Use casesFinancial institutions & fintech

  • Bank
  • Participation bank
  • Payment and e-money

Payment exceptions and investigations

A transfer comes back, sticks at an intermediary bank, or reaches a name that does not match; each inquiry means rebuilding the payment's path from messages and trackers before anyone can answer.

With the systemEach inquiry read, the payment traced through its messages, references matched; the officer answers and decides.

Who decides
Payments operations officer

Investigation request from the beneficiary bank · 4 November 2026

Request for information · customer transfer, euros

We are unable to apply the funds to the beneficiary's account. Beneficiary name does not match our account holder. Please confirm the beneficiary's full name and address.

→ same reference as the customer's transfer of 30 October

Illustrative The document as it arrives. In red, what the system marks for your expert.

1. The work

Today

  1. ReadInquiries arrive all day: a customer asking where a transfer is, another bank's request for information, a recall or cancellation request, a return, a query about charges; an officer reads each, the free text included, to see what is really asked.
  2. TraceThey find the original payment, then open the message history, the Swift tracker, the account, and earlier cases on the same payment, screen by screen.
  3. CheckReferences, amounts, currency, value dates, and charges are compared by hand across the messages, and the sender and receiver details against what the transfer rules require.
  4. WriteThe officer writes the answer to the other bank, a message to the customer, and the case note; a case waiting for the customer or a correspondent is chased by hand.
  5. DecideThe payments operations officer decides what to send and whether to return, recall, or ask for more; any sanctions or fraud sign goes to compliance or the fraud team.

With the system

  1. ReadThe SLM reads each inquiry, the free text of another bank's message included, and returns its type, the payment references, the amount, and what is asked.SLM
  2. TraceThe agent sends the case down the fixed path for its type and gathers, read-only, the original payment, every message on it, the tracker status, the account, and earlier cases.Agentic AI
  3. CheckYour written checks, run as code, match references, amounts, currency, value dates, and charges across the messages, and test the sender and receiver details against your rules.Your rule
  4. WriteThe SLM drafts the reply to the other bank, the message to the customer, and the case note in your formats, each fact tied to the message it came from.SLM
  5. DecideThe payments operations officer reads the file and decides, as before, what is sent and whether funds are returned or recalled; nothing leaves without them.Person decides

What lands on the desk

The officer gets the inquiry already read, the payment traced through its messages and the tracker, references and amounts matched by your own checks, and the replies drafted with each fact cited. The officer confirms with the customer, decides, and sends.

FIG. 1 · Investigation fileIllustrative
Outgoing customer transfer · euros · inquiry received 04.11.2026SME, textiles; supplier says nothing arrived

SLM · what is asked

  • CustomerAsks where the payment to the supplier isbranch email
  • Other bankBeneficiary name does not match its account holderinvestigation request
  • TypeBeneficiary query, payment held at the beneficiary bank

Agent · route: beneficiary query, read-only

  1. Original payment found, sent 30.10.2026payment hub
  2. Received by the beneficiary bank, not yet creditedSwift tracker
  3. No earlier case on this paymentcase history
  4. No sanctions hit at sendingscreening

Code · references and details

  • metPayment reference the same on every message
  • metAmount and currency match the order
  • metSender details complete under your rules
  • flaggedBeneficiary name: trade name, not the registered name

SLM · drafts

A message asking the customer to confirm the supplier's registered name, a reply to the beneficiary bank to send once confirmed, and the case note citing each message read. Nothing is sent.

  • SLM
  • Agent
  • Code

Why a person decidesEvery message to another bank or the customer, and any return or recall, is the officer's; a sanctions or fraud sign goes to compliance or the fraud team.

DecidesPayments operations officer

2. The system

For your technical team — skip it if you like.

How it works

In

  • Customer inquiries from branch and channels
  • Other banks' investigation requests
  • Recall, cancellation, and return messages
  1. SLM · reasons

    Inquiry reader

    Fine-tuned SLM; free text in Turkish and English

    Reads

    • Everything under In

    Out

    • Inquiry type
    • Payment references and amount
    • What is asked, and by whom
    • Confidence
  2. Agent · acts

    Investigation agent

    Router: one fixed evidence path per inquiry type; read-only

    Reads

    • The inquiry type and references
    • Your path for that type

    Out

    • The original payment message
    • Every message on the payment
    • Tracker status at each bank
    • Account and earlier cases
  3. Code · calculates

    Reference and amount checks

    Your written checks, run as code; no model

    Reads

    • References, amounts, value dates, charges
    • Sender and receiver details

    Out

    • References matched across messages
    • Amounts and charges reconciled
    • Missing transfer details flagged
  4. SLM · reasons

    Reply and case-note writer

    SLM, in your reply templates and case-note format

    Reads

    • The case file and its sources
    • Your templates

    Out

    • A draft reply to the other bank
    • A draft message to the customer
    • The case note, each fact cited
  5. Decision

    Payments operations officer

    Reads

    • The case file and the drafts

    Out

    • Decides what is sent, and to whom
    • Decides any return or recall
    • Sanctions sign goes to compliance
    • Fraud sign goes to the fraud team
How the agent's work is arrangedRouter

Router

The reader already names the inquiry type, and each type needs its own fixed set of evidence, so one routing step sends the case down the path written for 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

    200–400 past inquiries with the replies and case notes your senior payments officers graded: the type, what was asked, the facts used, and the answer.

  2. Dataset

    Past inquiries, the free text of bank-to-bank messages, the replies sent, and case notes; names and account numbers masked in training copies.

  3. Continued pre-training (CPT)If needed

    Only if the base model fails the test on payments vocabulary and the mixed Turkish and English of bank messages.

  4. Supervised fine-tuning (SFT)

    On reading inquiry types and on your reply templates and case-note format.

  5. Reinforcement learning (RL)Skipped

    Skipped: reading and writing in your formats are taught by fine-tuning, and no further reward is needed.

  6. Retrieval

    Retrieval supplies your procedures, correspondent terms, and message usage guides; references and amounts come from the code checks, never from the model.

  7. Monitoring

    The officer reviews every draft; a wrong type or a draft they rewrite joins the next test.

Where the agentic system stops

The payments operations officer decides what is sent to another bank or the customer and whether funds are returned, recalled, or released.

Show its tools, limits, and record
Tools and access
Reads your payment hub and message archive, the Swift tracker and case messages, core banking, earlier cases, and the screening results already on the payment; read-only; writes only the case file and the drafts.
May not
Send any message to another bank or the customer, return, release, recall, or reverse funds, post to an account, change payment data, dispose of a sanctions hit or a fraud alert, or grant itself any access.
Who decides and the written limit
Payments operations officerThe payments operations officer decides what is sent to another bank or the customer and whether funds are returned, recalled, or released. The system never sends a message, never moves or returns funds, and never disposes of a sanctions hit or a fraud alert: a sanctions sign goes to compliance, a fraud sign to the fraud team.
Stops when
Any sanctions or watch-list sign, a customer who says they were deceived, a recall for fraud, references that do not match, transfer details the rules require that are missing, or an inquiry the reader could not read with confidence: the case goes to the officer, marked; a sanctions sign also goes to compliance and a fraud sign to the fraud team.
Record
Every message and status read, the route taken, each check and its rule version, every draft, and who sent what and decided.

3. Value and conditions

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

Where the value comes from

  1. Inquiries answered sooner, with the facts in the first reply.
  2. Fewer rounds of messages between banks on the same payment.
  3. Officer hours freed for fraud recalls and the hard cases.

What your institution measures

  • Time from inquiry to answer
  • Cases the other bank or the customer chases again
  • Cases settled in the first reply
  • Officer time per case
  • Customer complaints about transfers
Published elsewhere
Swift estimates that exceptions and investigations cost the financial sector about USD 1.6 billion a year, driven by manual work, tied-up liquidity, penalties, and refunds.

Swift, as reported by PPI AG, "Rethinking E&I: The new Swift Case Management", July 2026 (vendor-reported) · vendor-reported

Measured in the Decision Blueprint

Today's time from inquiry to answer and officer time per case, from your case log, set the baseline; your institution sets the target.

Conditions

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

  • Under MASAK's measures regulation, a transfer message above the threshold it sets must carry the sender's name, account number or reference, and one identifying detail, verified, and the receiver's name and account number; a bank that receives a message without them returns the transfer or has the sending bank complete it, and every bank in the chain passes the details on (Tedbirler Yönetmeliği art. 24/1, 24/4, 24/6). The code check reads the current threshold from your rules.
  • Bank-to-bank investigations are moving to ISO 20022 messages through Swift's Case Management: an investigation request and its response (camt.110, camt.111), and a cancellation request and its answer (camt.056, camt.029). The agent reads them; only the officer sends one.
  • Any sanctions or watch-list sign stops the case for compliance, and the agent never disposes of a match; records are kept eight years (Tedbirler Yönetmeliği art. 46).
  • No machine-learning model is needed: whether a payment matches its messages is a check written in code. Messages and customer data stay on your servers.

Start here

We take a month of inquiries with your payments operations team, sort them by type, time each step from receipt to answer, collect your reply templates, and test the reader on your own past cases before anything is built.

Who builds it Every engagement is led by an engineer who has built machine-learning models that real decisions run on since 2012, and language models and agentic systems in production since 2024. About eqlx

Equalize the gap.

Bring work that takes expertise. Leave knowing the first step.

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

  1. You send a short descriptionNo confidential records, credentials, or production data
  2. The engineer checks whether it fits, and what it would needA model, a small language model, an agent, or none
  3. A proposed first step, or an honest “not a fit”A technical conversation with the engineer

https://eqlx.ai/use-cases/payment-investigations · Printed copy; the website is the current version.