Banking’s Exception Queue Is Its Most Valuable Unmanaged Asset

November 10, 202607 min Read

Two decades of payments modernization have pursued one goal above the rest, which is to move every payment from origination to settlement without a person touching it. Straight-through processing became the headline metric, and each new standard, rail and platform has been justified by its effect on that number. The payments that fall out of straight-through processing receive far less attention. They sit in the exception queue, where a person decides what a failed payment should have said and whether it is safe to release.

That queue is smaller than it once was and harder than it has ever been. The argument of this article is that it is also the most valuable and least governed operational asset in a bank's payments business, and that the next phase of automation will be decided by who holds the record of what happens inside it.

Exceptions Cost More Than the Repair

A survey of 150 global financial institutions by Datos Insights found that a third of banks say between 1 and 3 percent of their domestic payments incur charges from a banking partner because of payment failure, 24 percent say 4 to 5 percent, and 11 percent say 6 to 8 percent. For cross-border payments the figures are higher, with 23 percent of banks reporting that 12 to 15 percent of payments incur such charges and 18 percent reporting that it happens more than a quarter of the time (Datos Insights, 2025, as reported by American Banker).

The direct cost of a domestic repair typically runs between $6 and $10, yet 57 percent of banks charge customers $5 or less for it. The indirect cost is larger. Forty-five percent of banks estimate that customer attrition caused by failed payments costs them between $1 million and $5 million (Datos Insights, 2025, as reported by American Banker). The exception queue is therefore a cost center, a customer experience problem and a risk surface at once, and most banks measure only the first.

Better Standards Remove Errors but Not Judgment

ISO 20022 was meant to shrink the queue by carrying richer, structured data in every payment message. It is working in part. Seventy-two percent of banks report higher straight-through rates because fewer payments fail on missing or incomplete data. Yet 65 percent of banks still rely on manual processing, and analysts caution that structured fields improve consistency without giving a bank any additional ability to match data that was entered incorrectly (Datos Insights, 2025, as reported by American Banker).

Swift's own roadmap reflects the same reality. With the end of coexistence between legacy and ISO 20022 messages for cross-border payments in November 2025, the next phase moves exceptions and investigations onto structured case messages, with community adoption required by November 2026 and all institutions on the new messages by the end of 2027 (Swift, 2026). The conversation between banks about a troubled payment is becoming data.

"As straight-through processing rises, the exceptions that remain are the ones that carry judgment."

This is the point most modernization programs miss. A payment that fails because a field is empty can be fixed by a rule. A payment that fails because a beneficiary cannot be matched, a remittance cannot be reconciled or a counterparty looks unfamiliar needs someone to decide. As the easy exceptions disappear, those that remain are rarer, higher in value and more dependent on what the bank knows about its customers, its counterparties and its own appetite for risk.

Faster Rails Raise the Stakes

Instant payment networks make the effect sharper. FedNow reported that its transaction volume grew 460 percent in 2025, that nearly all of the nation's leading institutions now participate, and that its average payment reached $101,435. The Clearing House reported more than $1.3 trillion processed on RTP in 2025 after raising its transaction limit from $1 million to $10 million (PaymentsJournal, 2026). Payments that move faster and carry more value leave less time to decide, and each decision made under that pressure carries more consequence.

The Decision Record Is the Unmanaged Asset

Every exception resolution encodes the bank's judgment. It records which fix was proposed, which rule was checked, who decided, what they overrode and why. In most banks this lives in a ticket note, a workflow screen or a person's memory. It is rarely structured, rarely reviewed and almost never used to improve anything beyond the next ticket.

Agents change what that record is worth. A repair agent improves by learning from the decisions people make about the exceptions it sees, which means the record is also the training material for the capability the bank is adopting. The bank that holds the record can improve its own operation with it. The bank that does not can only use whatever the system returns.

Supervisors have not yet written the rule for this. SR 26-2, issued in April 2026, replaces the model risk guidance in SR 11-7 and SR 21-8, states that generative and agentic AI models are outside its scope, sets no enforceable standards and is most relevant to institutions above $30 billion in assets. For tools it does not cover, an institution's own risk management and governance should guide the controls, and for vendor products the principles still apply even where a bank may not receive code, data or methodology (Board of Governors of the Federal Reserve System, 2026). Until supervisors say more, the only evidence a bank can show about an agent-assisted release is the record it kept itself.

What Banks That Treat the Queue as an Asset Do Differently

They treat each resolution as structured data rather than a note, capturing the proposed fix, the rules checked, the person who decided, the override reason and the outcome. They keep that record in an environment they control and in a form they can inspect. They decide in advance where the learning from their overrides is allowed to go, so that the improvement an agent gains from their decisions accrues inside their own boundary. And they ask of every repair agent, whoever supplies it, whether they could show a supervisor the fix, the rule and the person for a single payment without asking anyone's permission.

Buying Fast and Owning the Record Are Compatible

The strongest objection is that vendor modules ship faster than anything a bank can build, and that a team facing a growing queue cannot wait. That is true, and buying is often the right decision. The choice, however, is not between speed and ownership. The record can live inside the bank when the module is bought, if it is written to the bank's environment in the bank's own schema, and if the arrangement makes clear in advance what leaves with the bank if it changes provider. The decision records, the governed context the agent works from and the operating knowledge of how exceptions are resolved are the parts that matter at exit.

Deferring the question has a cost that compounds. Every month of releases produces learning, and learning that accumulates outside the bank's boundary is difficult to bring back. The exception queue will keep shrinking. What the bank retains from it will determine how much of that shrinkage it can claim as its own.

Where to Begin

Banks and payments operators assessing where their exception records live today can begin with one conversation about it. Start the Conversation