Intelligence / Predictive models
Trained on your book, not the industry average
Severity, complexity, reserve adequacy and litigation likelihood — modeled on your own history, versioned, and promoted by a person with a reason recorded.
A coastal homeowners book and an urban commercial auto book do not behave alike. A model trained on somebody else’s mix will tell you so at the worst possible moment.
Score a real fileWhat happens when a model gets stale?
An alert goes to a human. It does not quietly retrain and re-promote itself, because a model that changes its own behavior without anybody deciding is a model nobody can explain later — and explaining it is the job.
Five stages, one human gate
The model lifecycle
- 01
Training
Runs as a background job on your organization’s own data. No cross-tenant training, no shared model learning from other books.
- 02
Registry
Every trained model is versioned and fingerprinted, so the score on a claim can be traced to the exact model that produced it.
- 03
Promotion
Candidate to active is a deliberate step, taken by a person, with a reason recorded. Nothing self-promotes.
- 04
Serving
The active model scores, and both the score and its reasoning are recorded on the claim rather than presented as an opaque number.
- 05
Staleness
Operators are alerted when an active model ages past threshold. The response is human review, not automatic retraining.
What is modeled
Four scores, and what each is for
Each of these changes a decision somebody makes early, which is the only point at which a prediction is worth anything.
| Score | Decision it informs |
|---|---|
| Severity | Initial reserve posture and which desk should handle it |
| Complexity | Whether the file needs a senior examiner from day one |
| Reserve adequacy | Whether the current figure looks light against similar closed files |
| Litigation likelihood | Early reassignment to whoever handles represented parties |
| Entitlement | Predictive scoring, off until bought |
What a score is not
It is an input to a person’s decision. It is never the decision, and no configuration makes it one.
It cannot act
No score sets a reserve, releases a payment, denies coverage or closes a file. Every one of those is a human action with an audit row.
Small books model poorly
A model needs history. If your closed-claim volume is thin, we will tell you the model will be weak rather than shipping a confident-looking one.
Bias is your risk too
A model trained on your history learns your history, including any patterns in it you would not defend. That is worth examining deliberately.