Case Study: Treaty Renewals in Days, Not Weeks (US Reinsurer)

The short answer: Treaty renewals slow down when teams manually reconcile multi-document renewal bundles (treaty wording, endorsements/addenda, I&Ls, schedules) across mixed-quality PDFs and scans. This reinsurer used Ultrassure to generate clause-level 'what changed' analysis with citations, standardize critical terms across renewals, and route ambiguity to review — cutting renewal cycles to days instead of weeks while improving auditability and consistency.

Key Takeaways

1. Documents processed: treaties, endorsements/addenda, I&Ls, schedules/exhibits; mix of native PDFs + scanned PDFs
2. Renewal cycle reduced from ~10–15 business days to ~2–4 business days (same scope, same team)
3. ~99% accuracy on critical structured fields after review pass; 'abstain + review' used for ambiguous clauses
4. Outputs: change map + cited summaries + structured exports with page/section citations

Results Snapshot

Accuracy definition: correctness of extracted structured terms with correct source citations, after a human review pass, with ambiguous cases explicitly flagged (not guessed).

What is insurance contract AI? Software that ingests policies, endorsements, and schedules; preserves document structure; reasons across multi-document stacks; and produces auditable answers with citations. This case study shows what happens when that approach is applied to treaty renewals. Learn why generic AI fails →

The Problem: Why Treaty Renewals Took Weeks

The reinsurer's renewal team was losing time to three repeat pain points:

1. Fragmented renewal bundles: Treaty wording, addenda/endorsements, I&Ls, and exhibits were scattered across mixed-quality PDFs (including scans).
2. Manual comparison burden: The team needed a reliable way to answer "What changed?" across versions (expiring vs renewal) without manual clause-by-clause reading.
3. Audit-ready outputs required: Outputs needed to survive internal scrutiny — structured term summaries and audit-ready evidence for counsel, underwriting leadership, and ops.

What documents were in scope (and what formats)?

This pilot focused on real renewal packages representative of their book.

In-scope document types

• Treaty wordings (quota share + XoL formats)
• Endorsements / addenda / amendments
• Interests & Liabilities (I&L) schedules
• Exhibits & reporting schedules
• Side letters where applicable (redacted)

File formats encountered

• Native PDFs (digitally generated)
• Scanned PDFs (mixed OCR quality)
• Schedule-heavy PDFs (multi-page tables)
• Slip-like summary pages embedded in PDF packs

Baseline Workflow (Before Ultrassure)

Their baseline process was typical for sophisticated reinsurers:

1. Triage & intake (1–2 days) Gather documents, confirm "complete set," de-duplicate versions
2. Manual comparison (4–8 days) Counsel/analysts compare expiring vs renewal, track changes in spreadsheets
3. Term summary drafting (2–4 days) Create internal renewal summary, circulate for review
4. Sign-off loop (2–4 days) Questions back to analysts; re-check source language; finalize approvals

Baseline cycle time: ~10–15 business days end-to-end (varied by complexity and number of endorsements)

New Workflow (After Ultrassure)

They implemented a repeatable "renewal bundle" workflow:

1. Upload & auto-organization (same day) Treaties + endorsements + I&Ls grouped into a renewal bundle with version labels (expiring vs renewal)
2. Automated extraction + citations (same day) Critical fields extracted into a standard schema; every field tied to source references
3. Clause diff generation (same day) Ultrassure produced a structured "what changed" report highlighting deltas by clause category
4. Review queue for edge cases (0.5–1 day) Ambiguous or scan-degraded sections routed for human confirmation rather than guessed
5. Export & sign-off (1–2 days) Counsel and underwriting sign-off using a single source of truth: diff + cited summaries + CSV

New cycle time: ~2–4 business days end-to-end (same renewal bundle scope)

What did they measure (methodology you can replicate)?

This case study is intentionally "evidence-forward." Here's the measurement approach the team used:

1) Renewal cycle time (before/after)

• Before: measured from "bundle complete" → "internal sign-off sent"
• After: same definition, same renewal team, same sign-off gate

Result: ~10–15 days → ~2–4 days

2) Accuracy on critical fields

They defined a critical-field checklist and labeled "ground truth" from the final signed set.

Critical fields tested (examples):

• Limits (per risk / per event / aggregate)
• Attachment points / retentions
• Key exclusions (named peril, cyber, war, etc.)
• Reporting requirements (timelines, data fields)
• Claims cooperation / control / notification
• Commutation clauses (presence + mechanics)

Scoring rules:

• Correct: extracted term matches ground truth + correct citation
• Needs review: model abstained or flagged ambiguity
• Incorrect: wrong value or wrong clause reference

Result: ~96–99% accuracy on critical structured fields after the review pass (with abstentions routed to human review rather than guessed)

Outcomes Beyond Speed

Speed was the headline, but three outcomes mattered just as much:

• Consistency: Standardized term summaries reduced "analyst style variance" between teams.
• Auditability: Every key term and conclusion could be traced back to exact source language.
• Renewal governance: Clause diffs created a durable record of what changed and why.

How Changes Were Flagged (Change Map)

Instead of long redlines, the team used a change map that categorized differences and prioritized review.

Change categories tracked

• Limits & structure — limits, attachments, aggregates, reinstatements
• Reporting & bordereaux — frequency, deadlines, required fields
• Claims control — notice windows, cooperation, consent
• Exclusions & carve-backs — new or modified exclusions
• Definitions & interpretation — terms that alter meaning elsewhere
• Conditions & warranties — obligations, remedies, time limits

Impact levels

• High: Likely to change coverage, economics, or obligations
• Medium: Operational or timing change requiring review
• Low: Clarification, cleanup, or formatting

Example: What a change map looks like

Critical Terms Checklist

Reinsurers should verify these fields are extracted accurately with correct citations:

Economics & structure

• Limits (per event / per risk / aggregate)
• Attachment points / retentions
• Reinstatements
• Participation / shares
• Term and sunset language

Operational obligations

• Reporting requirements
• Notice triggers and timelines
• Claims control / cooperation
• Audit rights

Coverage shaping

• Exclusions added, removed, or modified
• Carve-backs and sub-limits
• Key definitions affecting interpretation
• Conditions, warranties, and remedies

Buyer-grade rule: A term only counts as extracted if it includes both a normalized value and a citation to the exact source language.

What were the hardest edge cases (and how were they handled)?

No system should pretend every clause is easy. The reinsurer encountered three common edge cases:

1. Scan-degraded tables in I&Ls Handling: table-aware extraction + review queue for low-confidence cells
2. Manuscript endorsements with dense cross-references Handling: definition linking + cited evidence retrieval
3. "Same meaning, different wording" changes Handling: diff categorized as "semantic / potential impact" with human confirmation

The key is governance: abstain rather than guess and route ambiguous items to review.

What Other Reinsurers Should Copy

If you want similar results, replicate these three principles:

1. Treat renewals as bundles with version labels, not individual PDFs.
2. Score success on cycle time + citation quality + critical-field accuracy, not "demo accuracy."
3. Enforce "no-guessing" rules (abstain + review queues) for high-stakes clauses.

Planning a pilot? See the Buyer's Guide (2026) for evaluation scorecards, RFP questions, and practical pilot blueprints.