Loan decisions in 40 seconds: personal terms for every client, doubtful applications checked by a specialist
A specialist used to check every application by hand, which took 1–2 days, and everyone got the same cautious rate. Now the system assesses each application on the company's own loan history and shows personal terms in 40 seconds, and the specialist receives doubtful applications with the analysis already done.
- Client
- A non-bank lender≈40,000 applications a year
- Service
- application assessment and loan decisions
- Status
- In production at the clientevery application runs through the system
- Before → After
- Decision on an application: 1–2 days → 40 s1.8 pp fewer loans not paid back
per client data the client gave us this numberestimate we calculated it; the assumptions are next to it
Where the client started
The lender receives about 40,000 applications a year, and a specialist checked each one by hand. A decision took 1–2 days, and some approved clients borrowed elsewhere in the meantime.
Everyone paid the same rate, so it had to be cautious: too high for reliable clients, and still no filter for the ones who wouldn't pay back.
applications a year
per client datafrom application to decision
per client datathe same for every client
per client dataThe main ruleThe system calculates, and a person decides the doubtful cases: they go to a specialist with the analysis ready, and the specialist makes the call.
What the system does
- Takes the application on the website and collects everything up front: amount, term, the client's details and documents, so the specialist never has to chase anything.
- Checks documents and identity automatically, so a stolen or forged application never reaches a decision.
- Scores risk with a model trained on the lender's own loans and how they turned out, so decisions rest on how this lender's clients actually behave.
- Calculates personal terms: rate, limit and schedule for each client's risk, so a reliable client gets a better rate and a risky one a lower limit.
- Shows the client ready terms in 40 seconds with a payment schedule and a button to sign, so they don't go to another lender while they wait.
- Decides clear-cut applications on its own, so the specialist doesn't spend time on the obvious.
- Sends doubtful applications to the specialist with the analysis ready: what speaks for the loan, what speaks against it, and which terms are possible. The specialist makes the decision.
- Logs every decision by the model and by people, with the reason, so any loan can be explained and audited.
- Retrains every month on loans whose repayment is already known, so the assessment keeps up with how clients behave now.
- Puts a new model version to work only if it tests no worse than the previous one, so decision quality can't slip unnoticed.
How it is built
An application from the website goes to verification with its documents, the model scores the risk on the lender's loan history, and the system calculates personal terms. Clear-cut applicants see their terms in 40 seconds; borderline ones go to the specialist with an analysis.
- Data
- applications and documentsissued loans and their repaymentcredit historyspecialists’ decisions
- Roles
- specialist: doubtful applicationsclient: terms and signingrisk manager: new model versions
How it works
- client · websiteApplicationAmount, term, documents
- systemVerificationDocuments and identity
- modelRisk assessmentOn the lender's own loan history
- systemTermsRate, limit and schedule to match the risk
- system40 secondsClear-cut applications: terms go to the client at once
- specialistDoubtful caseA decision with the analysis ready
- clientPayoutSigning and the money
- modelLearningMonthly; only a version no worse than the last goes live
Defaults don't show up in the first week; they are counted on loans that have been running for several months. So −1.8 percentage points is the measured difference between loans issued after launch and loans issued before it, compared at the same age. It is not a model forecast.
Results and payback
from application to personal terms on screen
per client dataa smaller share of loans not paid back
per client dataapproved applications that end in a payout: clients no longer leave to wait elsewhere
per client dataProject payback: 6–9 months per client dataThe gain here is in portfolio quality. The difference in defaults shows once loans issued under the new model have been through several months of repayment, so payback follows the credit cycle: 6–9 months.
What's live
- live
Website applications with documents and automatic identity checks
- live
Assessing applications on the lender's own loan history, with a personal rate, limit and schedule in 40 seconds
- live
Gray-zone cases go to the specialist with the analysis ready, and every decision is logged
- live
Monthly assessment updates, checked against the previous version
A similar task in your business?Solomiia, Pavlo's AI assistant, will ask about the details and prepare your conversation with him.
Custom software
Systems built around the company's processes: application assessment, client portals, CRM. The connections to document checks and credit history are emilus. Integrator. Price after the brief.