 10-year litigation history. Same fate for a tidy "average contract value of about $18,000" (couldn't be tied to any single sourced passage) and "technician retention is strong" (the CIM said nothing about turnover).

That's the whole philosophy on one screen. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim doesn't reach you at all, it becomes a question.

What's actually in the pack

Twenty agents, ordered to follow a real deal across five stages. Feed the output of one into the next.

Sourcing: build a focused target list, sharpen a vague thesis into something testable, and write owner-direct outreach that gets replies without sounding like a broker.
Screening: read a teaser for what's really being sold, score it against your thesis in minutes, scan for the early red flags that kill deals later, and force a clean go/no-go so you don't drift into dead deals.
Diligence: read a full CIM into a cited, plain-English brief, turn financials into signal, stress-test customer concentration, sanity-check the market story, and run a Citation Checker that audits any AI or analyst summary back against the source.
Valuation: frame a range with comparable logic, sketch whether it works on leverage, and catch yourself before you overpay.
Process and close: generate the diligence question list you'd have forgotten, prep a sharp management meeting, and turn the whole thing into a clean IC-style memo.

The full bench, all twenty, each with what it does, is in the README.

How to run them (about thirty seconds)

Each agent is a single SKILL.md with a copy-ready prompt. Two ways in:

Copy-paste, no setup. Open claude.ai, copy the prompt from any agent, paste it as your first message, then attach the deal: teaser, CIM, P&L, your thesis. For one you'll reuse, drop the prompt into a Claude Project's custom instructions and every chat in that project is that agent.

Install as Claude skills. Each folder under skills/ is self-contained. Drop the ones you want into your skills directory and Claude reaches for the right agent when you describe what you're doing:

git clone https://github.com/MariusGithub13/open-deal-team.git
cp -r open-deal-team/skills/* ~/.claude/skills/

The one habit that makes any of this safe

When an agent gives you a number that matters, check the citation it points to. Every figure should trace back to a line in your document. If it can't, treat it as unverified.

That single habit, check the source every time, is the difference between using AI on a deal and trusting it. The agents are built to make that check fast: they tell you where each claim came from, and they tell you when something they needed was missing instead of papering over the gap.

From prompts to product, the honest version

I'll be straight about where this ends and where my product begins, because pretending otherwise would undercut the whole "cite it or cut it" point.

The agents give you the discipline: a prompt you run yourself, one document at a time, checking each citation by hand. That's genuinely most of the value, and it's free forever. If you only ever use the prompts, that's a win, they're yours, MIT, no strings.

The by-hand part is also the part that stops scaling. One agent, one document, one manual citation check is fine for a single target. Across a live data room, a CIM, three years of financials, tax returns, a QoE, a stack of contracts, all of which have to agree with each other, the checking is the work, and it's the work nobody has time for.

That's what I built Deal OS to do automatically: read a whole data room at once, verify every citation against the source, discard the claims it can't verify before you ever see them, and flag contradictions across documents, where the CIM, the financials, and the management call disagree, instead of you running a checker over each pair by hand.

If you're running one deal, use the prompts. If you're running several and the manual checking has stopped scaling, that's where the product earns its place. You can read a full source-cited brief on a sample deal before deciding either way.

One honest note

These are analytical tools to speed up your own judgment and help you read deals faster. They are not investment, legal, tax, or accounting advice, and they don't replace your own verification or your advisers. Always check an agent's output against the source documents, which is exactly what the cited rule is designed to let you do.

If the "cite it or cut it" idea is useful to you, the best thing you can do is star the repo so the next searcher who's drowning in CIMs finds it too. Fork it, adapt the prompts to your own thesis, and tell me which agent earns its keep and which one needs work.

Compiled by Marius Andronie, Devaland, Deal OS.