Back to case studies
Vertical SaaS

From document search to an auditable underwriting copilot

A 23-person insurance SaaS had four underwriters spending too much time searching submissions and policy manuals. We built a copilot that proposes a recommendation, cites the policy clauses, and keeps every accountable decision with a human.

Timeline:10 weeks · evidence model to production handover
Delivery:Production handover
AI integration dashboard showing workflow controls and business metrics

Project snapshot

The small-team operating context

The delivery and value signals below are anonymized and rounded so the operating scale is visible without exposing the client.

Client context

Insurance SaaS · 23 employees · 4 underwriters · ~40 submissions per week

Delivery team

2 engineers + 1 product lead with the underwriting manager and compliance lead

Sponsor

Underwriting Manager

Decision time

~60% reduction

Copilot-assisted underwriting path

Project-reported

Submission volume

~40 / week

Small-team operating baseline

Directional model

Final authority

100% human-owned

No autonomous accept or decline

Project-reported

Production signal

  • Underwriters reviewed recommendations with policy clauses and source records in context.
  • Complex medical, financial, and unusual risks continued to route to expert judgment.
  • Every accepted or rejected recommendation contributed to the evaluation and audit record.

What we learned

  • Start with the evidence and decision boundary, not the model prompt.
  • Reject unsupported citations before they reach the reviewer.
  • Measure extraction, retrieval, and recommendation quality separately.
Delivered: Auditable underwriting copilot · Evidence and reviewer workflow · Evaluation set, runbook, and audit handover

Actual process teardown

How the operating process changed

The important change is not a model in isolation. It is the full operating path around the model: inputs, orchestration, controls, review, and the system of record.

The challenge

Four underwriters were working through roughly 40 submissions a week. They searched long application packets, policy manuals, and internal references manually, then reconstructed the evidence trail after forming a recommendation. Faster decisions could not come at the cost of auditability or expert accountability.

The production response

We structured submissions and policy evidence, added clause-level retrieval and provenance, and built a copilot that proposes but never decides. Every recommendation passes through confidence checks and an underwriter review before a final decision is recorded.

Before · manual baseline

  1. 01

    Read the submission

    Underwriter

    Underwriters searched long application packets and supporting documents to find the fields relevant to the risk.

  2. 02

    Find applicable policy language

    Underwriter

    Relevant clauses, exclusions, and precedent were located across manuals and internal references.

  3. 03

    Form a recommendation

    Underwriter

    The underwriter synthesized evidence, missing information, and risk context from scratch.

  4. 04

    Document the decision

    Underwriter

    Reasoning and source references were recorded after the decision work, creating a second documentation pass.

After · AI-assisted production

  1. 01

    Extract and normalize

    AI + system

    Submission documents are parsed into typed fields with provenance and missing-data flags preserved.

  2. 02

    Retrieve policy evidence

    AI

    The copilot surfaces relevant clauses, exclusions, and supporting records with citations.

  3. 03

    Propose, do not decide

    AI

    The system produces a recommendation, confidence context, and unresolved questions for review.

  4. 04

    Approve and audit

    Human reviewer

    The underwriter makes the final decision and the system records evidence, rationale, and reviewer ownership.

Phase-by-phase delivery

What was delivered, in order

Each phase turns the process teardown into a concrete production deliverable, with a clear value created before the next layer is added.

  1. 01Delivery phase

    Baseline the submission journey

    Locate the time spent searching documents, finding clauses, synthesizing evidence, and recording rationale.

    Delivered

    • Submission journey map
    • Document and policy source inventory
    • Decision ownership and audit requirements

    Value created

    Speed is treated as an evidence-retrieval problem, not as permission to automate the accountable decision.

  2. 02Delivery phase

    Structure the evidence

    Make unstructured submissions and policy language available to the copilot with provenance intact.

    Delivered

    • Typed submission schema
    • Document extraction and validation
    • Policy retrieval with clause-level citations

    Value created

    Underwriters spend less time hunting for evidence and more time evaluating the risk itself.

  3. 03Delivery phase

    Build the recommendation loop

    Generate a recommendation that shows confidence, evidence, and unresolved gaps before review.

    Delivered

    • Recommendation drafting
    • Missing-data and confidence checks
    • Edge-case routing to an underwriter

    Value created

    The system accelerates preparation without turning a probabilistic model into an unreviewed decision-maker.

  4. 04Delivery phase

    Audit and improve in production

    Keep the final decision with the expert and use reviewed cases to improve the system safely.

    Delivered

    • Reviewer decision UI
    • Immutable evidence and rationale record
    • Evaluation set for extraction, retrieval, and recommendation quality

    Value created

    The copilot becomes more useful over time without weakening accountability or auditability.

How to read this teardown: the delivery sequence combines the existing public-safe project narrative with a practical view of the production work around it. It does not add client names, confidential implementation details, or unverified performance claims.

Measured and modeled impact

Separate the outcome from the value hypothesis

Project-reported outcomes are shown alongside directional models. Public benchmark figures are kept in their own evidence section below.

Underwriter decision time

Project-reported

~60% reported reduction

Before

Manual document and policy search

After

Copilot-assisted evidence review

Interpretation boundary

Anonymized and rounded project-reported outcome supplied for this case study; independently unverified in the public view.

Value model
Directional model

More decisions without removing judgment

For a small insurance SaaS, the return is more qualified submissions assessed by the same four underwriters while keeping risk judgment and audit ownership intact.

Model

Additional submissions assessed × conversion rate × expected premium or contribution value − incremental review cost

Assumptions to validate

  • The ~15-hour weekly capacity estimate is validated through time sampling.
  • Decision quality and loss ratio remain within agreed bounds.
  • Premium, contribution, conversion, and review-cost assumptions are validated by the underwriting business.

ROI scorecard

Value across time, money, resource, output, and error

Each teardown shows the practical ROI lens for the transformation. Existing project signals stay labeled; financial impact remains a transparent model until baseline volumes, costs, and quality data are verified.

Time
Project-reported

~60% reported reduction

Underwriter decision time is reported as down about 60% on the copilot-assisted path.

Money
Directional model

More assessed submissions

Directional model: additional qualified submissions assessed × conversion rate × expected premium or contribution value.

Resource
Directional model

~15 hrs/week potential

For four underwriters, less document search and evidence assembly can return roughly two working days per week to risk judgment.

Output
Directional model

More decisions per reviewer

A faster evidence loop can increase the share of the ~40 weekly submissions assessed without removing accountable review.

Error
Project-reported

Audit trail preserved

Citations, provenance, confidence context, and final reviewer ownership keep the decision traceable.

ROI should be finalized against the client baseline: loaded labor cost, transaction volume, unit economics, throughput, and measured error or denial rates. The directional model is a value pathway—not an invented client result.

System blueprint

The production system around the AI

The tool list shows what each component is responsible for across the application, data, AI, integration, orchestration, and governance layers.

Tools used

Application

TypeScript

Copilot workflow UI, reviewer states, and decision controls

AI + evaluation

Python

Document processing, extraction checks, and evaluation services

Model access

AWS Bedrock

Model access inside the approved cloud and security boundary

Data + audit

Postgres

Submission, evidence, recommendation, reviewer, and audit state

AI + processing

Document extraction layer

Typed fields, provenance, missing-data flags, and source references

Governance

Policy evidence retrieval

Clause-level retrieval, citations, confidence checks, and unresolved gaps

Orchestration

  1. 1Submission intake
  2. 2Document extraction
  3. 3Schema and provenance validation
  4. 4Policy retrieval
  5. 5Recommendation generation
  6. 6Human underwriting decision
  7. 7Immutable audit record

In-house data and boundary

The system is grounded in the client’s existing operational data rather than an unbounded external corpus.

Application and submission documentsPolicy manuals and clausesHistorical underwriting referencesStructured risk and customer records

Human in the loop

  1. 1Review every recommendation before a final decision.
  2. 2Handle complex medical, financial, or unusual risks.
  3. 3Reject unsupported citations, incomplete extraction, or missing evidence.
  4. 4Own the final accept, decline, or refer decision and audit rationale.
AI capabilities
Document extractionPolicy-grounded retrievalRecommendation draftingEvidence summarization

Public benchmark evidence

What the wider market is showing

These sources support the transformation pattern. They are not client-specific outcomes and should not be read as such.

Public benchmark

Swiss Re — AI at Swiss Re

Swiss Re describes AI for document summarization and information discovery while keeping expert oversight and accountability central to underwriting.

Why it matters here

Supports the copilot pattern: accelerate evidence handling, but do not automate the accountable underwriting decision.

External research provides context and benchmarks. It does not verify the anonymized client, implementation, or project-specific results described above.

Names, logos, and specific numbers are redacted in this public view. Full references are available under NDA on request.

Build the next one

Have a workflow that needs a closer look?

Start with the friction you already see. We will help you determine what is worth solving and how to move it safely into production.

Assess my workflow Talk to us