A fast-growing national construction lender needed to speed up how it approves and funds loans, while keeping the strict controls that protect it from bad bets.
The lender originates about 250 loans a year. Its origination process relied on experienced staff performing detailed reviews across multiple systems and documents, ensuring accuracy while becoming increasingly resource-intensive as loan volumes grew. For a growing business built on risk discipline, the leaders envisioned the next phase of growth as including an automated, auditable system of record.
The lender partnered with RapidCanvas to build origination as a single, governed, AI-assisted workflow: grounded document extraction, automated progressive validation, compliance gates and a complete audit trail, with people kept firmly on every material decision.
The Challenge
The construction-lending division originated roughly 250 loans a year through a highly structured, expert-led process that relied on manual review across each stage. As origination volume grew, the lender saw an opportunity to bring that same discipline into a modern, automated system.
Scaling growing loan volume. The manual, expert-led process held quality high but was hard to scale as volume climbed.
Re-keying across disconnected systems. Information was captured across multiple systems to support different stages of the lending process, creating opportunities to streamline data flow as the business expanded.
Quality control built around a structured review process. Cross-checks ran off a checklist, by hand, confirming that addresses, names, amounts and square footage agreed across documents. The team wanted to more than double the number of items, but only automation made that feasible.
Detailed legal review performed by counsel. Counsel read every title commitment line by line, opening each hyperlinked exception, checking easements against the survey and confirming entity and signer consistency, which created an extra step on every loan.
Limited access to portfolio-wide benchmarking. There was no practical way to ask “is this budget normal?” against comparable past loans. Reviewing each loan's line items at that depth was too time-intensive to be actionable.
Moving toward a unified system of record. Pipeline status stayed with certain individuals and a shared folder structure, which was harder to scale as the business grew. The need was for a formal, automated audit trail.
A Business Too Complex for Off-the-Shelf Tools
The lender's construction-lending business spans multiple loan types, over 35 document types and a nine-stage approval process, complexity that reflects the breadth of the loans it originates. That complexity meant any system had to earn the committee's trust: extraction had to be accurate and every decision had to remain governed and auditable.
The Solution
Together, the lender and RapidCanvas built one governed workflow that runs end to end: AI does the legwork at intake, validation, document generation and sync; the lender's expert team members own the material decisions, the committee vote, the title signing and the funding wire, with that boundary explicit at every step.
Step 1 - Intake and grounded AI extraction: A loan officer drops documents into the file. A fast AI model classifies each document type on upload, then extracts the fields specific to that type across 30+ per-document-type schemas, from appraised value and flood zone to appraiser license status and signer identity. The same call returns three things: structured fields, three to five freeform underwriting observations, and risk flags generated against the loan's own terms. Three grounding mechanisms keep it honest:
- A wrong-document heuristic. If a document doesn't yield enough of the fields it should, it's flagged for human review rather than saved automatically.
- Reference triangulation. When a field points elsewhere (for example, “See Exhibit A”), the system follows the pointer to resolve the real value.
- A structured legal review for title commitments.
Across all of this, the system only saves values it actually extracted, and shows a confidence score for each one.
Step 2 - A staged workflow with hard checkpoint gates: The loan advances through a defined lifecycle, from intake through validation, approval, closing and funding, with hard checkpoints it cannot cross until key preconditions are met, such as a verified closing package and confirmed downstream sync. Committee approval is a structured, role-gated step: votes are recorded per voter, the decision is computed from them, and every action writes to an append-only audit trail.
Step 3 - Progressive validation, checks streamed live: When a loan enters validation, a series of checks stream back in real time, covering builder financials, cross-document consistency, budget versus peers and overall readiness, so the officer sees each step resolve instead of waiting on a single review. Cross-document consistency uses fuzzy field matching to handle harmless variations, rather than flagging every small inconsistency as an error.
Step 4 - Closing, the Credit Approval Form and downstream sync: After the committee approves, the system auto-generates the Credit Approval Form, the document that previously took an account manager 20-plus minutes to build by hand. Documents flow to and from the title company with signing and recording tracked in the workflow, and at closing the loan's structured data is pushed into the draw-management system automatically, eliminating the final step. Officers kept asking pipeline questions on the fly, so an Ask-AI assistant was built for natural-language queries against the live loan database. Leadership quickly put it to use as a loan-committee tool, able to answer any loan's status mid-meeting.
What Made It Work, Beyond What Was Built
Executive mandate. The lender's leadership backed a full switch from day one. The friction that remained was integration timelines and terminology edge cases, solvable engineering problems, not trust problems.
Dual-track adoption. The lender ran real loans through the new system in parallel with its legacy process, giving its team a safe way to pinpoint exactly where extraction or rules needed tuning, with zero production risk to live deals.
A tight, in-app feedback loop. The lender's own power user marked up real documents in-app (thumbs-up/down and comment controls), feeding a tight feedback loop that took the system from rough to production-trustworthy in eight weeks.
Bounded human-in-the-loop by design. AI augments the decision but never makes the loan-approval call; committee voting, title signing and the funding wire remain human, with role-based access enforced at the API level, not just the UI.
Results and Benefits
Targeting a 70–80% cut in per-loan manual review. Routine extraction, cross-checking and quality-control activities are now automated, allowing experts to focus on higher-value review, with a target of cutting per-loan manual review effort by 70–80%.
QC checklist scaled from 20 to 50+ items. Automation makes 50-plus quality-control items feasible where manual review had capped the team at 20.
It caught a real discrepancy in live use. The system surfaced a genuine square-footage mismatch between the appraisal and the construction plans in live use, the kind of error that is critical to surface early.
Redundant data entry largely eliminated. Data now flows across downstream systems from a single structured source, reducing repetitive data entry; the Credit Approval Form is auto-generated and post-closing data syncs automatically.
Governance by design, and a compliance-ready audit trail. Leadership and the loan committee have role-gated touchpoints on every loan, a loan cannot advance until each gate's preconditions are met, and every action, vote and document writes to an append-only, compliance-ready audit trail.
AI legal review against a 9-point checklist. Title-commitment review runs against a nine-point checklist, so counsel confirms rather than discovers, replacing line-by-line manual reading.
Historical budget benchmarking. Budgets are benchmarked against comparable past loans, normalized per square foot, with outliers flagged, context the lender never had before.
Conclusion
By encoding its own rules as enforced constraints and keeping its people firmly on every material decision, the lender built a workflow it could stake its lending operations on, with RapidCanvas as its build partner. The goal was not simply to reduce manual effort, but to give the lender a scalable operating model that could support continued growth without compromising its underwriting discipline.
