AI agent · Capital markets · Read & extract
Corporate action capture
A corporate action notice holds a handful of terms that must be right before entitlements are calculated. The agent reads each notice, extracts those terms with a pointer to where they appear, and checks them against a second source before anything reaches your position system.
Typical volumes for this process, not a client figure.
Notices arrive as PDFs and get keyed into the position system.
Read, extracted, cross-checked against two sources, flagged when they disagree.
Where the time goes today
Corporate action notices reach you from custodians, paying agents, data vendors and issuers, often as PDFs or free-text messages, and often in several versions for the same event. Someone reads each one, identifies the event type and keys the terms into the position or asset servicing system: security identifier, ex-date, record date, pay date, ratio or rate, currency, options and their deadlines, and any conditions attached.
The work is not difficult, but it is unforgiving. A ratio keyed the wrong way round, a record date taken from an earlier version of the notice, or an election deadline missed because it sat on page three produces a wrong entitlement or a missed election. The error often surfaces only when cash or stock fails to arrive, weeks later. Voluntary events with several options, and events announced and then amended, take the longest and carry the most risk.
Most teams already compare two sources where they can. Done by hand for every event, that comparison tends to cover the fields that are easiest to see, not the ones that matter most.
How the agent works
- Collect each noticeThe agent picks up notices from the mailboxes, portals and feeds where they arrive, and groups those that refer to the same event and security.
- Extract the termsIt reads the notice and extracts event type, identifiers, dates, ratio or rate, options and deadlines, recording the page and passage each value came from.
- Cross-check sourcesEach term is compared with the same term from a second source, such as a vendor feed or the custodian's own announcement. Agreement is recorded; disagreement is shown with both values side by side.
- Detect amendmentsWhen a later notice changes a term, the agent identifies which fields changed and which positions and elections are affected, instead of creating a second event.
- Prepare the recordTerms that agree across sources are written to a draft corporate action record for approval, or straight to the system where your rules allow it for that event type.
What stays with a person
A person approves any record where sources disagree, where the notice is ambiguous, or where the event type is one your team always reviews: voluntary events with elections, events with tax consequences, unusual structures, or securities where you hold a large position. The agent shows both values and the passage each came from; the analyst decides which is right and, if needed, contacts the agent or custodian.
Elections, whether for clients or for the firm, stay with a person. The agent can track deadlines and remind the right people, but choosing an option is a decision, not data entry.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Custodian and paying agent notices, as PDF, email or message | A draft corporate action record, each term linked to its source passage |
| Data vendor corporate action feeds | A term-by-term comparison across sources, with disagreements flagged |
| Issuer announcements, where your team uses them | An amendment history showing what changed between notice versions |
| Your position system, to find affected holdings | A deadline list for elections and responses, by event and position |
| Events previously captured for the same security |
Controls that come with it
- No record is released unless two sources agree on the terms your team marks as critical, or a person approves it.
- Event types with elections, tax effects or unusual structures always go to review.
- Every extracted value keeps a link to the page and passage it came from.
- A weekly sample of released events is re-checked against the settled outcome.
- Records created by the agent are marked as such and can be withdrawn before entitlements are calculated.
How you know it works
- Extraction accuracy against past events with known-correct records
- Time from notice arrival to approved record
- Events where sources disagreed, and how each was resolved
- Entitlement errors traced back to capture, before and after go-live
Is your process ready?
- Written rules: your event types, the critical terms for each, and which source takes precedence are documented.
- Systems: notices can be collected where they arrive, and the position system accepts records through an interface or file your licence allows you to automate.
- Cheap check: a second source exists for most events, and the entitlement that follows confirms the capture.
- Volume: a steady daily flow of notices, most of them routine mandatory events, gives the agent enough work to repay the build.
- Consistent process: analysts agree on which version of a notice is authoritative when two exist.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Treating an amended notice as a new event, which produces duplicate entitlements.
- Ratios expressed differently by different sources, new for old or old for new, which look like disagreements when they are not.
- Poor-quality scanned notices, where the agent must stop and ask rather than guess a digit.
- Starting with complex voluntary events instead of high-volume mandatory ones, which delays any measurable result.
Questions we get
Which event types should we start with?
Start with frequent, well-structured mandatory events, such as cash dividends and simple splits, where a second source nearly always exists and the terms are few. Once extraction and comparison have been measured against your past events, move to mandatory events with options, then voluntary events. Complex reorganisations may stay with a person indefinitely, and that is a reasonable outcome.
What if we only have one source for an event?
Then the cross-check cannot happen, and the agent says so. You decide whether single-source events go to review by default, or whether some event types are simple enough to accept from one trusted source. Either way the record shows it was single-source, so nobody downstream assumes it was verified twice.
Can it read notices in several languages?
Language models read notices in most major languages, but accuracy should be measured, not assumed. During the build, the test set includes past notices in each language you receive, and extraction is checked against the records your team produced. A language where accuracy falls short of your threshold stays with a person until it improves.
Does it calculate entitlements?
No. That stays in your position or asset servicing system, which already does it. The agent's work ends at an approved, sourced event record. Keeping the calculation where it is means your existing controls on entitlements and payments are unchanged, and the agent can be switched off without affecting them.
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.