Use casesFinancial institutions & fintech

Request and complaint routing

Thousands of messages a day land in one queue; misroutes miss legal deadlines.

Kinds of AI
SLMs
Who decides
Head of operations
FIG. 1 · Intake routingIllustrative
Intake routingApp message · Turkish · 2 paragraphs

SLM · routing

  • CategoryCard dispute (unrecognized transaction) + possible fraud
  • DeskCard disputes, and the fraud team in parallel
  • PriorityUrgent: the customer reports the card is still active

SLM · suggested first step for the desk

Offer to block the card.

  • SLM

Written limitAuto-routes only above the confidence threshold; abstains and regulator referrals go to a named triage officer.

DecidesHead of operations

1. The desk

Today

  1. ReadApp messages, emails, call notes, regulator referrals, and branch forms arrive in one queue, and a triage team reads each one: complaint, fraud report, dispute, hardship request, data-subject request?
  2. RouteMisrouted items bounce between desks while the regulatory clock runs.
  3. DecideThe head of operations sees misroutes only as bounced items and missed deadlines; each desk decides the request itself.

With the system

  1. ReadThe SLM reads each message, scanned letters included, and returns its category, desk, priority, the regulatory clock that applies, flags, and a calibrated confidence.SLM
  2. RouteAbove the confidence threshold set for its class, the item is routed automatically; below it, with several intents, or as a regulator referral, it goes to a named triage officer.Your rule
  3. DecideThe head of operations decides the taxonomy, the thresholds, and which classes may auto-route, as before; the desk decides the request itself, and the model never closes one.Person decides

What lands on the desk

The desk gets the message already sorted: what it is, who owns it, how urgent it is, which deadline runs, and the flags that matter. Low-confidence, multi-intent, and regulator items go to a person instead of being guessed.

Routing is automatic only above the confidence threshold and only for routing; the head of operations owns the taxonomy, thresholds, and which classes may auto-route. Abstains and regulator referrals go to a named triage officer. Nothing is closed by the model.

FIG. 2 · Intake routingIllustrative
Intake routingApp message · Turkish · 2 paragraphs

SLM · routing

  • CategoryCard dispute (unrecognized transaction) + possible fraud
  • DeskCard disputes, and the fraud team in parallel
  • PriorityUrgent: the customer reports the card is still active
  • ClockDispute window; complaint answer deadline per policy

SLM · flags

  • Vulnerabilityno
  • KVKK requestno
  • Legal threatno

SLM · suggested first step for the desk

Offer to block the card.

  • SLM

Written limitAuto-routes only above the confidence threshold; abstains and regulator referrals go to a named triage officer.

DecidesHead of operations

2. The system

For your technical team — skip it if you like.

How it works

In

  • Message text: Turkish, English, or mixed
  • Scanned letters via text reading (OCR)
  • Channel and customer flags

  1. SLM · understands

    Routing model

    Fine-tuned to return category, desk, deadline, and flags

    Reads

    • Everything under In

    Out

    • Category, desk, priority, regulatory clock
    • Vulnerability, fraud, legal-threat flags
    • Data-protection (KVKK) request flag
    • A calibrated confidence

  2. Your rules

    Abstain threshold

    Set per class by the head of operations

    Reads

    • The routing output and its confidence

    Out

    • Above the threshold: routed automatically
    • Below it: human triage
    • Several intents, regulator referral: triage

  3. Decision

    The desk

    Reads

    • The routed item

    Out

    • The desk decides the request itself

None is needed: the model routes, and the ticketing system opens the item.

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. Model choice

    For single-label categories a fine-tuned classifier (BERT-class, such as a Turkish BERT) is often enough; a small language model (0.5–3 billion parameters) earns its place when the output is structured and policy-dependent (which clock applies, which flags, which clause).

  2. Test set

    1,000–3,000 recent messages labeled by senior triage staff with the final desk, clock, and flags; recall targets set separately for the fraud, hardship, and regulator classes.

  3. Dataset

    6–24 months of tickets labeled with the final resolving desk (not the first, possibly wrong, assignment), personal data redacted in training copies.

  4. Continued pre-training (CPT)Skipped

    Skipped unless the base model fails the test on vocabulary.

  5. Supervised fine-tuning (SFT)

    On the structured routing output.

  6. Reinforcement learning (RL)If needed

    Optional on hard classes once SFT plateaus; the reward is agreement with the final resolved desk and clock, which is reviewable.

  7. Monitoring

    Weekly drift; people review every abstain plus a sample of auto-routes; a taxonomy change triggers a retraining plan.

3. Value and conditions

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

Where the value comes from

  1. Time to the right desk.
  2. Fewer misroutes and missed deadlines.
  3. Triage hours freed.

What your institution measures

  • Misroute and bounce rate
  • Time to first touch at the right desk
  • Deadline breaches
  • Auto-route coverage at the target accuracy
  • Triage hours
Published elsewhere
A help-desk email ticket router in production at three major service providers assigned tickets with an accuracy close to 90% while covering at least 90% of email tickets.

Mandal et al., "Cognitive system to achieve human-level accuracy in automated assignment of helpdesk email tickets", arXiv 1808.02636

Measured in the Decision Blueprint

The baseline misroute rate comes from your reassignment history; your institution sets the target.

Conditions

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

  • Routing is not high-risk under the EU AI Act, but it touches complaint law: Turkish answer deadlines for financial consumers, and the EU joint complaint-handling guidelines.
  • Rare critical classes (fraud, vulnerability) are optimized for recall and never auto-closed.
  • The model runs on your servers, and its logs are redacted.

Start here

We measure today's misroute rate from your reassignment history, sit with your triage team on recent messages, and test a routing model on your own taxonomy before anything is built.

Request a Decision BlueprintWhat the Decision Blueprint covers
  • SLMs

    Corporate credit memo and financial statement analysis

    TodayAnalysts spread statements and write the memo by hand, in a fixed format, for every renewal.

    With the systemStatements spread with their sources, and the figures to check flagged.

    Credit analyst decides

  • ModelsSLMsAgentic AI

    SME credit referrals

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

    With the systemOne file with the rating, its reasons, and the authority level.

    Credit manager 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/request-routing · Printed copy; the website is the current version.