Payment exceptions and investigations
Reads payment messages (pacs, camt, MT), the ledger, investigation requests. Produces resolved cases, replies to counterparties. Code works out every figure, the Judgement Engine answers the narrow questions, and whatever stays uncertain goes to a person.
Narrow questions, each with a check behind it.
Why a payment failed or is queried; the action
Proven by code: message chain, camt/pacs, ledger. Code proves the answer from the data itself: totals tie, a match is exact, the quote is on the document. Most of these judgements settle without a person.
| Reads | Payment messages (pacs, camt, MT), the ledger, investigation requests |
| Produces | Resolved cases, replies to counterparties |
| Who signs | A person on your team, with every figure traced to its document and every call on record |
| Process | Claims, disputes and exceptions |
| Industry | Banking and payments |
Every line, decided where it's safest.
An example run. Each stage lights up as a line is decided there; pick a stage to see its lines.
| Ref | Item | Value | Decided by |
|---|---|---|---|
| INV-551 | pacs.004 return, AC04Account closed: return matches the original | 12,000.00 | Code · matched |
| INV-552 | Beneficiary says not receivedCredited 02.09 per camt.054 · 0.96 | 8,450.00 | Engine · proven |
| INV-553 | Duplicate MT103Same reference and amount 3 minutes apart · 0.94 | 25,000.00 | Engine · proven |
| INV-554 | Charges dispute OUR/SHACorrespondent charged twice? Unclear | 45.00 | To a person |
| INV-555 | Wrong beneficiary nameName mismatch: recall needed | 3,200.00 | To a person |
|
Settled 3
To a person 2
Settled wrong 0
| |||
Your volumes in. Your hours out.
In our tests, between 24% and 42% of items still went to a person after tuning. Set your own share; a pilot measures it on your data.
Start from what's closest.
Claims, disputes and exceptions in other work
Customer deductions and short-pay disputes
Valid deduction (promotion, damage) or not
BankingCard disputes and chargebacks
Network reason code; evidence enough to fight
BankingCustomer complaints handling (UK FCA)
Complaint category, redress due
More for banking
What buyers ask
What happens when Sumarity isn't sure?
Nothing settles below the cutoff, or when the check disagrees. The item goes to a person in the inbox with what the engine saw, its best answer and the runner-up. The person's ruling is kept and tunes the next run.
How are its judgements checked?
Code proves the answer from the data itself: totals tie, a match is exact, the quote is on the document. Most of these judgements settle without a person.
Can we change it to fit how we work?
Yes. Start from this workflow and describe your differences in plain words. The design assistant revises it, the validator checks it, your expert reviews it on the canvas, and your admin publishes it.
Does it read our files as they are?
Yes: spreadsheets and CSV in any layout, PDFs, bank formats and e-mail attachments. Sumarity suggests how each column maps, proves the mapping on a sample and remembers it once a person confirms it.
Where does our data go?
Sumarity runs in Zurich. Your data, backups and logs are stored in Switzerland.
How does it get better?
Your team's rulings become an answer key. Better questions and cutoffs are proposed, tested on data they haven't seen, and published only when your admin approves. It improves itself, with permission.
Start from this workflow. Run it on your data.
A pilot runs alongside your own process for a few weeks, at our cost, and ends in a line-by-line comparison.