How underwriting systems will evolve as lending scales
Table of contents
About the author
Utsav ShahAI and decision-systems operator with experience building large-scale systems at Uber and Cruise.
The U.S. Small Business Administration Office of Advocacy reports 36,207,130 small businesses in the United States as of February 2026, representing 99.9% of firms.
Most of them rely on equipment to run and expand, and acquiring that equipment usually requires access to financing.
ELFA's Q3 2026 outlook says equipment demand is on pace for a record and projects that new business volume among surveyed member companies should top $129 billion in 2026. For lenders pursuing that opportunity, operational capacity becomes part of the growth constraint.
Equipment loans tend to be higher in volume and tighter in margin. To serve this segment profitably, financial institutions must process applications efficiently while radically reducing the time spent per deal.
However, systems have evolved, but not enough to meet this demand profitably. Underwriting still involves manual and semi-manual work that doesn't directly assess risk.
Automated pipelines improved data retrieval, but underwriting isn't linear. Exceptions are routine. Static workflows struggle under real-world variability.
Scaling lending in 2026 and beyond requires moving beyond static pipelines toward a controlled preparation layer that coordinates specialized tasks while preserving human review.
In this blog, I'll explain how underwriting systems will evolve as lending scales and why institutions that adopt this architecture will grow without a proportional increase in operational scale.
Scaling Volume Will Expose Structural Limits
As lending volume increases, the friction that once felt manageable becomes visible.
In equipment finance, a single deal can include multiple files spanning dozens, sometimes hundreds, of pages.
That includes bank statements covering several months, tax returns, Articles of Organisation, equipment invoices, and more. Each document contains data points that need to be identified, extracted, and reconciled.
With limited volume, analysts can handle these inconsistencies manually. But as the number of deals increases, the effort scales faster than headcount can support.
In its Q3 2026 outlook, ELFA reports that equipment investment exceeded a 15% annual rate in both Q1 and Q2 and projects more than $129 billion in 2026 new business volume among surveyed member companies. As volume grows, institutions built around manual preparation should measure whether operational capacity keeps pace.
They will either need to expand staff at an unsustainable rate or limit volume, both of which undercut profitability in a margin-sensitive industry.
Agentic Credit Intelligence Is Not Just Another Toolkit
Most automation in lending today is rule-based. Data fields are extracted from structured documents and routed through pre-set logic. That approach works when everything matches expectations, but it breaks when files contain exceptions.
In reality, loan files rarely look uniform.
Bank statements differ by issuer. Tax data varies by business structure. Entity registrations appear in inconsistent formats. Invoices reflect different product categories and deal sizes.
Rules-based automation struggles here because it lacks the flexibility to reconcile contradictory signals or adapt to unfamiliar file structures.
Automated pipelines improved data retrieval, but underwriting isn't linear. Exceptions are routine. Static workflows struggle under real-world variability.
Agentic Credit Intelligence takes a different approach. It is built on an architecture of specialised agents, each designed to handle a specific part of the underwriting workflow.
Specialized components can gather entity-level information, classify bank statement transactions, cross-reference UCC findings, and compare identity or ownership details across documents. When evidence is missing or contradictory, the controlled behavior is to preserve the source and route the exception for review rather than silently resolve it.
The useful distinction is how the system handles the unexpected. Static automation expects structured, consistent inputs. An agentic preparation layer should identify variation, preserve uncertainty, and route unresolved issues to the appropriate human owner.
How Kaaj Changes the Preparation Layer
In a traditional scaling model, adding volume means adding analysts. Over time, operations costs grow in close proportion to deal flow.
With Kaaj, the preparation model is different. The system helps retrieve and reconcile data, organize financial analysis, and prepare a source-linked file for human review once the required documents are available.
Teams should measure the effect in a controlled pilot: preparation time by file type, exception rate, reviewer correction rate, traceability, and the amount of analyst work that moves from assembly to judgment. Staffing and unit-cost outcomes depend on the lender's volume, workflow, and deployment scope.
What's important to note is that this isn't about removing human oversight. It's about shifting human effort from data compilation and document review to higher-value evaluation.
The architecture should be evaluated on the lender's own file mix, policies, exception paths, and review requirements before broader rollout.
Evolution in Phases: From Static Workflows to Agent-Based Architecture
If you map out the road from where lending is today to where it needs to be to serve 33 million small businesses efficiently, the path becomes clearer when broken into stages. Lenders looking to protect margins while scaling will need to move through this evolution deliberately:
Phase 1: Centralised Task Consolidation
Designated ownership improves, but volume still outpaces analyst capacity.
Phase 2: Document Automation
OCR and templates extract fields, but they fail when the document format doesn't match the expected template.
Phase 3: Parallel Agent Architecture
Specialized components can prepare independent sections of the file in parallel instead of forcing every task through a single sequential queue. Actual processing time depends on file size, document quality, integrations, and the review path.
Phase 4: Reasoning Layer Integration
Agents can identify files, cross-reference details, and apply configured preparation logic. Conflicts and low-confidence results should remain visible for human review. This is where agentic credit intelligence becomes an operational preparation layer rather than an autonomous decision-maker.
For modern lenders—private credit funds, equipment finance providers, and revenue-based finance firms—this evolution is worth evaluating where manual preparation is constraining capacity. The case should rest on measured workflow outcomes, not an assumption that every lender needs the same architecture.
The evolution is not about replacing underwriters. It is about preparing more complete, traceable, decision-ready information while keeping exceptions and final judgment with the credit team. Any speed advantage should be measured by file type and review path.
Why This Is Relevant Now
Small business lending is high volume and relatively low margin. That makes processing efficiency a strategic priority, not just an operational concern. Agentic systems are the natural response to a market that is both growing in demand and structurally resistant to traditional scaling.
What Changes With Kaaj
| Function | System-of-record owner | AI-supported preparation | Human control and evidence |
|---|---|---|---|
| Intake and classification | LOS, CRM, or lender intake system | Inventory documents, classify files, and identify gaps | Operations confirms completeness; source files remain available |
| Business and document verification | Lender-approved verification systems and case record | Compare submitted details and surface mismatches or integrity signals | The lender resolves identity, compliance, and investigation questions |
| Financial analysis | Lender credit file | Prepare transaction classifications, reconciliations, and source-linked findings | The underwriter reviews assumptions and determines materiality |
| Credit memo | Lender-approved memo or credit record | Draft the factual sections, exceptions, and source references | The underwriter edits and owns final language and recommendation |
| Final decision | Lender LOS or decision system | No autonomous approval; present prepared evidence and open items | Authorized lender personnel make and record the decision |
Kaaj is designed for this exact shift. It deploys agents to make loan files decision-ready by processing documents, verifying business information, summarising bank statement data, and preparing the analysis that underwriters rely on to make final decisions.
Kaaj is designed to reduce repeated preparation work while leaving the lender's LOS or CRM as the system of record and credit authority with the lender. The economic and capacity effects should be established through a scoped pilot rather than assumed in advance.
- Prepares recurring document and reconciliation work for reviewer validation
- Produces structured summaries with links back to supporting evidence
- Surfaces data discrepancies before they are treated as resolved facts
- Routes exceptions to lender-defined human owners
- Preserves reviewer edits, overrides, and open questions in the workflow
- Works alongside the lender's existing systems and decision process
To evaluate the approach, compare SMB underwriting automation tools, review the Kaaj product workflow, or read the human-in-the-loop FAQ. To discuss a scoped pilot, book a demo.
Ready to see Kaaj in action?
Book a demo and walk through a live deal with our team — from intake to credit memo.
Book a demo