Home › Intelligence › Platform Intelligence

Platform Intelligence Engine

The Platform Gets Measurably Better
Every Time Someone Uses It

Every assignment outcome is labelled and linked to the coaching signals that preceded it. The models retrain. Predictions improve. Early warning patterns get more specific. Content that moves the needle gets amplified. This is the data moat.

The Core Argument

Every In-Person Trainer Walks Out With Nothing.

No data on whether the learning stuck. No signal on what helped or hurt. No ability to improve the programme based on outcomes. The trainer who ran a Japan workshop in 2019 is running the same workshop in 2025 — because they never saw what happened after the workshop ended. Bearings sees everything that happens after. And it feeds every signal back into the system.

🎯
Outcome Prediction
At week 4, predict completion probability at month 12. The model improves every time a new assignment outcome is labelled — more data, sharper predictions.
🔍
Pattern Discovery
AI analyses population data weekly for combinations not yet captured in the 10 seeded Early Warning patterns. Validated discoveries are promoted to the Early Warning Engine.
📊
Intervention Ranking
Tracks 30-day and 90-day outcomes for every intervention. Ranks intervention types by effectiveness per pattern — so the recommended action gets better with every cycle.
✏️
Content Effectiveness
Scores every coaching response type by downstream engagement and domain progress. The AI Mentor improves continuously as ineffective content is identified and replaced.

The Data That Makes It Work

The Outcome Data That Makes Predictions Possible

The outcome record links every assignment's final outcome to the coaching signals that preceded it at weeks 1–4, 5–8, and 9–12. This is what makes prediction possible. And it must be built and collecting data from the first client onward — every assignment that completes before this entity exists is training data permanently lost.

outcome record Fields
outcome_type · assignment_completed | early_return | repatriation_failure | retention_12m
weeks_1_to_4_features · engagement_rate, domain_scores, sentiment_avg, partner_enrolled, mission_rate
weeks_5_to_8_features · Same feature set at weeks 5–8
weeks_9_to_12_features · Same feature set at weeks 9–12
assignment_metadata · origin, destination, type, duration, industry, role_level, partner_enrolled, pre-departure training
interventions_taken · array of {type, week, 30d_outcome} for all interventions during assignment
⚠ Build This First
The outcome record must be built and active from the first client deployment. Every assignment that completes before it exists is training data permanently and irreversibly lost. The cost of building it early is low. The cost of not building it is losing 12–24 months of outcome data.
Model Training Pipeline
1
Outcome labelled
Assignment completes or ends early → outcome record row created with all feature fields
2
Trigger check
50th new label since last run → retraining job queued automatically
3
Model trains
Outcome predictor, intervention ranker, pattern discovery all retrain simultaneously
4
Analyst review
Bearings analyst reviews accuracy metrics before promoting new model version to production
5
Platform improves
New model deployed → predictions sharpen, intervention rankings update, new patterns queued for review
Current model: v2.4.1 · Trained on 847 completed assignments · AUC: 0.81

The Outcome Predictor

A Trajectory Assessment for Every Active Assignee

Available in the Assignee Detail view for any assignee with 6+ weeks of data and a medium-or-high confidence score. Shows completion probability as a directional gauge — not a percentage — with a plain-language interpretation and the top 3 risk factors described in plain language, not model terminology.

  • Completion probability as a colour-coded gauge: green (>75%), amber (50–75%), red (<50%)
  • AI-generated 1–2 sentence interpretation written without numbers, probabilities, or model terms
  • Top 3 risk factors described as "Lower-than-expected social engagement" not "social_rate_4w=0.23"
  • Confidence badge: High / Medium / Insufficient data — low confidence predictions are not shown at all
  • Permanent caveats note: "Directional guidance — not a definitive prediction. Individual circumstances vary."
Trajectory Assessment — Marcus R.
⚠️
Monitor Closely
Medium Confidence
Engagement has dropped sharply since week 8 and the social domain is not progressing at the rate expected for Tokyo. This pattern warrants attention now, before it becomes harder to reverse.
What Is Influencing This
▪Lower-than-expected social engagement for week 9 in this destination
▪Partner not yet enrolled — family domain below threshold for this stage
▪Declining check-in sentiment over last 3 check-ins

Privacy in Model Training

Cross-Client Intelligence With Full Privacy Separation

All model training uses anonymised cohort data. Individual identifiers are stripped before any record enters the training pipeline. Clients contribute to the benchmark pool under a data sharing addendum — and can opt out without losing access to their own programme analytics.

0
Individual or organisation identifiers included in any cross-client model training data
5+
Minimum assignment count required before any cohort benchmark is published
Opt-out
Available — clients who opt out keep their own analytics but lose benchmark comparison
7yrs
Minimum retention period for audit logs. Programme data retained per client contract terms.

See the Platform Intelligence Engine in Action.

We'll show you the Outcome Predictor, walk through how pattern discovery works, and explain the data flywheel that builds competitive advantage over time.

Book a Team Demo → See Pricing

FAQ

Common Questions

How does Bearings use client data?
Outcome data is anonymised before entering any cross-client model. Organisation names, user IDs, and any identifiable metadata are stripped. Only cohort metadata (origin, destination, type, industry sector) and outcome signals are used in cross-client training.
Can clients opt out?
Yes. Clients who prefer not to contribute to the cross-client benchmark pool can opt out. They still receive their own programme analytics (MA3) but lose access to the benchmark comparison lines. Opt-out is configured at account level.
How often do the models retrain?
Retraining is triggered automatically when 50 new outcome labels have been added since the last run. A Bearings analyst reviews the accuracy metrics before any new model version is deployed to production.
Is the Outcome Predictor available at all pricing tiers?
The Outcome Predictor is available at Leadership ($950) and above. It requires a minimum of 6 weeks of data and a medium-or-high model confidence score before showing for any individual assignee.
What happens when a new Early Warning pattern is discovered?
The AI Pattern Discovery job runs weekly. New patterns with confidence >0.4 are queued as PatternInsight records for Bearings analyst review. After validation, a new RiskSignalPattern row is created and the signal evaluation engine begins checking for it automatically.
Does the AI Mentor actually get better over time?
Yes. Content Effectiveness Scores measure every coaching response type by downstream engagement and domain score improvement in the 7 days following delivery. Content with effectiveness <0.3 is flagged for review and improvement. High-scoring content influences generation for similar topics.