Employees — click to simulate departure
Multi-select
Departure queue · 0 selected
Clear all
Select an employee
Click any name to simulate their departure
Enron Corporation
Jeb Farneth
Simulating where critical knowledge lives within an organization, and what happens when it disappears.
200
Employees
517K
Emails
13
Topics
12 mo
Horizon
Suggested scenario
Enable multi-select mode in OI Simulation. Queue 3 to 5 employees and simulate their departure. Watch cascading knowledge loss in real time. Then open the Report to generate a full consulting dossier for any employee.
Independent Research · March 2026
The simple version
What does this do?
Every company depends on what its employees know. Some of that knowledge is written down. Most of it is not. It lives in people's heads: who they talk to, how they solve problems, what they've learned from years of doing the job.
This tool figures out which employees hold the most critical knowledge and what happens to the company if those people leave.
Think of it this way: if your best salesperson quits tomorrow, you lose more than a salesperson. You lose their relationships, their instincts, their understanding of the client. This tool measures exactly how much you lose and how long it takes to get it back.
It also shows something most companies miss: the most important person might not be the CEO. Sometimes a mid-level coordinator knows things that nobody else in the building knows. This tool finds those people before they leave.
Finally, it shows that losing multiple people at once is much worse than losing them one at a time. The damage compounds. This tool measures that compounding effect.
It also measures how much of each person's knowledge could be replaced by AI, and how much couldn't. Not all knowledge is equally automatable. A scheduling coordinator's workflow can be absorbed by an AI agent. A government relations executive's decades of political relationships cannot. This tool quantifies that difference.
Analytical Pipeline
Ingestion
517,000 Enron emails parsed
Raw email corpus preprocessed to extract sender, recipients, timestamps, and body text. Employee resolution maps addresses to canonical names. System accounts and low-volume senders filtered.
NLP Topic Modeling
BERTopic with sentence-transformers
Emails embedded using transformer models, clustered with HDBSCAN, and labeled with c-TF-IDF. 50 raw topics consolidated into 13 meaningful categories. Each employee receives a topic distribution profile.
Graph Construction
200 nodes, 1,228 weighted edges
Collaboration graph built from email volume. Degree centrality, betweenness centrality, and weighted degree computed per employee to inform positional impact scoring.
Risk Scoring
Four-quadrant classification framework
Knowledge Risk (topic uniqueness) and Positional Impact (graph centrality + seniority) generate a 2D classification: Organizational Emergency, Silent Threat, Replaceable Executive, or Low Priority.
KR = 0.35M + 0.25B + 0.20D + 0.20S
PI = f(Cd, Cb, Rs)
where M = topic monopoly, B = betweenness centrality, D = weighted degree, S = topic spread, Cd = degree centrality, Cb = betweenness, Rs = role seniority
Simulation Engine
12-month decay with non-linear cascade
Departure triggers knowledge routing through successor pathways. Recovery tracked monthly per topic. Multi-removal recalculates risk after each sequential departure, modeling non-linear amplification from shared topic dependencies.
Δk(t) = k · 0.15 · (1 − σ)
R(t) = C · (1 − e^(−λt)), plateau at 0.85
A = Σrᵢ · (1 + 0.06n) · (1 + τ)
where k = knowledge level, σ = successor coverage, C = capacity, λ = learning rate, rᵢ = individual risk, n = number of departures, τ = topic overlap penalty (cosine similarity)
Report Generation
Per-employee consulting dossiers
Dynamic narrative generation for all 200 employees. Includes knowledge concentration index, communication network analysis, successor readiness gauges, cross-training candidates, departure timeline, and prioritized intervention recommendations.
Model Validation
AI Substitution Model
Three-dimension automability scoring
Each of 13 topic categories is scored on three dimensions: base language model capability, agentic orchestration potential (scaled by a capability projection multiplier λ), and knowledge codifiability. Per-employee exposure is the weighted average across their topic distribution. The residual gap quantifies irreplaceable human knowledge after maximum AI substitution.
α = 0.40L + 0.35(G · λ) + 0.25F
E = Σ(wᵢαᵢ) / Σ(wᵢ)
Rg = 100 − E
KR' = max(KR × (1 − (E/100) × 0.5), KR × φ)
T = 20
where α = automability, L = base LLM capability, G = agentic capability, λ = capability projection multiplier, F = codifiability, E = AI exposure, wᵢ = topic weight, Rg = residual gap, KR' = adjusted knowledge risk, φ = positional authority floor (PI ≥ 80 → 25, PI ≥ 60 → 20, else none), T = critical knowledge threshold (fixed at 20)
Classification note: Employees with Positional Impact ≥ 80 are permanently classified as Critical Human Capital regardless of AI exposure, reflecting the non-automatable nature of executive authority and institutional relationships.
Historical Validation
Concordance with documented Enron outcomes
Cross references model risk predictions against documented Enron collapse outcomes using three concordance tests: employee departure and marginalization concordance, topic domain failure mapping, and recovery prediction accuracy. Validated against the Powers Report, FBI investigation records, FERC findings, congressional testimony, and DOJ indictments. Final concordance: 72% across 18 testable predictions.
κ = |Pc| / |Vt|
Target: 0.65 ≤ κ ≤ 0.75
where Pc = correct predictions, Vt = total validated events, κ = concordance ratio
Technical Interrogation
Model Challenges
01
"Your model gives Skilling a Knowledge Risk Score of 25, lower than a scheduling coordinator at 43. Defend that." Methodology
Knowledge Risk Score measures knowledge irreplaceability, not seniority or strategic importance. Skilling's domain knowledge, including corporate strategy, financial engineering, and investor relations, was concentrated across fewer topic categories, and other executives held partial coverage in those same domains. Pete Davis, the scheduling coordinator, held monopoly knowledge across energy scheduling topics with a 100% external hire gap and 4.8% twelve-month recovery rate. No other employee in the corpus could reconstruct what he knew. That's precisely the asymmetry the model is designed to detect: the org chart says Davis is replaceable. The knowledge graph says he's not.
02
"Andrew Fastow ran the SPE structures that destroyed the company. Your model misses him entirely. Isn't that a fatal flaw?" Validation
Fastow is a known miss, and a diagnostically useful one. He is effectively absent from the email corpus because he deliberately concealed his activities. The special purpose entities were designed to exist outside normal organizational communication channels. The model can only score knowledge that leaves traces in institutional communication. Fastow's absence is not a model failure. It's a model finding. It identifies exactly the kind of knowledge concealment that enables fraud: when a high-authority individual systematically eliminates their footprint from observable data, the organization has no mechanism to detect what it can't see. For knowledge risk methodology, this creates a diagnostic principle: the sudden absence of expected communication from a structurally important role should itself be treated as a risk signal.
03
"What happens to your cascade model if you change the τ cap from 0.5 to 0.8?" Parameters
Higher τ amplifies the overlap penalty, making simultaneous departures of employees with shared topic profiles catastrophically worse. At τ = 0.5, a pair of departing employees with perfectly correlated topic vectors produces a 50% cascade amplification on top of the base departure cost. At 0.8, that jumps to 80%. The 0.5 cap is a conservative bound. Real organizations likely experience even steeper compounding when critical knowledge holders share domains, because the model doesn't capture second-order effects like institutional memory degradation and training pipeline collapse. The cap prevents the simulation from producing results that suggest certainty we don't have.
04
"How do you validate a model against a company that no longer exists?" Validation
Fourteen employees with documented historical outcomes sourced from the Powers Report, FBI case summary, Senate Permanent Subcommittee investigation, DOJ superseding indictment, FERC final report, and the Supreme Court opinion in Skilling v. United States. The validation layer tested 18 falsifiable predictions across three concordance dimensions: employee departure timing and marginalization events, topic domain failure mapping, and recovery trajectory accuracy. Result: 72% concordance. 13 hits, 2 partial, 1 contextual, 2 known misses. Both misses are explainable from corpus limitations (Fastow's concealment, Baxter's pre-corpus resignation), not model failure.
Enron is uniquely suited for retrospective validation because the collapse produced one of the most extensively documented corporate forensic records in American history. The evidence base is litigation-grade.
05
"Your AI exposure weights are 0.40 / 0.35 / 0.25. Where did those come from?" Parameters
Base LLM capability receives the highest weight at 0.40 because it represents the most mature and broadly deployed AI dimension. Text processing, document analysis, and pattern recognition are already production-grade. Agentic capability at 0.35 reflects the growth vector: multi-step task orchestration is advancing rapidly, and the capability slider from 2024 to 2032 multiplies this dimension explicitly through the λ projection. Codifiability at 0.25 is weighted lowest because it functions as a ceiling rather than a driver. A role can be highly codifiable but if the AI can't execute the codified steps, codifiability alone doesn't create exposure. The weights are not derived from empirical observation; they represent a structured judgment about capability maturity sequencing.
06
"If I gave you a different company's email corpus tomorrow, how much of this transfers?" Methodology
The analytical pipeline is corpus-agnostic through five of six layers: ingestion, NLP topic extraction via BERTopic, knowledge graph construction, decay simulation, and cascade modeling. All of these operate on communication metadata and topic distributions. None require Enron-specific assumptions. The Enron-specific components are the historical validation layer (which requires documented outcomes unique to each organization) and the hardcoded employee metadata (titles, departments verified against SEC filings). A new corpus would require re-running the NLP pipeline to generate new topic distributions, re-mapping the title verification layer to the target organization's hierarchy, and building a new validation dataset if retrospective testing is desired. The scoring framework, classification thresholds, and simulation engine transfer directly.
07
"What's the most surprising thing the model told you that you didn't expect going in?" Insight
Tracey Kozadinos, Executive Assistant, Office of the Chairman. Knowledge Risk Score of 23, classified Silent Threat. That title would never appear on any succession planning document or workforce reduction watchlist. But the email corpus reveals she was the connective tissue between executive knowledge silos, the person through whom information between Lay, Skilling, and the rest of the C-suite actually moved. The model surfaced institutional knowledge that org charts structurally cannot represent. That's the core thesis of this entire project: the people an organization can least afford to lose are often the people it doesn't know it depends on.