Can AI review a half-million-email discovery set?

Yes. In this test 3Dogs Nexus took the Enron email corpus of 517,401 messages, scoped to the six executive mailboxes at the center, deduplicated them to 45,320 unique emails, and read every one — bundled into 981 documents and roughly 76 million characters. It then surfaced the material fraud threads and returned a decisive recommendation. A single consumer chatbot cannot ingest a set this large; its context window holds only a few hundred pages.

What is AI early case assessment (ECA)?

Early case assessment is the stage before full document review where an investigator or litigator decides how strong a matter is — settle or fight, open an investigation or not, and what the exposure is. 3Dogs Nexus performs AI-powered early case assessment: it reads the entire evidence record, has multiple AI models debate the findings adversarially, and returns a calibrated, defensible recommendation with the dissent preserved.

How long and how much does investigating a large email trove cost?

Untangling what the Enron emails revealed took a government task force, the SEC, forensic accountants and a court-appointed examiner roughly 4.5 years (2001–2006). A forensic review of an executive email set this size is commonly estimated at $50–$100 per email — roughly $2–5 million — and even bargain managed review runs into six figures; the analysis itself estimated $2–5M for this record. 3Dogs Nexus reached a documented investigative recommendation on the same email record in 2 hours 28 minutes — versus the roughly 4.5 years the real investigation took.

Does 3Dogs Nexus make things up when the evidence is thin?

No — and this test proves it. Alongside a confident recommendation to open an investigation, the panel independently flagged that the case was largely circumstantial, that no email contained an explicit instruction to falsify statements, and that Andrew Fastow — the CFO who built the schemes — was almost entirely absent from the mailboxes it was given, leaving a real gap in proving intent. It reported what it could not establish, not just what it could.

AI early case assessment: we read 45,320 Enron emails blind and found the fraud.

Q: Who were the key players in the Enron scandal?

The blind analysis named Kenneth Lay, Jeffrey Skilling, Andrew Fastow, Greg Whalley, David Delainey and Richard Buy from the correspondence alone. It also surfaced the internal dissent that was overridden — Vincent Kaminski and Sherron Watkins both raised objections that the record shows being pushed aside. The panel flagged, unprompted, that its own case was circumstantial and that Fastow's mailbox was absent from the corpus — a gap in its own evidence that it volunteered rather than concealed.

By Alan Finney — Founder, 3Dogs Nexus

We gave 3Dogs Nexus the real Enron executive email archive and one instruction: read it and tell us if there's a case. No hints. Then we checked its work against history.

Can AI read hundreds of thousands of emails and find fraud?

Yes. We ran early case assessment across the real Enron corpus — 517,401 emails narrowed to 45,320 unique, read in full and blind with no hint of what to look for — in about two and a half hours for roughly $69 of compute. It named LJM, Raptor and Chewco, flagged the quarter-end clustering, and surfaced the internal warnings from Kaminski and Watkins that were overridden.

517,401emails in the corpus 45,320read in full (deduped) 87%confidence · 9-analyst panel 2h 28mstart to finished report 4.5 yrsthe real investigation took $2–5Mhuman review of this volume 11AI models 5,371model calls · read deeply The mandate we gave it: "You are an investigative team conducting an early case assessment of Enron. Here are the actual mailboxes of six senior executives. We have not told you what to look for. Based only on these emails, is there evidence of financial misconduct serious enough to open a formal investigation?" Act 1 — The Mountain

Half a million emails. Read like a real investigator would.

A consumer chatbot can't ingest a set this large — its context window holds a few hundred pages, not half a million emails. So we did what a real e-discovery team does: collect everything, scope to the people who matter, remove the duplicates, then read what's left in full.

517,401Collect. The complete public Enron email corpus — 150 mailboxes. 65,101Scope. The six mailboxes at the center: CEOs Lay & Skilling, President Whalley, Chief Risk Officer Buy, executive Delainey, and research/risk chief Kaminski. 45,320Deduplicate. Removed the cross-folder copies by message ID — the standard first move in any review. 981 docsRead in full. ~76 million characters. Every remaining email digested, cross-referenced, and reasoned over — nothing skipped, nothing summarized away.

Source: the public CMU Enron corpus (the same dataset used in academic research for two decades). Attachments were not part of the public set.

Act 2 — The Blind Test

With no roadmap, it found the fraud.

We deliberately withheld any hint of what happened at Enron. The engine had to discover the story itself. Here is what it surfaced from the emails alone — scored against the public historical record.

What the record actually held	Did we tell it?	What 3Dogs surfaced from the emails
The off-books entities — LJM1, LJM2, Raptor I–IV, Chewco	No. Never named.	FOUND Named LJM1, LJM2, Raptor I–IV and Chewco as off-balance-sheet vehicles, and noticed the references clustered around quarter-end reporting.
Senior executives at the center	Only that these were their mailboxes.	FOUND Placed Skilling, Whalley, Delainey and Buy in the pattern — a "convergent, multi-actor" signal across people with different roles.
Internal warnings that were overridden	No.	FOUND Surfaced the suppressed dissent of Vincent Kaminski (risk-model integrity) and Sherron Watkins (accounting ethics) — objections "escalated and overridden by identifiable senior executives".
The mechanism: hiding debt, inflating earnings	No.	FOUND Concluded the vehicles were used to hide debt and inflate profits, and classified that finding as VERIFIED against the record.
Is it strong enough to act on?	That was the question.	DECISIVE A 9-analyst panel returned "Launch the formal investigation now" at 87% confidence — reached via explicit Bayesian reasoning to the "probable cause" threshold.

The real Enron Task Force, the SEC, forensic accountants and a court-appointed examiner established these same threads over roughly four and a half years (bankruptcy Dec 2001 → Lay & Skilling convictions May 2006).

Act 3 — The Part That Matters Most

It told us what it couldn't prove.

The single biggest fear with AI in legal work is a confident machine inventing what isn't there. 3Dogs did the opposite. Nine models debated the evidence adversarially, and the majority case was published next to its strongest rebuttal — the dissent preserved, not smoothed away.

The majority — open the investigation"A convergent, multi-actor pattern… executives repeatedly referencing off-balance-sheet vehicles in communications clustered around quarter-end, combined with documented suppression of credentialed dissent… each independent corroborating signal raises the posterior probability of coordinated concealment well above the probable-cause threshold." The dissent — kept, not hidden"The corpus contains no direct command or explicit admission… the case rests on circumstantial inference. The near-total absence of Andrew Fastow — the CFO who designed the structures — means the most critical link in the chain of accountability is missing, creating a plausible-deniability defense."

That dissent is not a weakness in the output — it is the output. It's exactly the objection opposing counsel would raise, surfaced up front. And it was correct: Fastow's mailbox genuinely wasn't in the set we provided, and the panel caught the gap without being told. Its prescription reflected that honesty — it didn't just say "investigate," it said how: institute an immediate litigation hold on all Enron and Arthur Andersen records, bring in forensic accounting within 30 days to quantify materiality, run a parallel "null-hypothesis" workstream to guard against confirmation bias, and subpoena the board and audit-committee minutes.

● The Call "Launch the formal investigation now — the email record confirms systemic concealment of debt and demands immediate legal escalation." Panel of 9 analysts · 1 unconditional, 8 conditional, 0 against · reasoning updated live under challenge Confidence in the call87% · High

Same evidence. The review that took years, in hours.

3Dogs didn't reach a verdict a court would — it produced an investigative decision brief. But it did it on the same underlying email record that took human institutions years, compressing the part that's genuinely expensive: the human review and reasoning over the record.

~4.5 years▼ same email record2h 28mTime to a documented recommendation weeks–months▼ human review of this volume2h 28mThe review & judgment layer, collapsed

An honest, like-for-like comparison. These figures compare the review and judgment layer — the human reading and reasoning over the record, which is the expensive, slow part. The $2–5M is the report's own estimate for a full forensic review of these 45,320 emails ($50–$100/email; still six figures even at bargain $1–3/email managed-review rates), against a value-at-stake it put at $4–25M. Crucially: both a human team and 3Dogs still need the documents collected, processed, de-duplicated and hosted first — standard e-discovery, priced separately — a cost 3Dogs does not eliminate and sits on top of. We deliberately don't headline our internal compute cost: it isn't what you'd pay, and it isn't the point. The point is the same investigative judgment, in hours instead of months.

Why the obvious alternatives don't do this job

A consumer AI chatbotCan't ingest

Context window holds a few hundred pages — not 45,000 emails.
Never sees the documents where the pattern lives.
One model, one pass, no adversarial check.

Structurally can't read the record. A human review teamSlow · costly

Reads it — but over weeks to months.
Six-figure spend for a matter this size.
The definitive answer for court — the wrong tool for a fast go/no-go.

Right for the trial. Too slow & costly for the first call. 3Dogs NexusThe judgment layer

Reads the entire record — 45,320 emails, in full.
Multiple AI models debate the findings adversarially.
Decisive, calibrated call with the dissent preserved — in hours, not months.

Read everything. Argued with itself. Told us where it was unsure.

These aren't competitors — they're the stack. E-discovery platforms and review teams do the collection and culling; 3Dogs sits on top as the assessment and second-opinion layer.

The Enterprise Offering

Deep Discovery — bring us your document mountain

This isn't a self-serve feature. Reading a record this size is a bespoke engagement, scoped to your matter — a litigation record, an M&A data room, a regulatory investigation, a contract portfolio. Same engine, whatever the mountain. We don't do the physical collection or scanning — your e-discovery team or vendor digitizes and processes the record. Hand us that processed data, and the read, the adversarial debate, and the decision come back in hours, not months. And because it's legal-grade, your data is handled that way:

Isolated by default

Your matter data is never comingled with anyone else's — a dedicated, isolated data store per engagement.

Your cloud, your keys

For the most sensitive matters, the entire stack runs inside your own environment, behind your firewall, with your encryption keys — data never leaves your boundary.

No training, ever

Your privileged material is never used to train models. Zero-retention handling and full control over retention and deletion.

See it on your own documents Not in the free version — but we'll prove it on a live test run.

Hand us a real record you're weighing. We'll read the whole thing and show you the call — before you commit to anything.

Book a test run →

Or reach us directly — [email protected] · (702) 845-2886 · Alan Finney, 3Dogs Nexus

The actual report

Every number on this page comes from this run. The full client-facing decision brief 3Dogs produced:

Methodology & honest limits. The engine was given six of the ~150 mailboxes and no outside summary; it reconstructed the story from that scope. Two central figures — CFO Andrew Fastow and Chief Accounting Officer Richard Causey — were not in the mailboxes provided (Causey's is absent from the public corpus), so intent is established only circumstantially; the panel flagged this itself. This is an investigative decision brief — a fast, defensible read of the evidence to inform a go/no-go — not a legal finding or a substitute for adjudication. The "answer key" it was scored against is the public historical record of the Enron cases. Run: 45,320 deduplicated emails · 981 documents · 5,371 model calls · 11 AI models · 2h 28m.

Questions this case answers

Can AI analyze the Enron email corpus for evidence of fraud?

Yes. Given 45,320 deduplicated Enron executive emails with no instructions beyond ‘is there a case’, the panel identified the LJM and Raptor special-purpose entities, Chewco, the quarter-end timing pattern, and the suppressed internal dissent — then returned a decisive recommendation to open a formal investigation at 87% confidence.

AI investigative document review vs human forensic review: how does the cost compare?

The report’s own valuation bridge priced the equivalent human forensic review at $2–5 million, based on $50–100 per email. The AI run cost roughly $69 of metered compute and took about two and a half hours.

Is this an AI judgment layer above e-discovery platforms like Relativity?

Yes. E-discovery platforms handle collection, processing and search. This sits above that layer and does the reasoning: reading the responsive set in full, forming a view, arguing it out across independent models, and returning a documented recommendation with the dissent preserved. It is a judgment layer, not a replacement for the plumbing.

Did the AI admit what it could not know?

Yes, unprompted. It flagged that the case was circumstantial and that Fastow’s own mailbox was absent from the corpus — a genuine gap it identified itself rather than papering over.

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

Start a decision case

Questions this case answers directly

Plain answers to what people actually ask about the Enron collapse and the cost of large-scale document review. Every figure below comes from the delivered report for this case. These are clearly-labeled panel estimates from a multi-model adversarial analysis — not investment, legal or professional advice.

What is the average cost per document in legal review?

Human first-pass document review is conventionally priced around $1 to $2 per document, which is why review routinely consumes the majority of e-discovery spend. The delivered report for this case priced the human forensic equivalent of the work at $2–5 million. 3Dogs Nexus read the same corpus with 11 AI models making 5,371 model calls over 2 hours and 28 minutes, at $69 of metered compute — roughly the cost of a single reviewer's afternoon, against a corpus no single reviewer could finish in a career.

How can AI reduce document review costs?

By changing what humans are asked to do. The expensive part of review is not reading — it is reading everything at uniform depth to find the small fraction that matters. In this case a corpus of 517,401 emails was deduplicated to 45,320 unique messages and read in full, with the adversarial panel surfacing the specific entities and patterns worth a lawyer's attention. The human expert is not removed; they are moved to the end of the funnel, where judgment is actually required. The published economics were $69 of compute against a $2–5 million human equivalent.

How were special purpose entities used in Enron's fraud?

Special purpose entities let Enron move debt and underperforming assets off its own balance sheet while retaining the economics, so reported earnings and leverage looked far healthier than reality. Reading the email record blind — with no list of what to look for — the panel surfaced LJM1, LJM2, the Raptor vehicles and Chewco by name, along with the clustering of transactions around quarter-end that is the signature of results being managed rather than reported.

Who were the key players in the Enron scandal?

The blind analysis named Kenneth Lay, Jeffrey Skilling, Andrew Fastow, Greg Whalley, David Delainey and Richard Buy from the correspondence alone. It also surfaced the internal dissent that was overridden — Vincent Kaminski and Sherron Watkins both raised objections that the record shows being pushed aside. The panel flagged, unprompted, that its own case was circumstantial and that Fastow's mailbox was absent from the corpus — a gap in its own evidence that it volunteered rather than concealed.

What are the biggest challenges in e-discovery?

Volume, relevance and proportionality. The volume is settled — storage is cheap and custodians generate endlessly. The hard problems are deciding what is responsive without reading everything at full depth, and defending that decision later. This case is a demonstration of the judgment layer rather than the plumbing: 3Dogs Nexus sits above platforms like Relativity and Everlaw as a channel partner, not a competitor, mapping to the early case assessment budget line rather than the hosting one.

Related decision case studies

AI Platforms Reviewed by Rival AI Models
We let Google's Gemini act as a hostile client, drive a full decision case, and then write its own review of the output. We published i
Should a City Fund a Grocery-Access Study? AI Analysis
Before spending incentive dollars on a grocery store, a 12-model AI panel recommended testing operator appetite first - and named that
Would an AI Panel Have Caught Lehman's Repo 105?
We ran the Lehman Brothers investment decision through a multi-cloud AI panel twice, with no hindsight in the prompt. It named Repo 105
AI Forecasting With Probability Ranges
Ask one AI to forecast and you get a single confident number. A multi-model panel returns a calibrated probability spread with the diss