Peer-To-Peer Lending Models

AI agents aren’t just the next big thing - they are rewriting the rules of how we think, decide, and execute. Here's how they work. 𝟭.𝗧𝗵𝗲 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 AI agents operate like autonomous workers. A typical flow includes: A. Input from the environment: APIs, real-world sensors, or direct human prompts feed raw data into the system. B. Memory systems: Agents store and recall relevant context — just like a good colleague remembers past meetings or policies. C. Reasoning engine: They don’t just process — they “think,” applying logic and learned knowledge to make decisions. D. Orchestration: This is the control room, coordinating multiple steps, tools, or even other agents to complete a task. E. Guardrails: Built-in rules and policies ensure safe, compliant actions, especially in regulated environments. F. Agent-to-agent communication: Using emerging protocols agents now talk to one another to complete workflows collaboratively. 𝟮. 𝗘𝘅𝗮𝗺𝗽𝗹𝗲: 𝗦𝗠𝗘 𝗹𝗼𝗮𝗻 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 Imagine an AI agent working inside a bank’s SME lending unit: A. A small business applies for a loan. The agent pulls in data from application forms, bank account activity, credit bureaus, and open banking APIs. B. The agent recalls similar past cases, prior risk models, and policy exceptions. It understands the client’s history with the bank. C. It evaluates the applicant’s risk profile, compares loan terms, simulates repayment scenarios, and identifies anomalies (e.g., sudden revenue spikes). It flags one item as borderline and prepares justifications. D. The agent coordinates with other agents specialized in document verification, compliance, credit pricing. Together, they generate a complete credit memo. E. Built-in rules ensure the loan complies with internal risk limits, ESG criteria, and regulatory obligations. It escalates only if thresholds are exceeded. F. The agent shares the decision with the treasury and onboarding agents. Treasury adjusts funding allocation; onboarding prepares digital signatures and account disbursement. What once took weeks and five departments now happens in minutes. 𝟯. 𝗟𝗲𝘃𝗲𝗹𝘀 𝗼𝗳 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆 Not all AI agents are created equal. Capgemini proposes a 5-level maturity scale: Level 0: No AI. Think manual spreadsheets. Level 1: AI-assisted workflows. You’re still in charge, AI makes it faster. Level 2: Augmented decisions. AI provides options, you choose. Level 3: Integrated agents. They execute within controlled domains. Level 4: Multi-agent workflows. Like a team of bots handling customer onboarding while another handles compliance. Level 5: Fully autonomous. Human input shifts to governance and strategy only. Right now, most companies are stuck at Level 1, but leading firms are already scaling Level 2–3 implementations. Based on: Capgemini Research Institute / Rise of Agentic AI 𝐒𝐮𝐛𝐬𝐜𝐫𝐢𝐛𝐞 𝐭𝐨 𝐦𝐲 𝐧𝐞𝐰𝐬𝐥𝐞𝐭𝐭𝐞𝐫: https://lnkd.in/dkqhnxdg

Govern to Grow: Scaling AI the Right Way    Speed or safety? In the financial sector’s AI journey, that’s a false choice. I’ve seen this trade-off surface time and again with clients over the past few years. The truth is simple: you need both.   Here is one business Use Case & a Success Story. Imagine a loan lending team eager to harness AI agents to speed up loan approvals. Their goal? Eliminate delays caused by the manual review of bank statements. But there’s another side to the story. The risk and compliance teams are understandably cautious. With tightening Model Risk Management (MRM) guidelines and growing regulatory scrutiny around AI, commercial banks are facing a critical challenge: How can we accelerate innovation without compromising control?   Here’s how we have partnered with Dataiku to help our clients answer this very question!   The lending team used modular AI agents built with Dataiku’s Agent tools to design a fast, consistent verification process: 1. Ingestion Agents securely downloaded statements 2. Preprocessing Agents extracted key variables 3. Normalization Agents standardized data for analysis 4. Verification Agent made eligibility decisions and triggered downstream actions   The results? - Loan decisions in under 24 hours - <30 min for statement verification - 95%+ data accuracy - 5x more applications processed daily   The real breakthrough came when the compliance team leveraged our solution powered by Dataiku’s Govern Node to achieve full-spectrum governance validation. The framework aligned seamlessly with five key risk domains: strategic, operational, compliance, reputational, and financial, ensuring robust oversight without slowing innovation.   What stood out was the structure: 1. Executive Summary of model purpose, stakeholders, deployment status 2. Technical Screen showing usage restrictions, dependencies, and data lineage 3. Governance Dashboard tracking validation dates, issue logs, monitoring frequency, and action plans   What used to feel like a tug-of-war between innovation and oversight became a shared system that supported both. Not just finance, across sectors, we’re seeing this shift: governance is no longer a roadblock to innovation, it’s an enabler. Would love to hear your experiences. Florian Douetteau Elizabeth (Taye) Mohler (she/her) Will Nowak Brian Power Jonny Orton

Sitting with CTOs from 16 major lenders last week, I asked one question: "How well does your LOS handle complex decisioning?" Average score: Below 7. Not because their systems are broken. But because loan origination systems were never built to be decision engines. Here's what Rafi Goldberg from Sapiens explained on the Power House podcast that changed my perspective: AI decisioning isn't about replacing your underwriters. It's about competing on decisions. Think about what actually differentiates your lending: • That 20-year underwriter who knows when to make exceptions • The processor who catches patterns others miss • The branch manager with instincts you can't explain That institutional knowledge is your competitive advantage. Except it's trapped. The technical challenge isn't automation—it's translation. How do you convert decades of human pattern recognition into decision logic that scales? This is where the architecture matters: Traditional business rules approaches fail over time. They become brittle and inflexible, an albatross of technical debt unable to meet business needs. AI decisioning changes that paradigm. Combining declarative decision models with analytics and AI, your experts’ decision can now be converted to business assets at scale, with no loss in business intent and all the observability and adaptability you’ve come to need and expect. One CTO today said it perfectly: "Our LOS manages transactions. But our decisions happen in Excel sheets and email chains." That's the gap. While everyone races to perfect their point-of-sale experience, the real differentiator is decision velocity and precision. Your best people make hundreds of micro-decisions daily. Each one based on experience you can't hire off the street. When they retire, that knowledge disappears. Unless you capture it now. The mortgage industry keeps focusing on the wrong automation. We digitize applications. We automate verifications. We streamline workflows. But decisions? Those still happen in silos. What if your junior underwriter could access your senior team's pattern recognition? What if every loan officer could tap into your best performer's instincts? That's not replacing human judgment. It's amplifying it. The lenders who win the next decade won't have the slickest UI or the fastest application. They'll be the ones who turned their tribal knowledge into scalable, intelligent decision engines. Every lender in that room today knew their LOS wasn't built for this. The question is: Who's going to fix it first?

A credit application OCR is easy to build once. Making it run reliably for 5,000+ applications a year is a different game. When I took my prototype from a notebook to production-grade standards, the biggest gains came from designing for failure before it happened. Here’s the thinking behind each major decision: 1️⃣ 𝗦𝘁𝗮𝘁𝘂𝘀 𝗺𝗼𝗱𝗲𝗹 𝗶𝗻𝘀𝘁𝗲𝗮𝗱 𝗼𝗳 𝗮𝗱-𝗵𝗼𝗰 𝗳𝗹𝗮𝗴𝘀 • With multiple steps (upload → OCR → LLM → validation), failures were hard to track. • A formal 𝘴𝘵𝘢𝘵𝘦 𝘮𝘢𝘤𝘩𝘪𝘯𝘦 for both applications and extraction jobs meant any crash could be resumed exactly where it left off, no reprocessing entire batches. 2️⃣ 𝗔𝘀𝘆𝗻𝗰 𝗷𝗼𝗯𝘀, 𝗻𝗼𝘁 𝗹𝗶𝗻𝗲𝗮𝗿 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 • OCR and LLM calls could take minutes per document, blocking the whole process. • Using 𝘊𝘦𝘭𝘦𝘳𝘺 + 𝘙𝘦𝘥𝘪𝘴 for job orchestration kept the UI responsive and allowed failed jobs to be retried independently. 3️⃣ 𝗛𝘂𝗺𝗮𝗻-𝗶𝗻-𝘁𝗵𝗲-𝗹𝗼𝗼𝗽 𝗯𝘆 𝗱𝗲𝘀𝗶𝗴𝗻 • OCR and LLM errors can’t be fully eliminated. • A 𝘗𝘋𝘍 𝘰𝘷𝘦𝘳𝘭𝘢𝘺 with extracted fields and confidence scores let loan officers validate an entire application in minutes, not hours. 4️⃣ 𝗠𝗲𝘁𝗮𝗱𝗮𝘁𝗮-𝗱𝗿𝗶𝘃𝗲𝗻 𝗳𝗶𝗹𝘁𝗲𝗿𝗶𝗻𝗴 • Different credit types require different documents and fields. • Filtering by document type, stage, and confidence score meant irrelevant or low-confidence data never reached the final application form. 5️⃣ 𝗠𝗼𝗱𝘂𝗹𝗮𝗿 𝗰𝗼𝗺𝗽𝗼𝗻𝗲𝗻𝘁𝘀 • Vendor lock-in and future upgrades were a risk. • Each step (OCR, LLM, storage) runs as a 𝘴𝘦𝘱𝘢𝘳𝘢𝘵𝘦 𝘴𝘦𝘳𝘷𝘪𝘤𝘦 𝘸𝘪𝘵𝘩 𝘢 𝘤𝘭𝘦𝘢𝘯 𝘈𝘗𝘐, so swapping Azure OCR for another provider is a config change, not a rewrite. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗹𝗲𝘀𝘀𝗼𝗻 Robust AI systems aren’t built by adding “more AI”, they’re built by anticipating where things break, and making the system self-recover and easy to adapt. If you’re moving from a demo to production, design for failure before it happens. 💬 What’s one design choice you made early that saved you later? ♻️ Repost to help someone in your network

Scorecards vs Statements — Are We Lending on Insight or Guesswork? “Loan approved in 3 minutes. Defaulted in 3 months.” I read this on an internal note. A ₹2 lakh personal loan disbursed based on a fancy new AI scorecard. No documents, no bank statements, just behavioral and device data. Smooth process. Fast turnaround. Wrong customer. We’re at a turning point in Indian lending. On one side: Scorecard-based assessments — built on alternate data, mobile signals, UPI trails, and ML-based risk models. On the other: Classical credit underwriting — deep dives into financials, banking patterns, GST returns, and business cashflows. Credit bureaus and consultants are praising the new models. Fintechs are scaling with machine-driven decisioning. Government infra like Account Aggregator and OCEN is making access seamless. But pause for a moment. ? Are we replacing understanding with algorithms? As someone leading MSME lending at Yes Bank, I see both sides. We want speed. We want inclusion. We want to lend to the “new-to-credit” without saying no. But speed without seasoning? That’s risky. Have these scorecard models been tested across credit cycles? Can they catch subtle stress signals like declining working capital or bounced vendor payments? >> A bank statement tells you the story. >> A GST return shows the discipline. >> A good credit officer knows when something doesn’t “feel right”. >> Can a scorecard feel anything? Yes, AI helps. Yes, TAT improves. Yes, tech has a place. But credit is not just about who applies — it’s about who repays. If we're underwriting ₹50,000 loans in 5 minutes, great. But when we do ₹50 lakh business loans this way — we better be sure we’re not just guessing faster. Let’s not forget: ✅ Account Aggregator can now give us banking + GST data in minutes. ✅ OCEN is building pipes for contextual lending. ✅ UDYAM Assist is bringing informal entrepreneurs into the formal fold. The tools are ready. The question is: How are we using them? The Big Trade-Off Do we want a lending model that looks sharp today… But silently builds credit cost tomorrow? Or one that takes a little longer… But lasts longer too? There’s no easy answer. But here’s my view: AI + Human Judgment is the way forward. Not either-or. Both. Because when a loan goes bad, it’s not the algorithm that gets the recovery call — it’s us. What’s your take? Are we too quick to ditch classical underwriting? Will scorecards redefine credit — or just repackage risk? Let’s talk — openly, honestly, and without buzzwords. #CreditAssessment #AccountAggregator #OCEN #RiskManagement #DigitalLending

People tell me I'm sharing too much online. Giving away the secrets. Here's the secret: We do the boring work nobody else wants to do. We saved thousands per loan without sacrificing quality. That means better rates for borrowers and more comp for originators. Not through magic. Through obsessing over the basics and refusing to pass defects downstream. Every defect costs more to fix the further it travels. Problems that show up in closing didn't start in closing. Where you find the problem isn't where it was created. During the last six months of 2024, we increased closing and processing productivity by 20% by setting quality standards and automating everything we could. Today, we have almost zero drama between processing and closing. We used to have plenty. The usual corporate answer: "More training." That's been our biggest shift. We don't blame people, and the answer is never training. When someone in a meeting says "we need more training," I lose it. Management's job is to design a system that doesn't rely on constant training to avoid constant problems. Deming taught that 94% of quality problems come from the system, not the worker. He was right. Here's some examples of what we actually did: • AI and workflows require processors to complete VOE before submission to closing. • Early CDs go out automatically the day the loan is CTC. • Credit, appraisal, and VOE invoices are added to the eFolder automatically through APIs and pre-fill on the CD — 100% accurate every time. • Settlement agent contact info is extracted with AI, so it's never wrong. • AI compares title, HOI, CPL, purchase contract, AUS, and ID for name discrepancies — triggers a task if something doesn't match before getting to closing. The system flags HOI effective dates that fall after closing date and forces correction before submission. Mortgagee clauses are compared automatically and corrected early. All drama comes from sloppy agreements. We had no measurable standards between closing and processing. We were relying on people to "just know what to do." Now the system finds the mistakes and blocks submission until they're fixed. Processors are more efficient because they're not hunting for problems. Closers are more efficient because they're getting clean, consistent files. They're on the same team now instead of pointing fingers. Because productivity is up, costs are down. That's how borrowers get better rates and originators make more money. Not through clever tricks. Through doing the unglamorous work of fixing the system. We're doing this across every department. It's not one big change — it's a thousand small ones. The boring stuff everyone knows they should do but never actually does. That's the secret. Do the boring work. Fix the system. Stop blaming people. The results take care of themselves.

AI-driven underwriting is reshaping lending economics, and surprisingly few have caught on yet. I've been reflecting on why credit decisioning, especially to SMBs, remains so manually intensive. At first glance, you'd think regulation or tech limitations hold things back, but the core bottleneck is actually human labor. Banks still rely heavily on manual processing—reviewing outdated financial statements and Dun & Bradstreet reports, and depending heavily on human judgment to catch subtle risk signals. This problem feels familiar to one I worked on at Nauto (AI software for driver safety). Our models had to detect every crash perfectly (zero misses). But if we optimized strictly for perfect recall, precision plummeted. We flagged too many false positives, slowing down our human reviewers. So we built a human-in-the-loop system where AI pre-highlighted events, shrinking human review time down to just five minutes. The hybrid AI-with-human-oversight solution was key to managing scale and efficiency without sacrificing accuracy. Banks face the same recall-precision dilemma with underwriting. Traditional financial metrics, which are manually prepared, months old, and often incomplete, mean underwriters either miss important signals or drown in excessive manual reviews. At Slope, our hunch was that raw bank transactions could tell us more than quarterly financial statements ever could. So we built specialized LLMs trained on bank transaction data. With AI, we now construct credit-grade financials that are: ➡️granular (transaction-level) ➡️fresh (refresh daily) ➡️instantly verifiable (cannot be falsified) Then we layered on real-time signals from customer reviews and employee headcount changes that let us detect critical business shifts weeks or months before official reports. Our model dramatically cuts risk and cost. It opens up entirely new lending markets, segments previously labeled "too risky" or "not worth it." And this isn’t theoretical. Our models are assisting banks in underwriting millions of $ to real businesses, today. It reminds me of cloud computing replacing on prem services — a structural economic change, rather than a marginal improvement. If you're exploring similar shifts, reach out — I'd love to compare notes.

AI in Lending Isn’t Coming. It’s Already Here. Let’s be clear: AI is no longer theoretical. It’s operational. Tangible. Transformational. It’s not “the future of lending.” It’s today’s competitive advantage. Here’s how leading lenders are using AI right now, not in pilots or sandboxes, but in production: *AI-Powered Lead Management Conversational AI is now the “front door” to lending. Borrowers get instant answers, pre-qualifications, and appointments, even at midnight. One lender uses AI to predict with 89% accuracy whether a loan will close on the first call. *Document & Income Automation AI can classify over 1,000 document types in seconds. Income calculations, fraud checks, and inconsistencies are flagged instantly, no more manual stare-and-compare. *Fraud Detection in Real Time AI models are spotting altered documents and duplicate submissions that even seasoned underwriters might miss. *Proactive Servicing Speech analytics detect borrower stress during calls, alerting servicers before a missed payment happens—turning risk into retention. *Predictive Lending Intelligence AI is flagging refinance opportunities before the borrower even thinks to call. Some lenders are closing business before the competition even sees it. *UWM’s “Mia” Chatbot Mia handles borrower questions, schedules appointments, leaves personalized voicemails—and never sleeps. AI isn’t just improving mortgage operations—it’s redefining them. The organizations embracing AI today are: *Cutting costs *Speeding up cycle times *Delivering superior borrower experiences So here's the real question: Is your organization an Avoider, an Experimenter, or a true Leader in AI adoption? The future won’t wait. And the market isn’t pausing. Now is the time to decide: Will you adapt or be left behind? Eric Kujala Paul Orlando Jenna Nelson, CSM Ashley Gravano Fobby Naghmi Kathleen Mantych Ruth Lee, CMB Todd Feager Jake Vermillion Ana Cramer Faith Murphy, CMB® Suzy Lindblom Eileen Andersen Brian Vieaux, CMB Christine Beckwith Julia Brown Stew Scott Ed Kourany Jr., JD, MBA Dana Georgiou, CPLA, CFM Suha Zehl, CMB® Kortney Lane- Schafers

The mortgage industry doesn’t just have a cost problem — it has a loan quality problem that shows up as cost. What stood out to me in recent conversations with VPs of Mortgage Operations is how many critical defects still trace back to basics: missing income docs, unverified employment, assets that don’t tie out, and conditions that bounce between ops and underwriting. The fix isn’t another checklist. It’s smarter, upfront data collection at the point of sale, wired into underwriting and QC. A few concrete use cases we’re seeing: • Dynamic needs lists - Instead of static checklists, generate borrower‑specific needs based on product, income type, occupancy, and risk profile. Borrowers only see what’s truly required for *their* file, which reduces back-and-forth and early fallout. • Upfront income, employment & asset validation - Run VOI/VOE/VOA during application. Use document intelligence to read paystubs, bank statements, and tax returns, reconcile them against stated data and internal guidelines, and flag gaps before the file ever hits an underwriter’s queue. • Automated conditions clearing - When new documents arrive, AI agents re-check ratios, guidelines, and disclosures in real time and pre-tag conditions as “ready to clear” for the underwriter — turning manual detective work into a simple approve/decline review. • Pre-funding QC & repurchase defense - Continuously scan files for potential critical defects, document exceptions against policy, and maintain a complete audit trail. Fewer surprises in post-closing, stronger posture if a file is ever questioned. At Sei AI, this is exactly where our compliant AI agents for document intelligence come in — helping CX, ops, underwriting, and compliance teams build clean, audit-ready loan files from day one, so cost-to-originate comes down *because* loan quality goes up.

"Continuing on the loan approval process using Agentic AI" Agentic AI in lending is powerful. It's also unforgiving. One hallucinated APR calculation, one unexplained denial, one fair lending violation — and you're in regulatory territory no model card can save you from. So how do you actually build it safely? You stop trying to make one agent do everything. You decompose the workflow into specialized, governed agents — each with a clear job, clear boundaries, and clear escalation paths. Here's the architecture I've been working on Orchestrator — Stateful coordinator (not an LLM) that routes work across the loan lifecycle DAG. Data Ingestion Agent — Pulls and validates credit bureau and income data. Detects staleness and conflicts. Decisioning Agent — ML ensemble + rules engine. Generates SHAP explanations for every decision. Compliance Agent — Deterministic validation against Reg Z and fair lending rules. No LLM in the calculation path. Human-in-the-Loop Gateway — Escalates low-confidence decisions and exceptions with full context. The hard part is orchestration, governance, and knowing when autonomy ends and human judgment begins. How would you govern autonomy in your production AI systems? #AgenticAI #Fintech #LendingTech #ResponsibleAI #AIGovernance #MachineLearning