Case study

The yes comes before the account.

A retail and SME bank in Central and Eastern Europe sells cash loans of roughly EUR 1,000 to 30,000, over 3 to 60 months. Today, an eight-step application built and implemented by FlowX.AI takes a cash-loan request from first entry to final decision: applicant and co-borrower data, OTP-validated consents, credit bureau and tax authority checks, weighted income, scoring and debt-to-income, and one of three outcomes. Approved automatically, routed to an underwriter, or rejected.

  • Banking
  • Retail lending
  • Cash loans
  • Straight-through processing
  • Human-in-the-loop

Initial results from implementation and production

0.0K+
Loan applications already handled by the agentic solution
0w
From pilot go-live to the switch to Automatic Approval
0%
Of customers learn whether they qualify before opening an account
0
Steps from first entry to final decision, in one application

What if borrowers knew whether they qualified before opening an account, and underwriters only saw the applications that need their judgement?

Before FlowX.AI, parts of the credit decision ran through a master spreadsheet file and the judgement of experienced colleagues: correct rules, applied by people who knew them well, and invisible to everyone outside the process, the customer included. The sequence ran backwards, too. A customer who wanted to know whether they qualified first had to open a current account at the bank, then wait, hearing nothing between steps, for a decision that could vary with whoever made it.

Every application depends on two external systems, the credit bureau and the tax authority, that neither the bank nor FlowX.AI controls, and every automated decision has to be one the bank can explain. The engine behind it is not a model making a prediction. It is the bank’s own credit policy, executed the same way on every application.

The starting position

The pressure was sequence and trust, not volume.

Four uncomfortable truths defined the starting state.

01

The rules lived in a spreadsheet

Parts of the credit decision ran through a master Excel file. The policy was correct and well understood by the people applying it, but nothing enforced consistency, so decisions varied between the people making them.

02

The answer came after the account

A customer had to open a current account at the bank before the bank could tell them whether they qualified for the loan. Someone who only wanted to know if they were eligible first had to become a client.

03

The customer heard nothing

Between steps, there was no system to tell them anything.

04

Systems nobody in the project controls

Every application needs the national credit bureau and the national tax authority. Both are mandatory, both are external, and neither answers to the bank or to FlowX.AI. Whatever automation looked like here, it had to keep working when they were slow, and it had to produce decisions a bank answerable to a regulator could explain.

What changed

The credit policy, moved into the application.

What used to live in a master spreadsheet now lives where everyone working the file can see it. Each pressure from the starting state now has a concrete answer.

  1. Pain point

    Credit rules in a master Excel file, invisible outside the process.

    How we solved it

    Credit policy as configuration. Thresholds, income coefficients, knock-outs and the risk-to-DTI matrix live in the application, visible to the staff working the file.

  2. Pain point

    Decisions vary with the person making them.

    How we solved it

    One criteria table. Every application is scored against the same conditions, with three possible outcomes: approved automatically, routed to an underwriter, or rejected.

  3. Pain point

    A current account required before the customer could learn whether they qualified.

    How we solved it

    Decision first. The customer gets a credit decision without opening an account. The account follows the yes.

  4. Pain point

    No contact with the customer between steps.

    How we solved it

    Customers kept informed. Automated email at each main step carries the documents the customer needs for their own analysis or archive. Rejections arrive the same way, at whichever step they occur.

  5. Pain point

    Two mandatory external systems outside the bank’s control.

    How we solved it

    Built to survive dependencies. Asynchronous, retry-safe credit bureau and tax authority calls. Applications waiting on either hold their status and resume from where they stopped.

The agent

Eight steps, from first entry to final decision.

What the cash-loans straight-through processing agent does, in three phases: capture the application, consent and verify, score and decide.

Phase 01

Capture the application

  1. 1

    Credit request

    Product, amount, term, repayment type and discounts, with an informative calculation.

  2. 2

    Personal data

    Identity, residence, contact and household; age and residency knock-outs enforced on entry.

  3. 3

    Co-borrowers

    Up to two, with the same data and the same knock-outs.

Phase 02

Consent and verify

  1. 4

    Consents

    GDPR, tax authority and credit bureau consents per party, each validated by a 6-digit email OTP. Tax authority consent is valid for 5 days; after that, the system forces re-validation.

  2. 5

    Credit bureau

    FICO score, exposures and existing loans per party, with refinancing selection. Queried with a 30-second timeout and 3 automatic retries, 10 seconds apart.

  3. 6

    Income

    Asynchronous tax authority query on the same retry pattern, polled every fifteen minutes (configurable). Eligible income types and weighted net income per party.

Phase 03

Score and decide

  1. 7

    Offer

    Scoring and debt-to-income, showing the requested amount alongside the maximum available offer.

  2. 8

    Final decision

    Approved automatically, routed to an underwriter for manual analysis, or rejected. Every action is logged with user, timestamp and action type.

The hard part

A credit policy you can execute.

Automatic Approval is the part that asked the bank to change its mind about what software can be trusted with. It works because almost nothing in it is a judgement call. Applications that clear every knock-out but miss an auto-approval condition go to an underwriter.

Rule 1

Income is weighted, not counted

Each eligible income type carries its own minimum tenure, calculation method and coefficient. Stable income counts in full; variable income such as rent, self-employment or dividends counts at a discount. Salary is averaged over several months and capped against the prior year, so a recent raise cannot carry an application on its own.

Rule 2

Indebtedness follows the regulator, then tightens with risk

Maximum debt-to-income starts at the regulator’s 40% cap for local-currency loans and steps down by risk category. It is calculated at the higher applicable rate, so the decision still holds after any fixed-rate period ends.

Rule 3Explainable

Every threshold is configuration, not inference

Every automatic approval can be explained by pointing at the rule that produced it.

We appreciate that before a solution is offered, we also discuss how it will be implemented. The automated solution saved time for our client advisors, for our underwriting colleagues, and most of all for our customers. I think the appreciation for the current solution can also be measured by the new developments we keep putting on the table.

The bank’s Senior Product Manager, Consumer Lending
Delivery

From discovery to Automatic Approval.

Go-live with people making every decision, a pilot against the bank’s own underwriters, then Automatic Approval, switched on by the bank’s business users.

Month 1

Two-week discovery: cash-loan process, credit bureau and tax authority integrations, IT security, infrastructure readiness.

Month 4, day 1

Go-live. Full eight-step process, every integration running. Automatic Approval switched off; people make the credit decision.

Months 4-5

Pilot. Relationship managers and underwriters work live applications and compare the system’s output with what they would have decided.

Mid Month 5

Automatic Approval live, approved by the bank’s own business users. Same release: tax authority report in the risk department’s format, DTI adjustment by income band, automatic rejection on games-of-chance income.

The bank ran the pilot with people deciding, watched the system against its own underwriters, and switched on Automatic Approval when they were convinced.

What lived in a master spreadsheet and in the judgement of experienced colleagues now lives in a system that applies it identically, shows its working, and logs every action.

What comes next

Automatic Approval is the foundation.

The bank got there by treating automated credit decisioning as a governance and engineering problem. For a bank answerable to a regulator, an explanation is worth more than an accuracy score, and every automatic approval comes with one. The next phases of the project:

01

Head-office reporting

On volumes, approval rates and DTI distributions, now in delivery.

02

Supervisor views

Flagging applications that breach the one-hour decision target (credit bureau and tax authority delays excluded).

03

Optional loan insurance

With premium and consent handling.

04

Risk-based pricing

05

OCR ingestion

Of scanned application forms.

About FlowX.AI

The orchestration platform for regulated work.

FlowX.AI is the enterprise orchestration platform that enables regulated institutions to deploy deterministic, source-traceable agentic workflows on top of legacy infrastructure. Financial institutions run critical customer and operations journeys on FlowX.AI, including:

  • Commercial Onboarding
  • Underwriting Assessment
  • Commercial Lending
  • KYC & AML
  • Customer Churn & Retention
  • Fraud Investigation
  • Underwriting Financial Insights
  • and more
Next

Put an auditable agentinside your credit decisions.

Bring a lending journey and we will show you a path to production, on your core, with your credit policy executed as configuration and the audit trail intact.