CASE STUDY · FINTECH AI UNDERWRITING & CREDIT DECISIONING
Cutting loan approval time from 10 days to 15 minutes.
- 15 minutesStraightforward applications decided in about 15 minutes instead of 10 days
- 90Default rate unchanged in the 90-day test
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised digital lender. Name withheld under confidentiality.
Automated checks for straightforward applications, with complex cases sent to underwriters.
- 1Approving a loan took 10 days.
- 2Most applications were routine, yet every one went through the same manual queue.
- 3Applicants dropped out while they waited, and underwriters spent most of their time on cases that did not need them.
02 · WHAT WENT WRONG
What went wrong, and how we fixed it.
Every engagement hits problems. These were the ones that mattered here, and what we changed so they did not come back.
- PROBLEM 01A bureau data provider was slow during peak hours
Applications timed out, and the first version failed them as "unable to verify", which sent good applicants to decline. We changed the design so that a failed external check never causes a decline. It now retries, then refers to a person.
- PROBLEM 02We also added an alert on provider response times
- PROBLEM 03A second hurdle was cultural
Some underwriters distrusted the automated route and re-checked every automated approval, which cancelled the time saving. Instead of arguing, we gave them evidence. We showed a sample of automated decisions next to what they would have decided, and the agreement was high on the simple band.
- PROBLEM 04We then agreed a sampling rule: a fixed share of automated approvals is reviewed by a person each week, and the rest are left alone
The review caught a few edge cases that improved the rules.
03 · THE PROCESS
How we ran it.
We started by sorting a sample of past applications into simple and complex by hand, together with the underwriting team. That gave us the rules for routing.
- 01Step 1
Weeks of mapping: every step an underwriter took, what they checked, and what made them stop.
- 02Step 2
Build: automated checks for identity, income, bureau data and policy rules, with a clear pass, fail or refer outcome.
- 03Step 3
Shadow run: the automation made decisions in the background while underwriters worked as normal, and we compared results daily.
- 04Step 4
Live trial: straightforward applications were decided automatically while we tracked the default rate for 90 days.
A short daily review in the shadow period gave underwriters a say in what counted as straightforward.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
Underwriters spend their time on complex cases and on reviewing the automated route, not on routine files.
- NOW 02
Risk reviews the share of applications sent to review, and tunes the threshold with evidence.
- NOW 03
Changes to rules are versioned and approved before release, and the team watches default rate against the 90-day baseline.
- NOW 04
We handed over the rule set, a runbook for outages and a short guide on how to change a threshold safely.
- 15 minutes
Straightforward applications decided in about 15 minutes instead of 10 days
- 90
Default rate unchanged in the 90-day test
05 · WHAT WE BUILT
How it works.
- Document extraction reads statements and income proofs, and flags mismatches with what the applicant entered.
- Rules handle hard policy limits.Anything outside the limits is referred.
- A score handles the grey area, and borderline scores go to a person.
- Each application carries a decision record: checks run, results, and the route taken.
- Queues for underwriters show the reason for referral, so they start with context.
- Retries and timeouts for external data providers, with a safe refer-to-human fallback when a check cannot complete.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- Approving a loan took 10 days. Most applications were routine, yet every one went through the same manual queue. Applicants dropped out while they waited, and underwriters spent most of their time on cases that did not need them.
- What did Brandsmashers build?
- Automated checks for straightforward applications, with complex cases sent to underwriters.
- How was the work run?
- We started by sorting a sample of past applications into simple and complex by hand, together with the underwriting team. That gave us the rules for routing.
- What were the results?
- Straightforward applications decided in about 15 minutes instead of 10 days; default rate unchanged in the 90-day test.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
