Use casesFinancial institutions & fintech
Merchant onboarding and monitoring
New merchants sign up daily; one merchant set up to defraud or launder can cost more than many good ones earn.
- Kinds of AI
- ModelsSLMsAgentic AI
- Who decides
- Onboarding officer
SLM · identity, UBO, website
- metTrade registry
- metUBO chain: 2 levels, 1 natural person
- flaggedSite sells gift cards and top-ups: MCC mismatch
Model · risk score
Agent · options for the officer
- Option 1Onboard with a rolling reserve and a lower limit
- Option 2Decline
- Model
- SLM
- Agent
Written limitNever automatic: approval, decline, sanctions disposition.
DecidesOnboarding officer
1. The desk
Today
- ReadOnboarding officers read trade registry papers, tax registration certificates, authorized-signatory circulars, and the website by hand.
- ScoreThe requested merchant category code (MCC) is taken at face value, so a gift-card seller registers as electronics; after onboarding, monitoring looks at chargebacks only, after the loss.
- ScreenScreening, ownership, and links to closed merchants are pieced together by hand.
- DecideThe onboarding officer decides from whatever was gathered.
With the system
- ReadThe SLM reads the registry papers, ID documents, bank letter, and website pages into company facts, the ultimate beneficial owner (UBO) chain, and the goods actually sold, each with a citation.SLM
- ScoreThe merchant risk model scores chargeback, fraud, and insolvency risk and flags an MCC mismatch; for live merchants it runs monthly and alerts the risk team.Model
- ScreenThe agent runs sanctions and politically exposed person (PEP) screening, UBO resolution, a website crawl against the MCC, and link checks at once, and assembles one know-your-business (KYB) file; any hit goes to the officer.Agentic AI
- DecideUnder your policy, the onboarding officer approves, approves with conditions, or declines, as before; the risk team decides monitoring actions. Nothing is automatic.Person decides
What lands on the desk
The onboarding officer sees who owns the merchant, what the website really sells against the code it asked for, and every screening result, with the options set out. The decision itself stays with the onboarding officer.
SLM · identity, UBO, website
- metTrade registry
- metUBO chain: 2 levels, 1 natural person
- flaggedSite sells gift cards and top-ups: MCC mismatch
Agent · screening
- metSanctions / PEP screeningno match
Model · risk score
Agent · options for the officer
- Option 1Onboard with a rolling reserve and a lower limit
- Option 2Decline
- Model
- SLM
- Agent
Written limitNever automatic: approval, decline, sanctions disposition.
DecidesOnboarding officer
2. The system
For your technical team — skip it if you like.
How it works
In
- Trade registry, tax certificate
- Signatory circular, ID documents
- Bank letter
- Website pages
SLM · understands
Document and web reader
Reads
- Everything under In
Out
- Company facts
- Ultimate beneficial owner (UBO) chain
- Business description, goods actually sold
- A citation for each
Model · estimates
Merchant risk model
Calibrated GBM; live monitoring, transaction-laundering signals
Reads
- Facts, requested MCC, director links
- Early transaction data
- Live: transactions, refunds, chargebacks
- Live: website changes
Out
- Chargeback, fraud, or insolvency risk
- MCC-mismatch flag and reason codes
- Live merchants: alerts to the risk team
Agent · prepares
Screening agent
Parallel checks
Reads
- The facts and the score
Out
- Sanctions and PEP screening, UBO resolution
- Website crawl against the MCC
- Closed or terminated merchants, blacklist
- A know-your-business (KYB) file
Decision
Onboarding officer
Reads
- The KYB file and the score
Out
- The officer decides: approve, approve with conditions (a rolling reserve or a lower limit), or decline
How the agent's work is arrangedParallel
Parallel
Screening, UBO, web, and link checks are independent and must all finish before the officer sees the file.
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)
Test set
100–200 onboarding files graded by officers (UBO chain, business description, MCC fit).
Dataset
Past files with the officers' decisions.
Continued pre-training (CPT)Skipped
Skipped unless the base model fails the test on vocabulary.
Supervised fine-tuning (SFT)
On the KYB schema.
Reinforcement learning (RL)Skipped
Skipped: the KYB fields are checked by the test set and the officer's review.
Merchant risk model
Target
Chargeback, fraud, or loss within 6–12 months of onboarding.
Method and challenger
A gradient-boosted model against a logistic model.
Validation evidence
Precision at officer capacity, and calibration.
Monitoring
Monthly, with drift checks.
Where the agentic system stops
The onboarding officer decides every merchant; the risk team decides monitoring actions.
Show its tools, limits, and record
- Tools and access
- Read-only registries, the screening service, a web crawl in a sandbox, and the internal merchant database; writes the KYB file.
- May not
- Approve, decline, or dispose of a sanctions or PEP match.
- Who decides and the written limit
- Onboarding officerThe onboarding officer decides every merchant; the risk team decides monitoring actions. Approval, decline, and any sanctions or PEP disposition are never automatic.
- Stops when
- Any sanctions or PEP hit, an unresolved UBO, or a website that cannot be read: the file goes to the officer, marked incomplete.
- Record
- Every screen, crawl snapshot, and source document version.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Onboarding time for clean merchants.
- Fewer merchants set up to defraud or launder.
- Earlier monitoring signals.
What your institution measures
- Onboarding time
- Share of files complete at first review
- Chargeback and fraud loss rate in the first 12 months
- MCC mismatches found
- Alerts per confirmed bad merchant
Measured in the Decision Blueprint
Baseline onboarding time and the early-loss rate set the starting point; your institution sets the target.
Conditions
The rules, the data, and the known error, with its guard.
- Türkiye: Law 6493, the central bank's (TCMB) payment and e-money rules, and the obligations an obligated party owes MASAK, Türkiye's financial intelligence unit.
- EU: the Anti-Money Laundering Regulation (AMLR, 2024/1624) applies from 10 July 2027, with stricter customer due diligence and UBO rules.
- Web crawling respects site terms, and data stays on your servers.
Start here
We take recent onboarding files and merchants that later caused losses, and check what the documents and the website already showed at sign-up.
Similar desks
- 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
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
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/merchant-onboarding · Printed copy; the website is the current version.