case Study | Insurance & Reinsurance
From Manual to Intelligent: Agentic AI Rewrites the Underwriting Playbook.
How our solution turned a 4,800-hour underwriting bottleneck into a five-minute decision engine, and what it signals for insurers and reinsurers.
< 5 min
per case, down from 30 minutes of manual review
90%+
reduction in manual review time
10x
capacity expansion on the same headcount
> 95%
decision consistency across all cases
The problem hiding in plain sight.
Underwriting is where an insurer's risk appetite meets reality. It is also where a surprising amount of institutional value quietly leaks away. At a leading insurance provider in South East Asia, every year meant roughly 9,600 cases, 2,700+ manual reviews, and 4,800 hours of skilled underwriter time spent doing what skilled underwriters least enjoy: cross-referencing manuals, reconciling occupational and territorial risk tables, and hunting through reinsurer guidance to justify a single decision.
​
A typical case took 30 minutes of manual review. Multiply that across thousands of files and the cost is not just time; it is inconsistency. Two underwriters, working the same guidelines under different pressures, can reach different conclusions. In a business built on pricing risk accurately and defending those prices to regulators and reinsurers, that variance is expensive.
​
This is the exact terrain where agentic AI earns its keep.
01 What "agentic" actually means here
Software that thinks, decides, and acts, not merely responds.
Most AI in insurance to date has been a chatbot bolted onto a service desk, or rules-based RPA that breaks the moment a case falls outside its script. Our solution is built on a different premise.
01
Process Agents.
Automate the backend work of routing, analysis, and decisioning, and adapt as conditions change rather than breaking on the first exception.
02
Conversation Agents.
A natural-language interface so a human can interrogate the system in plain English and get a context-aware answer, instantly.
03
Compounding value.
The process layer surfaces real-time risk insight; the conversational layer makes it instantly accessible to the underwriter. For this insurer, that architecture became the Underwriting Copilot.
Six agents, one decision
01
Case Routing agent
Triages each incoming file and routes it into the right analysis path
02-05
Four Risk Analysis agents
Work the case in parallel from distinct angles: general, medical, occupational, and territorial risk, each grounded in the relevant reference material
06
Underwriting Insight agent
Synthesizes the findings into a recommendation, with every conclusion cited back to its source document and section
02 The results
30 min → < 5 min
Case handling collapses from half an hour of manual review to under five minutes, end to end.
90%+ less review
Manual review time drops by more than 90%, reclaiming 4,800 underwriter hours a year.
10x capacity
A tenfold expansion in underwriting throughput on the same headcount.
> 95% consistency
The number that matters most to executives. Consistency is not a productivity metric; it is a risk-control and audit metric.
A reinsurance story not just an insurance one.
It is tempting to read this as a back-office efficiency win. The more interesting implications sit at the cedent-reinsurer boundary. Reinsurers price treaties on the assumption that a cedent applies its underwriting guidelines consistently. When decisions vary case to case, the reinsurer absorbs that uncertainty and prices for it.
​
A platform that enforces >95% decision consistency, with every recommendation traceable to the exact manual or guidance section it relied on, changes the quality of the conversation. The cedent can demonstrate, not merely assert, that reinsurer guidelines were applied as written on every file.
​
That traceability is also a lever for faster, cleaner treaty audits and facultative referrals: the audit trail already exists at the point of decision. For reinsurers, a book underwritten this way is easier to trust and cheaper to monitor, which over time should translate into better terms for the cedent. The same agentic pattern generalizes across the value chain: automated facultative submission review, portfolio-level risk-drift detection, and consistency checks against reinsurer appetite are natural next deployments.
03 Built for the compliance reality
04 The economics
None of this survives contact with a regulator or a Chief Risk Officer unless it is auditable and secure by design.
AUDIT TRAIL
Every recommendation traceable to source document and section
ACCESS CONTROL
Role-based access so only authorized personnel touch sensitive data
DATA SECURITY
Encrypted processing with zero external data exposure
FRAMEWORKS
Aligned with ISO 27001 and SOC 2
For a regulated life insurer, that combination of automation and provable governance is the difference between a pilot and a production system.
The engagement was structured as custom development plus annual support, giving the insurer full ownership of the configured solution rather than a per-seat subscription. The build was a one-time USD 120,000 investment delivered across a five-phase methodology (conceptualize, comprehend, compose, calibrate, converge), running from workflow discovery through prototype, knowledge-base and model construction, piloting and fine-tuning, to training and go-live support.
​
Ongoing, USD 60,000 per year covers what an agentic system actually needs to stay healthy: model performance monitoring and periodic retraining to prevent drift, platform and security maintenance, and knowledge-base scalability. Year 1 totals USD 180,000; USD 60,000 per annum thereafter.
Set that against 4,800 hours of reclaimed effort a year and a tenfold capacity increase, and the payback case makes itself, before counting the gains in consistency, auditability, and reinsurer confidence.
Agentic AI, applied to a well-defined, guideline-heavy process, does three things at once that legacy automation could not: it collapses cycle time, it enforces consistency, and it makes every decision defensible.
The manual era of underwriting was defined by how much a skilled human could hold in their head. The intelligent era is defined by how well humans and agents work together, and how much of that judgment can finally be made consistent, traceable, and fast.