Mortgage Insurance

Every ML engineer at a mortgage insurer knows the dataset hunt. You need to build a credit scoring model. There are 47 tables in Snowflake that look relevant. Three of them have “customer” in the name. One is certified, one is a staging copy, and one is a raw extract that nobody owns. The metadata catalog tells you the tables exist, but not which one to trust, whether it is current, or what is missing for your use case. You spend two days figuring out which data you can actually use before writing a single line of model code. Every underwriter knows the tab-switching. A mortgage application is on your desk. You need the applicant’s risk factor from Snowflake, the LTV coverage rules from the policy guide, and the pipeline status to see how many other applications are waiting. Three systems, three logins, three different interfaces, just to make one approval decision. Every customer support agent knows the hold time. A borrower calls asking about their coverage. You need to look up their policy, check their claim status, and confirm the cancellation rules, all while the customer waits. If the answer is not in the first document you open, you put them on hold and start searching. The Mortgage Insurance Persona AI solves all three. It connects Actian DI (Zeenea) for metadata intelligence, Snowflake for live structured data, and Google Drive for policy documentation, unified into a single persona-aware assistant accessible from the browser and Microsoft Teams.

​

The knows your role before you ask

A Persona in IdeaBoxAI is the AI’s understanding of who you are, what you need, and what your data looks like. When you log in as an ML Engineer, you get a data intelligence layer tuned to dataset discovery, certification status, and lineage. When you log in as an Underwriter, you get a decisioning layer tuned to application pipelines, risk factors, and policy rules. When you log in as a Customer Support Agent, you get a resolution layer tuned to coverage questions, claim status, and escalation summaries. The Mortgage Insurance Persona AI ships with three purpose-built personas. Each connects to the data sources relevant to their role. The persona you are assigned to determines which data you see, which skills are available, and which scenario cards appear on your home screen. The following table summarises the three personas. Each persona also comes with pre-configured scenario cards on the home screen. Click any card to pre-fill the chat input with a common workflow prompt. You can also type your request directly in plain language; the Persona AI auto-detects the right skill. To browse all available skills, type / in the chat bar to open the skill picker.

​

ML Engineer scenario: dataset discovery for model building

Anika Patel · Senior ML Engineer Credit risk team · Building a new credit scoring model · Needs certified training data Anika has been asked to build a new credit scoring model. Before she can start, she needs to find which datasets are available, which ones are certified for production use, and whether any required features are missing. Before IdeaBoxAI, this meant opening the Atacama catalog, searching for relevant tables, then cross-referencing Snowflake to check freshness and row counts, a process that usually took 1–2 days. Anika types: “Which datasets are certified and trusted for building a credit scoring model?” The Persona AI responds:

Certified dataset discovery, credit scoring 47 tables in Snowflake match credit-related metadata · 8 are certified for production use · 3 are directly relevant to credit scoring Certified datasets for credit scoring

Table	Schema	Certification	Owner	Last updated	Row count
CREDIT_SCORE_FEATURES	PROD.RISK	✅ Certified	Risk Analytics Team	Apr 22, 2026 (daily refresh)	2.4M
LOAN_APPLICATION_HISTORY	PROD.UNDERWRITING	✅ Certified	Underwriting Ops	Apr 22, 2026 (daily refresh)	8.1M
BORROWER_PROFILE	PROD.CUSTOMER	✅ Certified	Data Engineering	Apr 21, 2026 (daily refresh)	3.2M

Why these three:

• CREDIT_SCORE_FEATURES contains the core feature set used in the current production model, FICO scores, DTI ratios, employment history, and 14 derived risk indicators.
• LOAN_APPLICATION_HISTORY provides outcome labels (approved, denied, defaulted) across 8.1M historical applications. Essential for supervised training.
• BORROWER_PROFILE links applicant demographics and LTV ratios to each application record.

Potentially relevant but not certified

Table	Schema	Issue	Recommendation
RAW_CREDIT_BUREAU	STAGING.RISK	Not certified, raw extract from credit bureau feed	Do not use for model training. This is the upstream source for CREDIT_SCORE_FEATURES. Use the certified downstream table instead.
CUSTOMER_PAYMENT_HISTORY	PROD.SERVICING	Certified for servicing, not for credit scoring	Contains 36 months of payment behaviour. Relevant for churn prediction but not labelled for credit scoring. Request cross-certification from the Risk Analytics team if you need it.

Data lineage note: CREDIT_SCORE_FEATURES is derived from RAW_CREDIT_BUREAU through 3 transformation stages. The most recent pipeline run completed at 2:14 am today. No quality alerts in the last 30 days. Suggested follow-ups: Show me the full column list for CREDIT_SCORE_FEATURES · What features are missing for a churn prediction model · Trace the lineage for BORROWER_PROFILE · Run a sample query on LOAN_APPLICATION_HISTORY

Anika follows up with a question that used to require opening Slack and asking the data engineering team. Anika types: “Is BANK_CUSTOMER the authoritative table or should I use RAW_CUSTOMER or STAGING_CUSTOMER?” The Persona AI responds:

Table disambiguation, CUSTOMER tables 3 tables match “customer” · 1 is authoritative

Table	Schema	Status	Refresh	Row count	Verdict
BANK_CUSTOMER	PROD.CUSTOMER	✅ Certified · Authoritative	Daily, 3:00am	3.2M	Use this one. This is the certified production table. Owned by Data Engineering.
STAGING_CUSTOMER	STAGING.CUSTOMER	⚠️ Not certified · Staging copy	Hourly	3.2M	Do not use. This is the staging table that feeds BANK_CUSTOMER. Data may include incomplete records from the current day’s ingest.
RAW_CUSTOMER	RAW.CUSTOMER	❌ Not certified · Raw extract	On ingest	4.1M	Do not use. Raw extract from the source system. Contains 900K records that failed validation and were filtered out in the staging-to-prod pipeline.

Key difference: RAW_CUSTOMER has 4.1M rows vs BANK_CUSTOMER’s 3.2M. The 900K gap is intentional, those records failed data quality checks (missing SSN, invalid state codes, or duplicate applicant IDs). If your model trains on RAW_CUSTOMER, it will include invalid records that the production pipeline already rejected. Lineage: Source system → RAW_CUSTOMER → STAGING_CUSTOMER → BANK_CUSTOMER. The BANK_CUSTOMER table is the end of the validated pipeline.

The results:

• Certified datasets identified in seconds, with ownership, freshness, and row counts from the live metadata catalog.
• Table disambiguation resolved without asking the data engineering team on Slack. The 900K record gap between raw and production tables explained with specific validation reasons.
• 1–2 days of dataset hunting reduced to a 3-minute conversation.

​

Underwriter scenario: application review and eligibility

James Wu · Senior Underwriter Reviews 15–20 mortgage applications per day · Pipeline review with manager at 11 am James reviews mortgage applications for approval. Each decision requires checking the applicant’s risk profile in Snowflake, confirming eligibility against policy rules in the coverage guide, and tracking how long the application has been in review. Before IdeaBoxAI, this meant three browser tabs and a spreadsheet to track aging. James types: “Show me all pending mortgage applications waiting for my approval.” The Persona AI responds:

Application pipeline, James Wu 18 applications pending · 3 flagged for attention 🔴 Priority, aging or high risk

Application	Applicant	Loan amount	LTV	Credit score	Risk factor	Days in review	Flag
APP-28471	Martinez, Elena	$340,000	92%	618	14.2	8 days	Aging, exceeds 5-day SLA. Risk factor above threshold.
APP-28455	Chen, Robert	$520,000	88%	642	12.8	6 days	Aging, exceeds 5-day SLA.
APP-28502	Johnson, Tanya	$275,000	95%	580	16.1	2 days	Highest risk factor in queue. Requires senior review per policy section 4.2.

🟡 Standard review, 12 applications All within SLA. Risk factors between 4.0 and 10.0. Average days in review: 2.4.

Risk range	Count	Average loan amount	Average LTV
4.0–6.0 (low risk)	5	$310,000	78%
6.1–8.0 (moderate)	4	$385,000	84%
8.1–10.0 (elevated)	3	$420,000	89%

🟢 Ready for fast-track, 3 applications All three have risk factors below 4.0, LTV under 80%, and credit scores above 740. These qualify for automated approval under policy section 2.1. Suggested follow-ups: Show me the full details for APP-28471 · Check eligibility for APP-28502 · How many applications this week have a risk factor above 12 · Show me my pipeline aging trend over the last 30 days

James clicks into the highest-risk application. James types: “I have an applicant with a 580 credit score and 10% down, are they eligible?” The Persona AI responds:

Eligibility decision, 580 credit score, 10% down payment

Criteria	Applicant	Policy requirement	Status
Credit score	580	Minimum 620 for standard approval	❌ Below threshold
Down payment	10% (LTV 90%)	Minimum 5% required	✅ Meets requirement
LTV ratio	90%	Maximum 97% for insured loans	✅ Within limit
Debt-to-income	Not provided	Maximum 45% per policy section 3.4	⚠️ Needs verification

Eligibility verdict: Not eligible for standard approval. The 580 credit score is 40 points below the minimum threshold of 620 defined in policy section 3.1 (Minimum Credit Requirements). This is a hard floor, it cannot be waived at the underwriter level. Alternative paths:

• If the applicant has compensating factors (12+ months of reserves, no derogatory marks in 24 months), the application may qualify for exception review under policy section 3.1.3. Exception reviews require VP-level sign-off.
• If the credit score is within 90 days of a rescore event (dispute resolution, paid collection), recommend the applicant reapply after the rescore. The current application should be placed on hold, not denied.

Policy source: Coverage Guide, Section 3.1, Minimum Credit Requirements. Last updated: January 2026. Suggested follow-ups: Show me the full policy on exception reviews · Check if this applicant has compensating factors · What is the maximum LTV ratio for a 620 credit score · Pull the DTI requirements for this loan amount

The results:

• 18 applications surfaced with risk scores, aging status, and SLA flags, no spreadsheet tracking needed.
• Eligibility decision returned in seconds with the specific policy section, threshold, and alternative paths.
• 3 applications auto-identified for fast-track approval, freeing underwriter time for the complex cases.
• Pipeline aging visible at a glance, with applications exceeding the 5-day SLA flagged automatically.

​

Underwriter scenario: risk filtering before pipeline review

James Wu · Senior Underwriter Pipeline review with manager in 30 minutes · Needs to know which applications carry the highest risk James needs to walk into his pipeline review with a clear picture of risk concentration. His manager will ask how many high-risk applications are in the queue and what the team is doing about them. James types: “How many applications this week have a RISK_FACTOR above 12?” The Persona AI responds:

Risk filtering, RISK_FACTOR above 12 Week of April 21, 2026 · 74 applications received · 6 have a risk factor above 12

Application	Applicant	Risk factor	Credit score	LTV	Loan amount	Days in review	Assigned to
APP-28502	Johnson, Tanya	16.1	580	95%	$275,000	2	James Wu
APP-28471	Martinez, Elena	14.2	618	92%	$340,000	8	James Wu
APP-28519	Okafor, David	13.8	625	91%	$410,000	1	Sarah Lin
APP-28487	Patel, Neha	13.1	630	88%	$295,000	4	James Wu
APP-28510	Reeves, Mark	12.6	640	90%	$380,000	2	Sarah Lin
APP-28455	Chen, Robert	12.8	642	88%	$520,000	6	James Wu

Risk concentration analysis:

• 6 of 74 applications this week (8.1%) are above the 12.0 threshold. This is slightly above the trailing 4-week average of 6.8%.
• 4 of the 6 are assigned to James Wu. Consider redistributing APP-28487 or APP-28455 to balance the high-risk review load.
• The common pattern: all 6 have LTV above 88% and credit scores below 645. The risk is driven by thin equity combined with borderline credit, not a single outlier factor.

SLA status: 2 of the 6 (APP-28471 and APP-28455) have exceeded the 5-day review SLA. Both are assigned to James. Suggested follow-ups: Show me the risk factor trend over the last 4 weeks · Check eligibility for APP-28502 · Redistribute my high-risk applications · Compare this week’s risk profile to the same week last quarter

James walks into his review with the exact numbers: 6 high-risk applications, 8.1% concentration (above the 4-week average), two past SLA, and a clear pattern linking thin equity to borderline credit.

​

Support scenario: coverage question and claim lookup

Diana Reyes · Customer Support Agent Handles 40–50 borrower calls per day · Current call: borrower asking about coverage Diana is on a call with a borrower who wants to know what their mortgage insurance covers and whether they can cancel it early. Before IdeaBoxAI, she would put the borrower on hold, open the coverage guide PDF, search for the relevant section, and hope the search term matched. Diana types: “What is the coverage for a homebuyer who put 15% down?” The Persona AI responds:

Coverage lookup, 15% down payment A homebuyer who put 15% down has a loan-to-value (LTV) ratio of 85%. The following coverage applies.

Coverage attribute	Detail
LTV at origination	85%
MI coverage percentage	12% of the loan amount
Monthly premium (typical range)	0.30%–0.55% of the loan amount annually, paid monthly
Coverage type	Borrower-paid mortgage insurance (BPMI)

What the coverage means in practice: If the borrower defaults and the lender files a claim, mortgage insurance covers 12% of the original loan amount. On a $300,000 loan, that is$36,000 in coverage to the lender. Cancellation eligibility:

• The borrower can request cancellation when LTV reaches 80% (through principal paydown or home appreciation). This requires a current appraisal.
• Automatic cancellation occurs when LTV reaches 78% based on the original amortization schedule, per the Homeowners Protection Act.

Policy source: Coverage Guide, Section 2.3, Standard Coverage Percentages by LTV Band. Last updated: January 2026. Suggested follow-ups: At what LTV ratio does MI automatically cancel · Can a borrower cancel early with a new appraisal · What is the coverage for a 10% down payment · Look up a specific policy number

The borrower asks a follow-up. Diana does not need to put them on hold. Diana types: “Can a borrower cancel their mortgage insurance early and if so, how?” The Persona AI responds:

Cancellation rules, early cancellation A borrower can cancel mortgage insurance early through two paths. Path 1: Borrower-requested cancellation (LTV reaches 80%)

Requirement	Detail
LTV threshold	80% or below
How LTV is calculated	Based on current principal balance vs original property value, or a new appraised value if the borrower requests a reappraisal
Payment history	Must be current on payments. No 30-day late payments in the last 12 months. No 60-day late payments in the last 24 months.
Process	Borrower submits a written request to the servicer. Servicer orders appraisal if needed. Cancellation takes effect on the first day of the month following approval.

Path 2: Automatic cancellation (LTV reaches 78%)

Requirement	Detail
LTV threshold	78% based on the original amortisation schedule
No borrower action needed	The servicer is required to cancel MI automatically on the date the principal balance is scheduled to reach 78% of the original property value.
Legal basis	Homeowners Protection Act of 1998 (HPA)

Key distinction: Borrower-requested cancellation at 80% can use a new appraisal to account for home value appreciation. Automatic cancellation at 78% uses only the original amortization schedule; home appreciation is not factored in. What to tell the borrower: “You can request cancellation now if your balance is at or below 80% of your home’s current value. If you believe your home has appreciated, you can request a reappraisal through your servicer. Otherwise, your MI will automatically cancel when your scheduled balance reaches 78% of the original value, no action needed on your end.” Policy source: Coverage Guide, Section 6.1, Cancellation and Termination. Homeowners Protection Act of 1998. Suggested follow-ups: Look up a specific borrower’s current LTV · When will MI auto-cancel for a specific loan number · What is the reappraisal process · Explain the difference between BPMI and LPMI

Diana answers the borrower’s question in real time, no hold, no document search, no transfer to a senior agent. Diana types: “What is the current status of claim number 48291?” The Persona AI responds:

Claim status, #48291

Field	Detail
Claim number	48291
Policy number	PMI-2024-07-38291
Borrower	Williams, Karen
Loan amount	$285,000
Claim filed	March 14, 2026
Claim type	Default, 90+ days delinquent
Status	Under review: assigned to Claims Analyst Team B
Documents received	3 of 4 required (missing: updated property valuation)
Expected resolution	Within 15 business days of complete document submission

What is holding this claim: The updated property valuation has not been received. This document was requested from the servicer on March 18. It has been 26 business days with no submission. Recommended action: Contact the servicer to follow up on the outstanding property valuation. Once received, the claim moves to final review and should resolve within 15 business days. Suggested follow-ups: Draft a follow-up email to the servicer about the missing valuation · Show me all claims for this borrower · Summarise this claim for escalation to a senior agent

The results:

• Coverage question answered in seconds with the specific policy section, premium range, and cancellation rules, no PDF searching.
• Borrower-facing language provided alongside the technical answer so the agent can respond naturally.
• Claim status pulled with the specific blocker identified (missing document, 26 days outstanding) and the next action ready.
• Average handle time reduced by eliminating hold time, system switching, and manual document lookup.

​

Skills that power all three personas

Each persona in IdeaBoxAI is powered by a set of Skills, AI capabilities that define how the Persona AI handles specific types of requests. The Mortgage Insurance Persona AI ships with 15 skills across the three personas. You can trigger a skill in three ways:

• Click a scenario card on the home screen to pre-fill the prompt and run the skill immediately.
• Type your request in plain language. The Persona AI auto-detects the appropriate skill from your message.
• Type / in the chat bar to open the skill picker and select a skill by name.

​

Data and ML Engineer skills

The following skills handle dataset discovery, data quality validation, and analytical queries against the warehouse.

​

Underwriter skills

The following skills handle application reviews, risk assessment, and policy-based eligibility decisions.

​

Customer Support Agent skills

The following skills handle borrower-facing questions, claim lookups, and escalation preparation.

​

Connected to your data infrastructure

The Mortgage Insurance Persona AI sits on top of three data sources, unified into a single persona-aware assistant. It does not replace your existing infrastructure, it makes it accessible through natural language. When an ML engineer asks “which datasets are certified for credit scoring?”, the Persona AI queries the Actian DI (Zeenea) catalog for certification status and lineage, then cross-references Snowflake for freshness and row counts. The answer arrives in seconds, with the specific tables, their owners, and the reasons to use or avoid each one. Each persona only sees the data relevant to their role. An ML engineer queries the metadata catalog and Snowflake. An underwriter queries Snowflake application data and Google Drive policy documents. A support agent queries Snowflake claim records and Google Drive coverage guides. The platform connects through the following layers:

• Actian DI (Zeenea) connected via API, loads metadata context, data lineage, and certification status from the enterprise catalog.
• Snowflake connected as a structured knowledge base, NL-to-SQL queries run directly against the warehouse for live application, claim, and risk data.
• Google Drive connected via RAG pipeline, retrieval-augmented Q&A over policy documents, coverage guides, and underwriting manuals.
• IdeaBoxAI browser app for all personas, dedicated web interface with full conversational and dashboard capabilities.
• Microsoft Teams bot integration, Teams-native queries for users who work inside Teams throughout the day.

​

Getting started

Setting up the Mortgage Insurance Persona AI for your team takes less than a day.

​

Next steps

Explore these guides to learn more about the platform capabilities behind the Mortgage Insurance Persona AI.