Use casesFinancial institutions & fintech
Behavioral score and limit management
Limits are set once at application; good customers outgrow them and quietly deteriorating ones keep full, undrawn lines.
- Kinds of AI
- ModelsAgentic AI
- Who decides
- Underwriter
- Where it runs
- Your servers, banking data kept in Türkiye
Agent · actions this run
- Limit increaseBand 1–2, utilization high for 3 months, clean bureau, step within policy cappolicy rule · tens
- Increase above capSame rule, above capunderwriter · few
- Freeze undrawnArrears elsewhere or a returned check in 30 dayspolicy rule · few
Model · one account: SME, food wholesale
- Model
- Agent
Written limitCaps used this run: within limits. Kill switch: armed. PSI: stable.
DecidesUnderwriter
1. The desk
Today
- ScoreCustomers are scored at application, then mostly forgotten until they miss a payment.
- ReviewLimit reviews are manual, at renewal.
- ActChanges wait for renewal: undrawn lines are drawn just before default, and growing customers move to another bank.
- DecideThe underwriter answers drawdown requests on revolving SME lines on judgment.
With the system
- ScoreThe behavioral model scores every account each month (weekly for cards and revolving SME lines) with reason codes; large drops go to early warning.Model
- ReviewEach account lands in one cell of the strategy matrix the credit policy committee writes.Your rule
- ActThe agent submits only pre-authorized cells through the bank's limit service, under written caps, and queues every other cell with its evidence.Agentic AI
- DecideThe underwriter decides the queued actions with their evidence, inside the committee's matrix, as before.Person decides
What lands on the desk
Each run reads as a list of actions, each tied to the rule that allows it and to who acts on it, with the caps used and the state of the kill switch. Account lines show why one customer moved and what happens next.
Model · migration vs. application
- MajorityStable
- Small groupDown two or more bands, sent to early warning
Agent · actions this run
- Limit increaseBand 1–2, utilization high for 3 months, clean bureau, step within policy cappolicy rule · tens
- Increase above capSame rule, above capunderwriter · few
- Freeze undrawnArrears elsewhere or a returned check in 30 dayspolicy rule · few
- Limit decreaseBand 8–10 and low utilizationunderwriter · few
Model · one account: SME, food wholesale
Agent · on this account
- Undrawn limit frozenrule F-2
- Routed to early warning
- Relationship manager notified first
- Model
- Agent
Written limitCaps used this run: within limits. Kill switch: armed. PSI: stable.
DecidesUnderwriter
2. The system
For your technical team — skip it if you like.
How it works
In
- Monthly snapshot (weekly for revolving)
- Utilization, payments, days past due
- Overdraft days, inflows, returned checks
- Bureau refresh
Model · estimates
Behavioral PD model
Logistic WoE scorecard or monotonic GBM, by product and tenure
Reads
- Everything under In
Out
- Behavioral band
- PD
- Reason codes
Code · calculates
Band migration
A transition matrix against the application band
Reads
- Behavioral band
- Application band at origination
Out
- Up, stable, or down two or more bands
- Large drops sent to early warning
Your rules
Policy strategy matrix
Written by the credit policy committee
Reads
- Band and utilization
- Profitability
- Exposure at default (CCF)
Out
- One cell per account
- Increase, decrease, or freeze undrawn
- Drawdown yes, no, or refer; or no action
Agent · prepares
Limit-action agent
A state graph over the matrix
Reads
- The cell for each account
Out
- Only pre-authorized cells submitted
- Per-cell amount and count caps
- Other cells to an underwriter, with evidence
- Notices from approved templates
Decision
Underwriter
Reads
- Queued actions with their evidence
Out
- The underwriter decides; the agent records the decision
How the agent's work is arrangedState graph
State graph
Every account sits in one matrix cell, and the policy writes which transitions are pre-authorized, with what caps, and which need a person; a graph keeps that explicit and auditable.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Behavioral model and strategy
Target
90 or more days past due in the next 12 months from each monthly observation point; features over 1-, 3-, 6-, and 12-month windows; a separate model for accounts younger than six months.
Method and challenger
A logistic WoE champion against a monotonic GBM challenger; usage and CCF models only where the book is mature enough.
Strategy
Back-test the matrix on 24 months of history, then run champion/challenger on a small, randomized share of the book.
Validation evidence
Gini, calibration, stability, and cell-level outcomes (balance, revenue, 6- and 12-month bad rate), for the validation unit.
Monitoring
Monthly PSI and cell performance; a quarterly strategy review; annual revalidation.
Where the agentic system stops
The underwriter decides inside the matrix the credit policy committee writes; the committee owns the matrix, caps, and notice rules.
Show its tools, limits, and record
- Tools and access
- Reads core banking and the bureau refresh; submits limit changes only through the bank's limit service and only for listed cells, each under a maximum amount per account, a maximum count per run, and a monthly cap on the share of the book that can be cut; sends notices from approved templates.
- May not
- Decrease a performing customer's limit, act on a group-connected exposure, or exceed a cap: these go to an underwriter.
- Who decides and the written limit
- UnderwriterThe underwriter decides inside the matrix the credit policy committee writes; the committee owns the matrix, caps, and notice rules. Inside the cells it pre-authorizes, the bank's limit service applies the change automatically within written limits, prepared and recorded by the agent; increases above the cap, any decrease on a performing customer, and any group-connected exposure go to the underwriter. Model changes go through validation, and the agent never holds credit authority.
- Stops when
- The policy owner pulls the kill switch, any cap is reached, a bureau or data feed is missing, or the population stability index (PSI) rises above its alert level: all actions queue for a person.
- Record
- Score version, cell, rule ID, action, notice template, and timestamp, per account.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Revenue from good customers' line growth, and faster drawdown answers.
- Lower exposure at default on deteriorating accounts, with an earlier hand-off to early warning.
- Fewer manual renewals.
What your institution measures
- Gini of the behavioral score
- Balance and revenue in increase cells against a control group
- 12-month bad rate of increased accounts against a control group
- Undrawn exposure removed before default
- Attrition of good customers
- Drawdown decision time
In Federal Reserve supervisory data on US credit cards, about 75–80% of credit limit increases were initiated by the bank rather than the customer, whether weighted by the size of the increase or by the number of cards.
Measured in the Decision Blueprint
A back-test of the matrix on 24 months sets the baseline; your institution sets the target, confirmed in a live champion/challenger.
Conditions
The rules, the data, and the known error, with its guard.
- The banking regulator's (BDDK) caps on consumer cards and loans sit above the matrix.
- SME revolving-line contracts must allow unilateral changes; the consumer protection law (Law 6502) governs consumer notices.
- Bureau enquiries are paid: batch monthly, daily only for watched accounts.
- Cuts are pro-cyclical in a downturn, hence the cap on the share of the book.
Start here
We back-test a draft matrix on your last 24 months of accounts, list which cells your policy could pre-authorize, and check the cost of the bureau refresh before anything is built.
Similar desks
- ModelsSLMsAgentic AI
TodayCalls reach customers who would have paid anyway, while the ones who cannot keep up need a plan, not ten calls.
With the systemHardship told apart from won't-pay, with the plan's arithmetic on file.
Collector decides
- ModelsAgentic AI
TodayMost alerts are false, and each one still takes an analyst real time to rule out.
With the systemEach alert arrives with one note; sanctions hits stay with a person.
Fraud analyst decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/behavioral-score-and-limits · Printed copy; the website is the current version.