Straight-through processing is the handling of a transaction from start to finish without anyone re-keying data or stepping in by hand: the instruction is captured once, checked automatically, and passed through each system to completion. Items that fail a check fall out of the straight-through path into an exception queue for a person to deal with.
Why does straight-through processing matter once AI agents act?
Classic straight-through processing runs on fixed rules. If the data is complete and valid, the item flows; if not, it drops out. The exception queue is where cost and delay build up, because every item in it needs a person to read, interpret and repair it.
Agents change what can be done with those exceptions. They can read an unstructured instruction, find the missing reference, or propose the right account from a name and a description. That is useful, and it is also different in kind. A rule gives the same answer every time; an agent applies judgement. When an agent repairs an item, it has not passed straight through. It has been handled by something new, which needs its own limits and its own record.
How does it work in practice?
- Capture: the instruction arrives in structured form, or is turned into structured data, often by intelligent document processing.
- Validate: formats, mandatory fields and reference data are checked.
- Enrich: missing standard data is added from reference sources.
- Control: limits, sanctions screening and fraud checks run.
- Execute: the transaction is posted or sent.
- Reconcile: the outcome is matched back to the instruction.
Any failure sends the item to exceptions. That is where an agent usually sits.
What does that look like in an example?
A bank's incoming international payments arrive with beneficiary details that do not always match an account exactly. Rules process the clean ones straight through. For the rest, an agent compares the name, address and reference with customer records and proposes the matching account, with the facts behind its proposal. Where the evidence is strong and the amount is small, the bank's limits let the agent apply the payment at depth 3. Where it is not, the proposal is held for an operations officer. Each repair is filed as agent-handled, not straight-through. Refunds work similarly, as in how to let an AI agent issue refunds through your payment gateway.
How is it different from processing in the data protection sense?
"Processing" in data protection law means any operation on personal data: collecting, storing, using, sharing. A straight-through flow that touches customer details is processing in that sense, and must appear in the firm's records of processing. The obligations are covered in GDPR Article 30 records of processing for AI agents, and whether an agent can help keep those records in can an AI agent maintain a record of processing activities? The shared word causes confusion in agent briefs; define it.
Which related terms should you know?
- Exception queue: the items that fall out of the automated path.
- Payment repair: correcting an instruction so it can complete.
How agent-handled exceptions fit among the other limits in a bank is set out in AI agents in banking.
Questions people ask
Does adding an agent raise the straight-through rate?
It reduces the exceptions people must handle, but report agent-handled items separately. Counting them as straight-through hides a judgement step that supervisors and auditors will want to see.
Should every transaction go straight through?
No. Some should stop by design, such as large payments or new payees, because a person's check is the point. Aim for straight-through on routine work only.