AI agent · Capital markets · Match & reconcile
Settlement break triage
A settlement break usually has an ordinary cause: an unmatched instruction, a partial delivery, a wrong account. The agent works out which cause applies to each break, gathers the evidence, and leaves your team the breaks where the answer is not obvious.
Typical volumes for this process, not a client figure.
Someone opens each break, checks three systems, and writes a note.
The software matches, classifies the break, and opens only the ones that need a judgement.
Where the time goes today
Each morning the overnight reconciliation between your book and the custodian's records produces a list of breaks: trades and positions where the two sides disagree on quantity, settlement date, account, counterparty or cash amount. Someone on the settlements team takes each break in turn, opens the trade in the booking system, looks up the custodian's status message, checks the counterparty's confirmation and writes a note on what they found.
Most of that time goes into finding things, not deciding things. The causes are familiar: an instruction sent to the wrong account, a partial settlement, a quantity that differs because a corporate action has not been booked, a counterparty that has not matched, a standing settlement instruction that changed. Establishing which one applies means reading three screens and an email thread. The notes are free text, written differently by each analyst, and hard to search when the same cause returns next week.
The order of work matters too. Breaks that must be fixed before today's deadline sit in the same queue as breaks that will clear overnight, and where late settlement carries a penalty or a buy-in risk, working the queue in arrival order costs money.
How the agent works
- Load both sidesThe agent takes your reconciliation output, or builds it from your book and the custodian's position and transaction files, and pairs records position by position and trade by trade.
- Gather the evidenceFor each break it pulls the trade record, the custodian status message, the counterparty confirmation and any related corporate action or instruction change, and keeps a reference to each.
- Classify the causeIt assigns the break to one of your cause categories and records which fields and messages support that choice. Where the evidence points to more than one cause, it says so rather than picking one.
- Order by deadlineBreaks are ranked by settlement date, value and the consequence of missing the deadline under your market's rules, so the queue opens on what must be fixed today.
- Act or hand overFor causes your team has cleared for routine handling, such as requesting a status update or chasing an unmatched counterparty, the agent acts and logs it. Everything else becomes a case with the evidence attached and a draft note.
What stays with a person
A person decides anything that moves money or changes a booking: amending a trade, cancelling and re-instructing, agreeing a claim with a counterparty, accepting a penalty. The agent can draft the amendment and show why it is needed, but the commit belongs to someone with the authority to make it. Breaks where the evidence conflicts, where the counterparty disputes the facts, or where the value is above your threshold also go to a person, with everything the agent read attached.
Your team also owns the cause categories. When a new kind of break appears, someone has to name it and decide how it is handled; until then the agent reports it as unclassified.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Your internal book of trades and positions | A classified break list, ranked by deadline and value |
| Custodian position and transaction statements | A case for each break that needs a person, with evidence and a draft note |
| Custodian settlement status messages | A log of the routine follow-ups the agent sent |
| Counterparty confirmations and affirmation records | A daily summary of breaks by cause, counterparty and age |
| Standing settlement instructions | |
| Corporate action records for the securities concerned |
Controls that come with it
- Only cause categories your team has cleared for routine handling are acted on without review; all others reach a person.
- Every break above a value you set is reviewed, whatever its classification.
- A weekly sample of classified breaks is re-checked by a senior member of the team.
- Each classification records the fields compared, the messages read and the rule applied.
- The agent has no permission to amend or cancel trades; it drafts, a person commits.
- Any category can be returned to full review at once if its error rate rises.
How you know it works
- Classification accuracy against past breaks with known causes
- Time from a break appearing to first action, by cause
- Breaks still open when their settlement deadline passed
- Analyst time per break, before and after go-live
- Breaks left unclassified, and the reason
Is your process ready?
- Written rules: your team can list its break categories and what it does for each. If they live only in people's heads, write them down first.
- Systems: the booking system and custodian files can be read by interface or file, and your custodian and vendor agreements permit automated access.
- Cheap check: whether a break was classified correctly becomes clear within a day or two, when it settles or does not.
- Volume: hundreds of breaks a day, most with ordinary causes, repays the build. A desk with a handful a week does not need an agent.
- Consistent process: two analysts investigating the same break check the same things in roughly the same order.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Cause categories that are too broad, such as 'other' or 'counterparty issue', which hide the pattern the agent is meant to reveal.
- Stale static data: a changed settlement instruction nobody updated, producing breaks that look new every day.
- Status message formats that differ by custodian and market, each needing its own mapping and tests.
- Allowing the agent to chase or re-send before its classifications are trusted, which repeats a wrong fix at scale.
Questions we get
Does the agent replace our reconciliation tool?
No. If you already run a reconciliation engine, the agent starts from its output and does the investigation that follows, which is where your team's time goes. If you do not, the agent can do the matching itself from position and transaction files. Either way, its value is in classifying the break and assembling the evidence, not in the comparison of two numbers.
How does it cope with several custodians and markets?
Each custodian's files and status messages are mapped during the build and tested against past breaks from that custodian. Deadlines and penalty rules are configured per market, not assumed. Adding a custodian is a mapping exercise rather than a rebuild, but it still needs its own set of past breaks before the agent handles its flow unsupervised.
What happens when a classification is wrong?
The case shows what the agent read and why it chose that cause, so the analyst who corrects it can see the reasoning. The correction is logged and added to the set of past cases used for testing. Because the agent cannot amend or cancel trades, a wrong classification costs time, not a wrong booking. A category that keeps failing goes back to full review.
Can it help with penalties and claims?
It records, for each break, the settlement deadline and whether the break was still open when it passed, which gives you the input for reconciling any penalties your market applies. Agreeing or disputing a penalty or a claim with a counterparty stays with a person. The agent prepares the file: trade history, messages exchanged, and the dates that matter.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.