The plan is plausible. The system is not stable.
Where does the success story break?
Pressure-test Northstar Mutual’s AI transformation as an interacting system of incentives, dependencies, controls and human adaptation. Stressors combine; defenses change the failure path. Outputs are a structured discussion aid, never an empirical forecast.
4,800US mutual
$12M18 months
42 pointsscarce
4uneven maturity
ACausal failure map
Choose a node for its mechanism and signal. Edges distinguish causes, amplifiers, dependencies, detection and mitigation.
FailureStakeholderInitiativeSystem / control━ cause · ╌ amplifier · → dependency
This simplified graph is backed by explicit mechanism scores. Select a node to inspect it.
Seed inventory · actors, ownership, quality and dependencies
All 16 failure mechanisms · category, causal story and weak signal
BCompose pressure
Choose at least three stressors. Interactions amplify shared control congestion, vendor dependence and legitimacy pressure.
Metric corruption lab · what leadership rewards becomes visible
Same inputs → same ensemble; no external data.
COutcome envelope
Exploratory model, not calibrated to Northstar or industry base rates. Ranges are 10th–90th percentiles of explicit ordinal assumptions plus seeded perturbations. They express relative sensitivity, not real-world probabilities.
DFailure-path replay
Same scenario and defense state as outcomes. Walk decision points; graph state follows.
— / —
Run a scenario to generate a path
A representative path makes weak signals and intervention windows concrete.
Intervention windows appear at each step.
EMitigation portfolio
Defenses consume budget, calendar and scarce change capacity. Some shift costs or create new risks.
Budget used $0 / $12MCapacity 0 / 42
FAdversary lenses
Distinct actors, duties and evidence; not six versions of “be careful.”
GPre-mortem & dissent
Month 18 failure story; distinguish seed facts, scenario assumptions and interpretations.
“We scaled the easy denominator.”
Likely sequence · simulated interpretation
Uncomfortable recommendation
Dissent to preserve
Seed fact stated in brief Model assumption hypothetical
HDecision package
Download local JSON: scenario, assumptions, defenses, ranges, replay, dissent and unresolved questions.
Remaining unknowns
Actual case-mix drift; vendor recovery; regulator interpretation; reviewer throughput; workforce adaptation; peak-load viability of manual fallback.
Generated locally; no data is sent.
Model assumptions & interpretation limits
- Mechanism intensity is ordinal (0–100), seeded by the fictional scenario.
- Stressors add pressure; named interaction terms amplify congestion, legitimacy loss and vendor concentration.
- Mitigation coverage weakens specific paths; delay and side effects are explicit.
- 512 deterministic perturbations sample assumption sensitivity; they are not observed frequencies or forecasts.
- Financial result is a simplified illustrative proxy; time-to-value is modeled months to usable benefit.