Products
Pricing Blog Join the waitlist
← Back to BehaviorGraph
The Model

The Large Behavior Model.

A pretrained AI model of how organizations actually work. It reads approved work metadata and answers the questions your org chart cannot: who owns this, who has capacity, which document is the truth, where risk is forming.

Agents are smart. They just don't know the building.

🧭

The context gap

Companies are putting AI agents into more workflows, but agents still lack the context human operators use every day: who knows what, who is trusted, who has capacity, who owns the latest answer, and where risk is forming.

🔁

Rebuilt manually, everywhere

Today that context is reconstructed by hand in every company and every workflow: routing tables, escalation configs, tribal knowledge. It breaks every reorg.

🧬

Work behavior can be modeled

Work events form sequences, and a model can learn from them the way a language model learns from sequences of words. With enough breadth, it learns patterns that carry over to new companies.

How LBM works before a company has mature ONA Collaboration data 5 signal families meetings · messages · work items knowledge artifacts · peer/org signals LBM behavior layer behavioral tokens → shared representation who knows what · who relies on whom what is treated as true · where risk forms Use-case probes agent routing diligence and integration knowledge and workforce risk More signal families raise confidence 1-2 families first behavior map 3-4 families richer probes 5 + ONA / peer strongest calibration Concept-level only: token schema, source weighting, and calibration recipes are withheld.

The deployment path. Ordinary collaboration data becomes a first map of how work moves; existing network analysis and peer feedback make the results more precise but are not required to begin.

Six things simple activity counts miss

Not a social graph. Not a productivity score. A model of work behavior that adjusts for role and activity level, so the busiest or most connected person is not mistaken for the expert.

Expertise

Who gets pulled into tickets, incidents, and design reviews when the team needs the answer that unblocks work.

Influence

Whose involvement actually moves a decision: when they weigh in, things get built, delayed, or rolled back.

Source-of-truth

Which runbook, doc, or ticket people actually open before shipping, debugging, or answering customers.

Disengagement & attrition

Who stops reviewing, misses standups, skips planning, or quietly hands off work, months before they formally leave.

Agent routing

Where an agent should send a failed deploy, an unclear ticket, a support escalation, or a question no doc answers.

Integration risk

Which team, system, or single person quietly holds a migration, reorg, or acquisition plan together.

CollaborationeventsBehavioral tokensequenceShared LBMrepresentationExpertise · routingKnowledge source-of-truthDisengagement · attritionInfluence · key-person riskIntegration · hidden valueVALIDATION SURFACEREAL · TRANSFERheld-out seniority  r ≈ 0.50 (0.25 net)REAL · EXTERNALEnron influence  2.12× liftREAL · BENCHMARKpeer-value benchmark  SIM · CONTROLLEDattrition  0.94 AUCSIM · CONTROLLEDdocument truth  0.88 AUC

One model of work behavior, many uses. Each use case reads the same model; validation includes new-company tests, public external data, and controlled simulation.

Validated against real outcomes

Every result is tested against simpler methods and checked to make sure the answer does not come from accidental clues in the data.

0.88 AUCfinds the document people actually treat as canonical, so agents stop citing stale knowledge
r ≈ 0.50applied to a company it never trained on, the model's reads still hold
0.94 AUCpredicts likely departures months ahead, on a scale where 0.90+ is excellent
0.83 AUCcatches quiet disengagement early, detecting withdrawal before it becomes visible
CapabilityResultWhy it matters
Hidden influence: sway beyond the title2.12x better than chanceSurfaces the people whose real sway their title does not explain: risk, compliance, and diligence
Works on a company it has never seenr ≈ 0.50, about 0.25 after adjustmentThe model keeps working on a new company without custom training
Attrition early warning: who may leave0.94 AUCFlags likely departures months before standard HR reports
Same model, flat or steep org0.94 in both org shapesSteep hierarchy to flat org, with no extra tuning
Which document is the truth0.88 AUCStops people and agents using stale knowledge
Quiet disengagement, caught early0.83 AUCDetects withdrawal before it becomes binary
Who can cover a role3.2x better than chanceSuccession planning and operating resilience
Case study 1: Enron, the test the model cannot game.

About 517,000 emails, roughly 170 mapped employees, 1999 to 2002, where who held power and who was prosecuted is court record. Reading communication patterns alone, without being told who was later charged or convicted, the model:

  • Ranks the indicted insiders at more than twice the rate chance would produce, without being told any outcome.
  • Detects the trading desk hardening into an insular ring through the California energy crisis, then dissolving afterward.
  • Surfaces internal code names tied to the fraud about a year before public exposure, clustered around the employees later convicted.

A modern all-digital workplace leaves a far richer signal than phone-era email. For risk, compliance, and diligence teams, this is an early warning system that reads collusion and shadow authority from structure alone.

Case study 2: a real 900-person tech company.

Tested on a real technology company of about 900 people the model had never trained on: fast-moving, socially driven, project-based, under heavy delivery pressure. It read only the company's permissioned calendars and Slack. From those two signals alone, the model rebuilt the company's operating map:

  • About 70 go-to experts the company truly relies on, plus about 100 hidden experts, genuinely relied on, with titles that understate their real role, that a title chart alone would miss.
  • About 35 key-person risks whose exit would break a live workflow, concentrated in senior specialists rather than junior staff.
  • Backup depth: about 70% of people have a peer who could step in, but about 25 critical roles have no real backup, and roughly a quarter of the priority projects depend on a single person.
  • About 85 cross-team bridges holding otherwise separate groups together, who also carried the highest overload and burnout risk, and about 7 clusters at risk of becoming isolated silos.

Two checks compared the model's answers with facts it was never shown: job function, identified from work patterns alone, and departures, predicted from past calendar behavior and confirmed against independent external records. No surveys, no documents, no answer sheets handed in.

One model. Every use case is a product.

Connect permissioned data once, evaluate the model on your own organization, then use any capability through the REST API or expose it to your agents over MCP.

Agent routing

Send a task, question, or exception to the person who actually knows it and has capacity.

Workforce-risk monitoring

Likely departures and disengagement, months before they show up in HR.

Knowledge governance

Which document is the source of truth and who owns it.

Change management

Who actually drives or blocks a rollout.

Compliance & risk

Shadow authority and collusion patterns read from structure alone.

M&A diligence

How work actually moves beneath the org chart: integration risk and knowledge continuity.

Evaluate it on your own organization

Start with the free evaluation: connect calendar metadata, get one full diagnosis back over MCP, judge the results before you pay anything. The full design brief and validation protocol are available for technical diligence.

Start the evaluation See pricing