Use casesFinancial institutions & fintech

SME credit referrals

Referred files wait while analysts rebuild the firm from four systems and last year's committee notes.

Kinds of AI
ModelsSLMsAgentic AI
Who decides
Credit manager
FIG. 1 · Credit referral fileIllustrative
SME, metal parts · group of 3 firmsRenew the cash working-capital line; raise the letter-of-guarantee line

Model · rating

5was 4 twelve months ago

  • down: margin
  • up: net debt / EBITDA

SLM · what the documents say

  • AuditorUnqualified; emphasis on a related-party receivableFS-2025 p.14
  • GazetteOne partner exitedgazette notice
  • BureauCash risk at other banks up; no arrears; check index one band lower

SLM · prior committee conditions

  • metPartner guaranteemet
  • not metQuarterly interim statementsnot met
  • Model
  • SLM

Why a person decidesAbove the engine's limit; a prior condition not met.

DecidesCredit manager

1. The desk

Today

  1. GatherThe engine refers the file with a one-line reason; the analyst opens core banking, the bureau, the trade registry gazette, and the shared drive by hand.
  2. ReadStatements are spread by hand and reconciled to the e-ledger and tax return; last year's committee conditions are checked only if someone remembers them.
  3. WriteThe memo is typed into the bank's format, and two analysts read the same firm differently.
  4. DecideThe credit manager receives an uneven file, and clean small files wait behind hard ones.

With the system

  1. GatherThe agent pulls the group's exposure and the documents through read-only connectors, and asks the relationship manager for anything missing.Agentic AI
  2. ReadThe SLM spreads the statements and reads last year's conditions, each figure tied to its source and page.SLM
  3. WriteMemo sections are drafted in the bank's format with open questions listed; the proposal is left to the analyst.SLM
  4. DecideThe credit manager receives one file with its rating, reasons, and authority level, and decides as before.Person decides

What lands on the desk

The analyst starts from a file that already cites its sources: the rating and why it moved, what each document says, the condition last year's committee set and whether it was met, and what is still open. The proposal itself is left to the analyst.

FIG. 2 · Credit referral fileIllustrative
SME, metal parts · group of 3 firmsRenew the cash working-capital line; raise the letter-of-guarantee line

Model · rating

5was 4 twelve months ago

  • down: margin
  • up: net debt / EBITDA

SLM · what the documents say

  • AuditorUnqualified; emphasis on a related-party receivableFS-2025 p.14
  • GazetteOne partner exitedgazette notice
  • BureauCash risk at other banks up; no arrears; check index one band lower
  • NewsNothing matched to this tax number; 2 name-alikes excluded

SLM · prior committee conditions

  • metPartner guaranteemet
  • not metQuarterly interim statementsnot met

Agent · authority from the matrix

  • Credit managerrule R-4.2, group band C

Agent · open questions

  1. Growth of the related-party receivable
  2. Q2 and Q3 interim statements
  3. The guarantor

Draft memo

Sections 1–7 drafted; the proposal is left to the analyst

  • Model
  • SLM
  • Agent

Why a person decidesAbove the engine's limit; a prior condition not met.

DecidesCredit manager

2. The system

For your technical team — skip it if you like.

How it works

In

  • Balance sheets and interims
  • The consented e-ledger (e-defter)
  • Tax returns

  1. SLM · understands

    Statement reader

    Schema-guided extraction

    Reads

    • Everything under In

    Out

    • Spread fields
    • Source, page, and quote for each
    • "Not stated" where a figure is absent

  2. Model · estimates

    Rating and PD model

    Logistic on ratios with a behavioral module; boosted challenger

    Reads

    • The spread
    • Account behavior
    • The bureau (KKB / Risk Merkezi)

    Out

    • Internal rating and PD
    • Reason codes
    • Change since the last review

  3. SLM · understands

    Document reader

    The same SLM: one reader task per document type

    Reads

    • Committee notes and conditions
    • Board resolutions, gazette notices
    • Sector research, news, the bureau

    Out

    • Dated, cited facts
    • Prior conditions met or not met
    • Contradictions flagged
    • Name-alikes excluded by tax number

  4. Agent · prepares

    File-building agent

    Orchestrator–workers with a checker

    Reads

    • The facts
    • The rating
    • Group exposure

    Out

    • Exposure recalculated in code
    • Policy checks
    • Authority level from the matrix (a rule)
    • Memo sections, open questions

  5. Decision

    Decision engine and credit authority

    Reads

    • The complete file

    Out

    • Inside policy and under the written limit, the engine approves and records
    • Every other file and every decline goes to the credit manager or committee
How the agent's work is arrangedOrchestrator–workers + Check and revise

Orchestrator–workers

Each firm brings a different set of documents, so a coordinator plans one reader task per document; a checker then ties every figure and condition in the draft to its source before release.

Check and revise

One step drafts; a second checks the draft against sources and rules, until it passes or a round limit stops it.

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

Small language model (SLM)

  1. Test set

    100–300 past files graded by senior analysts before training (extraction per field, condition checks, memo sections).

  2. Dataset

    Decided files with personal data removed; extraction pairs and accepted memo sections.

  3. Continued pre-training (CPT)If needed

    On Turkish banking and accounting language (the Uniform Chart of Accounts, TFRS and tax-law statements), only if the base model fails the test on vocabulary.

  4. Supervised fine-tuning (SFT)

    On the bank's spread template and memo format.

  5. Reinforcement learning (RL)Skipped

    Skipped unless figures still drift from their sources after SFT; then the reward is "every figure matches its cited source", which can be checked.

Rating model

  1. Target

    12-month default: 90 or more days past due, or unlikely to pay.

  2. Method and challenger

    A logistic champion against a gradient-boosted challenger with monotonic constraints; no method is presumed better.

  3. Validation evidence

    Gini, calibration by grade, stability, and override analysis, prepared for the bank's own validation unit.

  4. Monitoring

    Monthly, after go-live.

Where the agentic system stops

Where the bank's credit policy allows automatic approval, the bank's own engine approves files inside policy and under its written limit, and the file is recorded.

Show its tools, limits, and record
Tools and access
Read-only connectors to core banking, the bureau, the gazette, news, and the document store; runs spreading and rating through the bank's own services; may request missing documents from the relationship manager by template.
May not
Approve, decline, price, waive a condition, contact the customer, or add a data source.
Who decides and the written limit
Credit managerWhere the bank's credit policy allows automatic approval, the bank's own engine approves files inside policy and under its written limit, and the file is recorded. Every other file, and every decline, goes to the credit manager or the committee the board's authority matrix names. The agent never holds credit authority.
Stops when
A required document is still missing at the reminder deadline, the checker cannot tie a figure to a source after two revise rounds, or a name match falls below its threshold: the file goes to the analyst, marked incomplete.
Record
Every tool call, document read, model version, checker verdict, and analyst edit.

How agentic systems stay inside limits

3. Value and conditions

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

Where the value comes from

  1. Analyst hours per file, and time to decision.
  2. Fewer document requests to the relationship manager, with prior conditions tracked on every file.
  3. Consistent memos, and hidden weaknesses (related-party flows, partner exits) found before committee.

What your institution measures

  • Median and 90th-percentile time from complete application to decision
  • Analyst hours per file
  • Document requests per file
  • Share of files with every prior condition checked
  • Committee return rate
  • 12-month watch-list rate of approved files, by grade

Measured in the Decision Blueprint

A time study of 30–50 recent files sets the baseline; your institution sets the target, confirmed in a shadow run on the same files.

Conditions

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

  • The banking regulator's (BDDK) information-systems rules and KVKK Art. 9 keep the system and its data in Türkiye, on your premises.
  • e-ledger access needs the firm's consent and a confirmed channel; banking secrecy under Banking Law Art. 73 covers committee notes, and access follows the credit-file role.
  • Name disambiguation is the main reading error.
  • Analysts may stop reading sources: red flags require a citation click, and a quality sample is reread.

Start here

We sit with your analysts on recent referred files, map the documents and systems each one touches, and test the reader on your statements before anything is built.

Request a Decision BlueprintWhat the Decision Blueprint covers
  • ModelsSLMsAgentic AI

    Early warning on the loan book

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

    With the systemA file, not a spreadsheet: signals checked in parallel and gathered in one place.

    Watch-list committee decides

  • ModelsSLMsAgentic AI

    Early arrears collections

    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

Equalize the gap.

Bring one decision. Leave knowing the first step.

A short description of the decision is enough; no confidential records.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/use-cases/sme-credit-referrals · Printed copy; the website is the current version.