Skip to main content
Industries · Insurance

Autonomous AI discovery built for insurance.

Continuous AI discovery across your policy, claims, channel, and reserves data. Every finding ships with the underlying queries, the cohort definition, and the cost of the run. The cohort-level book questions that never make the actuarial deep-dive queue, answered before your next book review.

Built for Chief Underwriting Officers, Heads of Claims, Chief Actuaries, and Heads of Pricing at P&C, life, and health carriers.

CRITICALFrequency drift
Auto comp claim frequency rose 31% for the under-25 cohort over six weeks.
Agent flagged the shape week 3 of the trend. 8,400…
From a real DecisionBox run
Why DecisionBox

The cohort-level book questions that never make the actuarial deep-dive queue, running overnight against your data.

Your team owns the rate filings, the reserves, the regulatory posture. The agent runs unattended against your policy, claims, channel, and reserves data — and ships findings into a queue your team triages. Every claim carries the underlying queries and the data-validated number. The agent reads where your data lives; it never writes back, never bypasses RBAC, and logs every step for the rate-filing record.

See how the agent reads cohort frequency and severity
The cost of late signals

The combined ratio looks fine. The sub-line underneath is where the book actually moved.

Three benchmarks every underwriting, claims, and actuarial leader recognizes. Each points to the same gap: your data already holds the signals — frequency drift, channel-quality drag, sub-line economics — but they only surface when somebody asks the right question. Most weeks, nobody does.

Acquisition vs renewal
Cost to acquire a new policyholder vs renew an existing one. The renewal cycle is where the book's economics are made or lost — and the segment-level retention drift rarely surfaces before the renewal letter goes out.
Bain & Company, Customer Loyalty research
Claims leakage
10%
Estimated leakage rate across P&C claims spend. The gap lives in subrogation misses, fraud signals not flagged, and reserves set against the wrong exposure shape.
Accenture, Claims Operations Benchmark
Annual shopping
30%
Of personal-lines policyholders shop their coverage each year. The aggregate retention rate hides which segments are most at risk — and which are quietly already gone.
J.D. Power, Insurance Shopping Study
From real runs

Three findings the agent surfaced, paired with the recommendation it shipped.

Three shapes the book sees every quarter: a single-line frequency drift the aggregate concealed, a cross-system channel-quality pattern the underwriting score didn't catch, and a sub-line economics shift the combined ratio averaged away. Same auditable trail every time — severity, the underlying queries, the cohort definition, the cost of the run.

CRITICALFrequency drift
Auto comprehensive claim frequency rose 31% for the under-25 cohort over six weeks.
period: 6-week rolling, Q2 · under-25 cohort…
Cross-checked against policies.auto_in_force…
single-line frequency drift
P1Recommendation
Re-rate cohort + investigate competitive pricing
Re-rate the under-25 segment ahead of the next renewal cycle;
Expected impact
target: align under-25 pricing to actual frequency…
Target: Re-rate cohort +…2 concrete actions
CRITICALChannel quality
Aggregator-bound homeowners policies carry 2.1× the 18-month claim severity of direct-bound.
aggregator-bound in-force: 12,400 policies…
Cross-checked across policies.homeowners_in_force…
cross-system pattern
P1Recommendation
Add channel as a rating factor + renegotiate partner economics
Re-rate aggregator-bound homeowners business;
Expected impact
target: align homeowners pricing to actual loss shape…
Target: Add channel as a…3 concrete actions
HIGHSub-line masking
Combined ratio held at 96 aggregate.
aggregate combined ratio: 96 (target ≤97) · cyber…
Cross-checked across policies.in_force_by_subline…
aggregate hides the sub-line
P1Recommendation
Reprice cyber + review reinsurance + board-materials flag
Reprice cyber against actual severity development;
Expected impact
target: stop the GL cross-subsidy of cyber · effort…
Target: Reprice cyber +…3 concrete actions

Point the agent at your book of business for a week. Triage what it finds Monday morning.

Use cases

Five surfaces the team already owns, one continuous discovery.

Same engine, same data, five categories of work your team is already doing — or wants to do, if the queue ever clears. Pick the surface that matters this quarter.

The cohort whose frequency moved, and the aggregate that hid it.

Frequency and severity drift inside cohorts the dashboard reports in aggregate. The agent reads claims.fnol_daily, policies.in_force, and cohort definitions (age band, vehicle class, peril, geography) together — and surfaces the segments where frequency or severity is moving before the next reserve review catches it. The deep dive lives in how the agent decides what to investigate.

Sample analysis areas
  • Cohort-level frequency drift across age band, vehicle class, geography, and peril
  • Severity development by accident-year cohort
  • Catastrophe vs. attritional loss decomposition
  • Frequency masking inside aggregate-flat lines
  • Loss-cause clustering and emerging-risk detection
Why DecisionBox

Built for carriers who want an analyst running cohort math overnight, and a record the rate filing can rely on.

Three things every finding from DecisionBox carries. None of them changes how your team works in your policy admin system, your claims platform, or whatever data tools and BI you already standardized on.

Sees what nobody had time to investigate

The agent picks which cohorts, channels, sub-lines, and reserve segments to investigate from your policy, claims, and reserves data. Findings ship into a queue your team triages in the morning. The cohort-level frequency drift stops sitting in the data waiting for the next reserve cycle.

Every number you'd defend at the book review

The underlying queries the agent ran, the cohort definition, the data window, the cost of the run, the validation checks. Your team re-runs anything, extends anything, overrides anything. No black boxes, no numbers the actuarial review can't reproduce, no "trust the AI."

Reads your data where it lives. Filing-ready record.

Self-host on Docker, Helm, or Terraform — or managed cloud on your terms. The agent reads your data where it lives, never writes back, never bypasses RBAC, and logs every query and reasoning step for the rate filing, the actuarial review, and the audit. Snowflake, BigQuery, Databricks, Redshift, PostgreSQL. Open source, AGPL v3. Built for the regulator, the model validation review, and the data-residency posture you already enforce.

The next book review, with the sub-line shifts already mapped.