Industry · Insurance
A policyholder-event transformer pre-trained on millions of quote-to-claim paths, fused with a document-grounded encoder for unstructured claim flow. Downstream heads fine-tune per carrier, line, and regulatory framework — and agents act on each score across acquisition, retention, and claims operations.
01 · SharpAI
Predictive models
11 models · foundational + downstream
02 · AUTONOMY
Agents
20 agent patterns in this vertical
03 · FDE
Engineering services
3 anonymized engagements
01 · SharpAI · Insurance
A sequence-transformer over policyholder life events (quote, bind, endorsement, claim, payment, renewal) plus a document-grounded encoder for unstructured claim data. Both pre-trained self-supervised; downstream heads attach to the embeddings and fine-tune per carrier, line, and regulatory framework.
Foundation · what the backbone tokenizes
Millions of quote-to-claim paths across personal and corporate lines. Each dot is a raw signal the Policyholder Event Transformer tokenizes — hover or tap to see what it is.
Downstream · what the head consumes
The renewal-retention head reads the frozen PET embedding plus a focused set of decision-time features. Hover any node — these inputs drive the retention call and the offer.
Foundation models · self-supervised pre-training, frozen backbone reused downstream
Tokenizes the policyholder lifecycle — from quote to bind, endorsement, claim, payment, renewal, cancellation — into a single event sequence the encoder reads. The output embedding feeds conversion, retention, and household cross-sell heads.
Transformer encoder · multi-head attention over policy event tuples · 256-dim embedding · line-of-business aware embeddings
Pre-training objectives
Corpus · Millions of policyholder paths across personal + corporate lines · multi-carrier · multi-year history including renewal and cancellation events
Document-grounded encoder for unstructured claim flow — narratives, medical records, photos, adjuster notes. Built on a domain-tuned text foundation; fine-tuned with structured fields jointly attended over the PET embedding.
Domain-tuned text encoder · joint attention with PET embedding · multimodal expansion for photos via vision encoder
Pre-training objectives
Corpus · Hundreds of thousands of claim narratives + structured records · multi-carrier corpora respecting confidentiality and PHI boundaries
Downstream heads · fine-tuned on the frozen foundation embedding per institution and use case
Conversion propensity tuned to personal auto, home, life, and health lines.
Fine-tuned head over PET embedding · line-conditioned
Lead scoring for corporate / commercial lines — segmented by firm size and sector.
Hierarchical head · firm-level pooling
Multi-product propensity across the household — auto + home + life + ancillary.
Two-tower retrieval over household-pooled embedding
Predicts non-renewal in personal lines; surfaces the retention conversation before the lapse window.
Multi-horizon survival head
Predicts cancellation in pension and life portfolios months in advance.
Survival head · long-horizon calibration
Pre-decision risk surface for underwriters with explainable factors and comparable cases.
Fine-tuned head + retrieval of nearest-neighbor cases
Tier-1 claims classification — fast-track, standard, complex, suspected fraud.
Multi-class head over CDE + PET attention
Anomaly head over the claim foundation — flags suspect patterns for SIU review.
Reconstruction anomaly + contrastive divergence signal
Surfaces broker / agency conversion and retention performance vs. peer cohort.
Aggregated head over broker-pooled embeddings
+98%
Prospect-to-customer conversion lift · top 10 global insurer
7 of 10
Retirement closures predicted 90+ days in advance · top 5 BR insurer
Pre-action
Suitability and policy compliance built into the score
Multi-line
Personal + corporate portfolio coverage
02 · AUTONOMY · Insurance
Acquisition, retention, claims, and back-office — each agent below pairs a predictive score with an executed action, and reports back with a measurable line in the carrier's P&L.
20 agent patterns · ordered by deployment maturity
Scores prospects on conversion likelihood; routes the highest-yield to outbound; surfaces tailored channel and message.
Trigger · Lead inbound
↳ +98% conversion lift · top 10 global insurer
Detects household cross-sell opportunities (auto + home + life + ancillary); drafts the offer through the broker or direct channel.
Trigger · Policy event or annual review
Predicts cancellation in pension and life portfolios; opens pre-emptive retention plays months before the cancellation window.
Trigger · Cancellation score threshold
↳ 7 of 10 retirement closures predicted 90+ days advance · top 5 BR insurer
Detects non-renewal risk on personal-line policies and routes the right retention offer (discount, coverage adjustment) through the right channel.
Trigger · Renewal window + score
Tier-1 claims processing — reads documents, queries policy data, recommends fast-track / standard / complex / SIU.
Trigger · FNOL (first notice of loss)
↳ Back-office throughput
Pre-decision risk surface for underwriters — explainable factors, comparable cases, recommended action.
Trigger · Underwriting submission
↳ Faster, better-grounded decisions
Flags suspect claim patterns and routes to SIU with case rationale and supporting evidence.
Trigger · Anomaly pattern + threshold
Detects subrogation opportunities from claim narratives and structured data; routes to the recovery team with case file.
Trigger · Claim closure
Reads policy applications, claim documents, medical records, and contracts; extracts structured fields and posts back to core systems.
Trigger · Document arrival
Tier-1 conversational resolution across SMS, email, WhatsApp, voice — coverage questions, payment plans, ID cards — escalates on judgment.
Trigger · Inbound contact
Predicts up-sell from mandatory to voluntary retirement / insurance plans; routes targeted communication to existing plan holders.
Trigger · Plan event or annual window
Surfaces under-performing or under-covered brokers/agencies; drafts coaching content and identifies hand-offs.
Trigger · Weekly performance window
Schedules personalized outbound contact across SMS, email, voice — frequency-capped per household and per offer.
Trigger · Score + campaign
Detects abandoned quotes and re-engages the prospect with a tailored offer through the channel they used.
Trigger · Quote abandonment + decay window
Detects life events from data signal (move, marriage, new vehicle) and surfaces endorsement opportunities.
Trigger · Life-event signal
Pre-checks every customer-facing recommendation against suitability, anti-discrimination, and regulatory rules before it reaches a channel.
Trigger · Pre-action gate
↳ Pre-action compliance
Reads inbound complaints, classifies by severity and regulatory risk, drafts first response, and routes critical cases to humans.
Trigger · Complaint inbound
Reads payroll, sales, and policy data to recommend premium-audit adjustments on commercial lines.
Trigger · Audit cycle
Activates during cat events — surfaces affected policies, drafts proactive outreach, prioritizes adjuster routing.
Trigger · Cat event declared
Coordinates triage + document + fraud + subrogation + customer-resolution agents on a single complex claim — one audit trail.
Trigger · Complex claim flag
↳ Coordinated claim resolution
+98%
Conversion lift · top 10 global insurer
7 of 10
Retirement closures predicted 90+ days advance
Pre-emptive
Retention plays open before the cancellation window
End-to-end
Acquisition → claims → recovery coverage
03 · FDE · Insurance
Senior engineers embedded with the carrier's acquisition, underwriting, claims, and platform teams — integrating with core policy systems, claim platforms, and broker portals, against the regulatory framework the carrier operates under.
Read-only access into policy admin (Guidewire / Duck Creek / proprietary) and claim platforms — via API, secure file transfer, or direct database — with schema-mapping automation.
Reads policy applications, medical records, claim narratives, photos, and contracts; extracts structured fields and posts back into core systems with confidence and lineage.
Models and agents pre-check against the carrier's suitability framework, anti-discrimination policies, and regional regulator expectations — explainability per decision logged.
Insight lands inside the broker portal or agency CRM — propensity, retention risk, cross-sell opportunities — without a separate dashboard to log into.
Cat-event response patterns — affected-portfolio detection, proactive outreach, adjuster routing — built into the deployment, not improvised when the event happens.
Every score, decision, and resulting action captured with lineage — supports supervisory exams, internal audit, and customer disputes.
Anonymized engagements · same engineering team end-to-end
Top 10 global insurer
Personal lines · global
Prospect-to-customer conversion modeling, lead scoring, and outbound orchestration. Engineering team embedded with the insurer's digital acquisition and broker channels.
+98% prospect-to-customer conversion lift versus prior approach.
Top 5 Brazilian insurer
Pension + life lines
Cancellation prediction across pension and life portfolios with pre-emptive retention agents through advisor and direct channels.
7 of 10 retirement closures predicted 90+ days in advance.
Top US financial services + insurance
Multi-line carrier with retirement business
Up-sell from mandatory to voluntary retirement plans with targeted communication to existing plan holders, coordinated with claims and underwriting downstream.
$11.17B AUM up- and cross-sell opportunity identified.
Team shape — typical FDE composition for this vertical
Embedded senior engineers, predictive models in production, and agents that execute on every score.
