Settlement got faster and investigations did not
Swift, which carries payment instructions between banks around the world, reported in 2025 that cross-border settlement times are a quarter of what they were five years ago. In the same announcement it said investigation times are unchanged. When a cross-border payment goes wrong, the investigation still takes five to ten working days, and those investigations cost the industry more than 1.6 billion dollars a year.
The rails got faster. The part that explains what happened on them did not.
Surveys this year show the back office still runs by hand
Two industry surveys published in 2026 show the same pattern from inside institutions. Both were run by software vendors, so read them as indicators rather than census data.
Aqua Global surveyed 150 European bank IT leaders in July 2026:
- 74% named manual handling as the biggest obstacle to real-time payments.
- 13% were fully automated across core payment functions.
- 29% take more than a day to resolve a reconciliation exception, the case where two records of the same payment do not match.
AutoRek surveyed 250 finance managers in the UK and US in February 2026:
- 69% said manual processes are their biggest constraint on scaling.
- 67% said instant payment rails are increasing the need for real-time controls.
Instant payments, domestic schemes, digital wallets and open banking all arrived within a decade. The operations that have to agree with all of them are still held together by people.
The work sits between systems that were never built to agree
A single payment is recorded by several systems at once: the processor that handles it, the bank that settles it, the ledger that books it and the compliance system that screens it. Each one is correct on its own terms, and each holds only its part.
When those records disagree, nobody's software owns the disagreement. Someone opens several systems and rebuilds the story by hand. Most of that work comes down to two questions, asked all day long. Why doesn't this record agree with that one? And: I expected this to happen, so why did something else happen instead? The settlement due at noon that has not arrived. The fee charged that is not the fee agreed.
Every new payment rail, partner and regulatory obligation adds more records that have to agree, so the work grows with the business.
Institutions spend an estimated $530 billion a year on that work
Cordant's own analysis puts the annual cost of operations teams inside banks and payment companies worldwide at about $530 billion, within a range of $450 to $650 billion. These are the people who reconcile accounts, chase failed payments, work compliance cases and close the books. By the same analysis, roughly five dollars in every six of that is salary rather than software.
The figure is an estimate built from public bank cost data and industry research, not a measured total. What it describes is plain enough: most of the cost of making financial systems agree is paid in people's time.
Agents run into the same problem people do
Financial institutions are now piloting AI agents in exactly this work. The early evidence says the agents hit the same obstacle the people do.
Deloitte surveyed 501 US senior managers and executives whose organizations are already piloting agentic AI, between April and June 2026, and published the results on 12 August:
- 72% lack unified, accessible data across their systems.
- 70% cannot adequately trust and govern their agents.
- Only 5% say their business processes are highly prepared for agents.
Gartner put the reason directly in May 2026. In the words of Rita Sallam, Distinguished VP Analyst: "Without context, a clear understanding of the specific relationships and rules within an organization's data, AI agents cannot operate accurately and are far more likely to hallucinate, introduce bias and produce unreliable results."
An agent can automate work inside one system. Once the work crosses systems, it needs what a person needs: a reliable answer about what happened and what was supposed to happen. The firms in that survey are not short of models. What most of them lack is the part that knows what happened across their systems.
Financial infrastructure has solved this kind of problem before
The industry has faced fragmentation like this several times, and each time a new layer arrived above the fragmented systems:
- 1970, card networks. Thousands of separate card programs came under one network.
- 1973, SWIFT. Banks sending telex messages in their own formats moved to one shared messaging network.
- 1973, the Depository Trust Company. Paper share certificates moved by hand gave way to one central depository.
- 2002, CLS. Foreign exchange trades settled one side at a time moved to a system that settles both sides together.
Each layer left the underlying systems in place and gave the industry one reliable account above them. The fragmentation in today's back office is creating demand for the same kind of layer, and the arrival of agents makes the need more urgent.
Ask your operations team two questions
If you run, fund or oversee a financial operation, ask your team two questions this week. When something goes wrong, how many systems do they open to find out why? And how long does it take to resolve an exception from the moment it is found?
If the answers are "several" and "more than a day", you are in the same position as nearly a third of the banks in the Aqua Global survey. The order that fixes it is the same whether the work is done by people or by agents: first know what happened across every system, then check it against what was supposed to happen, then send the difference to the person who owns it with the evidence attached. Agents come last, within limits a person sets.
Money already moves in real time. The decisions should too.
