TL;DR
Modern mortgage LOS architecture requires 7 specialized AI agents working in production: document processing, income calculation, underwriting, closing, compliance, voice AI, and post-closing QC. Each owns end-to-end work in its stage and acts on the loan record (not just retrieves data). Vendor marketing AI is mostly point integrations or chatbots; this guide names what to ask for and how to test it on demo day. Confer ships 9 production agents on this pattern.
What separates a real AI agent from a chatbot wrapper?
Three tests. Action — can the agent write to the loan record, generate documents, and submit to upstream systems? Or does it only retrieve information about the loan? Explainability — does every agent action have an immutable audit trail with operator, timestamp, inputs, and rationale? Or is the reasoning opaque? Determinism where it matters — is the math (income calc, tolerance buckets, ATR factors) reproducible loan to loan, or does it drift because an LLM is in the math path?
Use these three lenses to evaluate every claimed agent below.
The seven agents, in workflow order
Agent #1 — Document Processing Agent
What does a document processing agent actually do in a mortgage LOS?
Job
Classifies every uploaded document, extracts structured data, and routes the file to the correct E-Folder container. Replaces the 40–60 minutes of manual stare-and-compare per loan that legacy LOS forces on processors.
Required capabilities
- 3-tier classification: pattern matching → LLM → computer vision fallback
- Confidence threshold (90%+) with auto-escalation to human review below it
- Structured parsers per document type (W-2, pay stub, 1040, bank statement, driver's license)
- Encompass E-Folder container mapping (U1–U15) for hybrid deployments
How to test on demo day
Bring 30 representative loans on demo day. Ask the vendor's agent to classify them in real time. Measure: per-document time, confidence distribution, and how many require human review. Anything above 5 seconds per document or below 90% confidence on standard types is not production-ready.
Agent #2 — Income Calculation Agent
Why does income calculation need its own dedicated agent?
Job
Computes Fannie Mae 1084-compliant income across W-2, Schedule C, Schedule E, K-1, investment, retirement, and other income types. Applies add-backs deterministically. Runs 2-year trending. Flags variances over 10%.
Required capabilities
- Deterministic math — no LLM in the calculation path
- Per-income-type calculators with full Fannie 1084 add-back coverage
- 2-year trending with declining-income flag
- Reproducible outputs — same inputs produce identical results loan-to-loan
How to test on demo day
Send the vendor a Schedule C with depreciation, business use of home, and 50% meals deduction. Ask for the 1084 add-back calculation. If the answer is 'our LLM will figure it out' or 'we partner with an income vendor,' the agent is not native. Income/employment defects are the #1 critical defect category per ACES Q4 2024 — this matters.
Agent #3 — Underwriting Agent
What does an underwriting agent do that an AUS doesn't?
Job
Reads the loan record, runs the credit, income, asset, and DTI checks, generates conditions tied to specific deficiencies, and submits to DU/LPA via MISMO 3.4. The AUS gives a finding; the underwriting agent shapes the loan to get the finding.
Required capabilities
- Built-in AUS engine that mirrors DU/LPA logic for pre-submission validation
- Auto-condition generation — every finding maps to a tracked condition
- MISMO 3.4 export ready for submission without re-formatting
- Audit trail: every decision logged with rationale and inputs
How to test on demo day
Hand the agent a borderline DTI scenario. Ask: what would clear this to approve/eligible? A real underwriting agent will name the exact lever (asset reserves, gift letter, alternate income source) and show its work. A chatbot wrapper will say 'consult an underwriter.'
Agent #4 — Closing Agent
How does a closing agent enforce TRID timing without breaking?
Job
Generates the Closing Disclosure, runs the 3-business-day delivery clock, manages redisclosure triggers, and orchestrates the closing package. The hard part is durability — the timer must survive server restarts, deploys, and crashes.
Required capabilities
- Durable workflow engine (Confer uses Temporal) — timers survive restarts
- Atomic state transitions — clocks cannot be lost mid-deploy
- Tolerance bucket reconciliation between LE and CD
- eClose / hybrid close support with eNote and MERS eRegistry hooks
How to test on demo day
Ask the vendor: 'If your servers restart while a TRID 3-day timer is running, what happens?' If the answer involves 'cron job,' 'scheduled email,' or 'queue worker,' the timer is fragile. Durable workflows are a specific architecture; rule-based reminders are not the same thing. A late CD costs $5K–$25K per CFPB incident plus 3+ extra closing days.
Agent #5 — Compliance Agent
Why is a compliance agent different from a compliance module?
Job
Continuously verifies QM/ATR (8 factors per 12 CFR § 1026.43(c)), HOEPA high-cost detection, ECOA adverse-action timing, and HMDA LAR field completeness — at every state transition, not at the end. Catches defects before close, not after.
Required capabilities
- All 8 ATR factors verified per loan, not sampled
- HMDA 110+ LAR fields auto-populated from origination data
- ECOA 30-day decision timer with adverse-action notice generation
- Immutable audit log: operator, timestamp, before/after state on every change
How to test on demo day
Walk a loan through the platform with a deliberate ATR factor missing (e.g., 7-year employment history not yet collected). Does the system block close until it's resolved, or does it surface as a post-close QC finding? Block-before-close is workflow compliance. After-the-fact catch is module compliance. Asia Frost calls this 'enforcement as architecture' — that's the difference.
Agent #6 — Voice AI Agent (Kylie)
Does a voice AI agent actually help with mortgage origination?
Job
Handles inbound borrower status calls, walks through outstanding conditions, prompts document uploads, and transfers to a loan officer when needed. Also runs outbound reminder calls. Replaces 30–90 minutes of LO/processor time per loan.
Required capabilities
- Real-time speech recognition with mortgage vocabulary fluency
- Connection to live loan record — knows status, conditions, and next steps
- Outbound calling with consent + DNC compliance
- Smooth handoff to human LO with full call context
How to test on demo day
Call the vendor's voice AI yourself. Ask: 'What's outstanding on my loan?' If the agent says 'I'll have someone call you back,' it's an IVR with TTS, not a voice AI agent. Real voice agents pull live data and answer the question.
Agent #7 — Post-Closing Agent
What does a post-closing agent do that a QC team doesn't?
Job
Runs continuous QC on every loan rather than sampling 10% post-fund. Validates ULDD/MISMO export against investor delivery requirements, prepares HMDA filing, and runs reproducibility checks on income and underwriting decisions.
Required capabilities
- 100% loan QC, not 10% sampling
- Pre-delivery ULDD validation against Fannie/Freddie/private investor specs
- HMDA filing prep continuously, not as a year-end batch
- Reproducibility checks — same inputs to income calc must produce same outputs
How to test on demo day
Ask the vendor: 'What percentage of loans get full QC review on your platform — and is that pre-fund or post-fund?' Pre-fund 100% is what reduces repurchase risk. Post-fund sampling is a discovery mechanism, not a defense. ACES Q4 2024 industry critical defect rate is 1.79%; pre-fund continuous QC targets under 0.5%.
What about open protocol — should AI agents be exposed via MCP?
Yes, especially for lenders who want to avoid vendor lock-in. The Model Context Protocol (MCP) is the emerging standard for exposing tools and data to LLMs. A mortgage LOS that exposes its agents via MCP lets any LLM — Claude, GPT-5, Llama, or a purpose-built model — interact with the platform through a standard interface. Confer ships 32+ MCP tools across underwriting, document AI, and compliance domains so your IT team can connect downstream BI, customer service, and executive dashboards without per-vendor SDK work.