Use casesFinancial institutions & fintech

Early warning on the loan book

The borrower is current; other banks are already cutting its limits, and you see its quarter only at renewal.

Kinds of AI
ModelsSLMsAgentic AI
Who decides
Watch-list committee
FIG. 1 · Early-warning fileIllustrative
Commercial borrower, textilesEarly-warning file
  • Proposed stageGreen → Watchfor the committee

Model · rating re-run with the Q2 interim

7was 4

IFRS 9: passed to the staging process as input; not decided here.

Agent · proposed actions, none executed

  • Freeze undrawn
  • Q3 trial balance and 13-week cash forecast
  • Management meeting
  • Model
  • Agent

Why a person decidesStage, action plan, and next review: the watch-list committee.

DecidesWatch-list committee

1. The desk

Today

  1. DetectThe early-warning list is a set of rules (30 days past due, a returned check) that fires too late and too often.
  2. ReadA firm whose sales fell in the second quarter still looks fine until the annual review.
  3. PrepareRelationship managers (RMs) answer alerts with "customer contacted, no issue".
  4. DecideThe watch-list committee gets a spreadsheet, not a file.

With the system

  1. DetectThe deterioration model scores monthly signals with reason codes, and the policy's hard triggers fire regardless of the model.Model
  2. ReadThe SLM reads interims, gazette notices, news, RM notes, and the last minutes into dated signals, each with its source; interim figures are marked unaudited.SLM
  3. PrepareThe agent checks the signal families in parallel, requests current interims through the RM or the portal, re-runs spreading and the rating, and builds the watch-list file with a proposed stage.Agentic AI
  4. DecideThe monitoring officer reviews the file, and the watch-list committee decides stage, action plan, and hand-over, as before.Person decides

What lands on the desk

The monitoring officer gets one file per name: why it fired, what each source says with its citation, the rating re-run on fresh interims, and the actions proposed but not executed. The stage, the action plan, and the next review are left to the watch-list committee.

FIG. 2 · Early-warning fileIllustrative
Commercial borrower, textilesEarly-warning file
  • Proposed stageGreen → Watchfor the committee

Model · why it fired

  • up: Model: low → high band
  • Hard triggers: none
  • Soft triggers: 3

The signals

  1. Turnover through our accounts down sharply vs. the same months last year
  2. New short-term cash loans at 2 other banks in 60 daysbureau
  3. Revolving line near full for 8 weeks
  4. Q2 interim, requested by the agent and received 17 days later: revenue down, EBITDA margin halved, unauditedinterim p.2–4
  5. Managing partner left the boardgazette notice
  6. RM note: largest export customer moved ordersCRM note

Model · rating re-run with the Q2 interim

7was 4

IFRS 9: passed to the staging process as input; not decided here.

Agent · proposed actions, none executed

  • Freeze undrawn
  • Q3 trial balance and 13-week cash forecast
  • Management meeting
  • Collateral review

Agent · alert history

  • flaggedRM reply “temporary”: no action documentednot accepted
  • Model
  • Agent

Why a person decidesStage, action plan, and next review: the watch-list committee.

DecidesWatch-list committee

2. The system

For your technical team — skip it if you like.

How it works

In

  • Monthly signals (per ECB guidance)
  • Turnover vs. the same months last year
  • Utilization, overdraft days, returned checks
  • Bureau, group, and sector signals
  1. Model · estimates

    Deterioration model

    Logistic champion, GBM challenger, discrete-time survival

    Reads

    • Everything under In

    Out

    • P(watch list or default) in 6–12 months
    • Reason codes
  2. Your rules

    Hard-trigger rules

    Written by policy; they fire regardless of the model

    Reads

    • The same monthly feed

    Out

    • 30 days past due anywhere
    • An enforcement notice
    • A concordat filing
  3. SLM · understands

    Signal reader

    Reads

    • Interim financials, auditor notes
    • Trade registry gazette notices, news
    • RM notes, last committee minutes

    Out

    • Dated signals, each with its source
    • Interim figures spread, marked unaudited
  4. Agent · prepares

    Watch-list file agent

    Parallel signal checks, then a tool loop with a stop rule

    Reads

    • The score
    • The hard triggers
    • The SLM signals

    Out

    • Current interims requested via RM or portal
    • Spreading and rating re-run on arrival
    • The watch-list file
    • A proposed stage, e.g. Green → Watch
  5. Decision

    Watch-list committee

    Reads

    • The file, reviewed by the monitoring officer

    Out

    • The committee decides the stage, the action plan, and the next review
    • IFRS 9 staging receives the file as input; it is never decided here
How the agent's work is arrangedParallel + Tool loop

Parallel

Signal families (account, bureau, legal, financial, group) are independent, so they are checked in parallel; once something fires, the next step depends on what comes back (interims received or not, rating moved or not), which is a loop over approved tools.

Tool loop

The agent picks the next tool from an approved list, reads the result, and repeats until a stop rule fires.

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

Deterioration model

  1. Target

    Entry to the watch list or Stage 2, or 90 days past due, within 6–12 months; features include deviation from the firm's own history and the age of its last financials.

  2. Method and challenger

    A logistic champion against a GBM challenger; a survival model if lead time is reported.

  3. Validation evidence

    Hit rate, false-alert rate, median lead time, Gini, and calibration.

  4. Monitoring

    Monthly, after three to six months in shadow beside the current list.

Small language model (SLM)

  1. Test set

    Interim statements and notes graded by monitoring analysts before training: field extraction, signal class, and date.

  2. Dataset

    Past interims, gazette notices, RM notes, and minutes, with personal data removed.

  3. Continued pre-training (CPT)If needed

    Only if Turkish financial vocabulary fails the test.

  4. Supervised fine-tuning (SFT)

    On the spread template and the signal schema.

  5. Reinforcement learning (RL)Skipped

    Skipped: there is no checkable reward beyond the extraction check, which SFT already targets.

Where the agentic system stops

The monitoring officer reviews every flagged name and prepares it; the watch-list committee decides stage, action plan, and hand-over, and the decision is documented.

Show its tools, limits, and record
Tools and access
Read-only access to core banking, the bureau (daily for watched names), the gazette, and the document store; may send the interim-financials request and reminders by template through the RM or the portal; may re-run spreading and the rating.
May not
Set a stage, book a provision, restructure, accelerate, or contact the customer outside templates.
Who decides and the written limit
Watch-list committeeThe monitoring officer reviews every flagged name and prepares it; the watch-list committee decides stage, action plan, and hand-over, and the decision is documented. The agent may apply only protective actions the policy names (for example freezing an undrawn line on a hard trigger, if the bank delegates it). Nothing reaches the customer automatically.
Stops when
Interims are not received by the policy deadline: the file escalates. Three loop rounds pass without new evidence: the file goes to the officer as it is. Any hard trigger fires: the file goes straight to the officer.
Record
Every signal, request, reminder, re-rating, and closure reason; no alert is cleared until a person records the action taken (ECB guidance on non-performing loans, Annex 7).

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. More lead time, so more cure options and lower loss.
  2. Fewer false alerts.
  3. A complete file for the committee, and a stage process supervisors can follow.

What your institution measures

  • Hit rate (defaults flagged at least six months before)
  • Median lead time
  • False alerts per eventual default
  • Share of alerts with a documented action
  • Cure rate of watch-listed names
  • Share of names with financials under six months old

Measured in the Decision Blueprint

A back-test on 24–36 months (when would each list have fired before each default?) sets the baseline; your institution sets the target, confirmed in a shadow run.

Conditions

The rules, the data, and the known error, with its guard.

  • Grounded in the EBA guidelines on loan origination and monitoring (EBA/GL/2020/06, section 8: monitoring, early-warning indicators, watch lists) and the ECB guidance on non-performing loans.
  • Interims are unaudited and can be dressed up, and the file says so; access to enforcement data is limited, so the system relies on the gazette, public notices, and the client.
  • Concordat filings cluster by sector, so portfolio indicators matter too, and alerts are capped per RM to avoid fatigue.
  • The system runs on premises, and every alert and action is logged for the banking regulator's (BDDK) inspection.

Start here

We back-test your current list and a draft model on past defaults, see how early each would have fired, and read a sample of interims and RM notes before anything is built.

  • ModelsSLMsAgentic AI

    The lease file

    TodayA dealer sends a proforma invoice; you must know if the lessee is good, the price is right, and the machine is not already financed elsewhere.

    With the systemThe lessee, the price, and other financing checked in one file.

    Credit committee decides

  • ModelsSLMsAgentic AI

    Merchant onboarding and monitoring

    TodayNew merchants sign up daily; one merchant set up to defraud or launder can cost more than many good ones earn.

    With the systemEach new merchant checked against your policy before anyone signs. Nothing is automatic.

    Onboarding officer 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/early-warning · Printed copy; the website is the current version.