Path A
Hire more processors
The math didn’t work: every hire needed 60 days of onboarding, and growth outpaced their ability to recruit and train.
All work
3 brokers now do the work of 8.
A Canadian mortgage brokerage was losing deals because doc chasing and prep took 8+ hours per deal, sometimes reaching 12 hours. We rebuilt their intake (pulling docs from email threads, parsing them into the LOS, scoring discovery calls, writing follow-up messages, and retrieving docs) and put hours per file back into their brokers' hands.
Brokers, the most expensive seat in the building, were doing processor-grade work.
Lead flow was healthy. The bottleneck was downstream: every approved file took ~8 hours of human time to prep for underwriting: chasing W-2s, paystubs, statements, and tax returns through email threads, classifying each document, validating it against the LOS schema, and flagging what was missing.
The cost wasn’t just the eight hours. The CEO had hired through the bottleneck twice; each new hire bought three months of relief, then volume caught up. By the time we walked in, they were quietly turning away referrals.
We scoped three paths. We took the hard one.
Path B
Off-the-shelf OCR tool
Already tried it. OCR was passable on clean docs but failed on the 30% of borrower scans that were phone photos, partial pages, or handwritten annotations. The vendor’s fix was 18 months out.
Path CChosen
Custom AI intake + call analysis
Higher upfront cost, but the LLM approach handled the messy documents that broke OCR-only systems, and scored discovery calls before file prep even started.
We took Path C, not because we’d shipped it before, but because the unit economics worked even at conservative success-rate assumptions, and the brokerage was willing to run a phase-gated build with explicit stop-decision points.
Document intake + discovery-call analysis.
Document intake engine
Watches a shared inbox. On submission it pulls attachments, classifies each (W-2, paystub, statement, tax return, ID), runs a hybrid OCR + LLM pipeline to extract structured fields, validates against the LOS schema, and either writes verified data straight into the LOS or flags ambiguous fields. Processors only see the ~8% of docs the system isn’t sure about.
Call-analysis layer
After each discovery call, an LLM pipeline scores it against pre-qualification criteria, writes a 4-line summary to the CRM, sends a full-scale follow-up to the client with a summary, and surfaces a fit-score before a processor ever touches the file. Deals that would DQ later in underwriting get caught in the first 24 hours.
Six months post-launch, the numbers held.
−95%
File prep per loan (8 hrs → 25 min)
27
Originations per LO / month (was 10)
13
Days to collect all documents (was 21)
8%
Under the budget cap
Every number moved in the same direction.
| Metric | Before | After |
|---|---|---|
| File prep time per loan | ~8 hours | ~25 minutes |
| Brokers doing processor work | 6 of 8 | 1 of 8 (overflow only) |
| Files DQ'd in underwriting | 22% | 9% |
| Originations per LO / month | 10 | 27 |
| Days to collect all documents | 21 days | 13 days |
They told us not to do AI for one of our workflows. We were sceptical, but hired an outsourced resource instead of automation. That worked and saved us a lot. That’s how we knew we could trust them on the others.
Nick S. · CEO, Alltrust Mortgage
Built on your stack, not a black box.
- Models
- GPT-4 (doc classification & extraction), Claude (call summarization & scoring), local model for PII anonymization
- OCR
- AWS Textract + an LLM cleanup pass for low-confidence regions
- Orchestration
- n8n for workflow logic, custom Node.js services for LOS integration
- LOS integration
- Velocity
- Call recording
- CallCentral + Fireflies
- Vector store
- Postgres + pgvector for borrower-history retrieval
8 wks
Timeline · 2-wk prototype + 6-wk build
3
Senior team · AI, backend & delivery
$48K
Budget cap · billed weekly
$44.2K
Actual spend · 8% under the cap
4 hrs
Hand-off training · docs + runbook
The numbers can be yours
Want results like these? Let’s scope it.
Tell us where your team is doing processor-grade work. We’ll tell you whether AI moves the needle, or whether it doesn’t.
No deck · No demo · No sales pressure