End-to-end payments agents can unlock up to 60%

Complexity in payments is not one large problem, it's eight to ten smaller ones: initiation, verification and validation, enrichment and routing, sanctions and AML screening, exchange with the network, clearing and settlement, exception handling and repair, accounting and posting, and end-of-day reconciliation. Multiply complexity across payment types and the number of process variants grows faster than the number of products a bank sells.

This is precisely the profile that agentic AI is suited to: judgment-heavy but rule-bounded work spread across many systems and currently routed to a person because no single application owns the decision. It is also why single-process pilots stall: automating one of ten stages leaves the other nine.

With an end-to-end perspective, the opportunity to apply agentic AI in payments is substantial. Activity-level uplift ranges from 5–15% where the answer is process simplification, to 20–40% for traditional automation, to 40–80% where an agent can carry the work end to end. Weighted across an assessed function, that puts 30–60% of manual effort in payments back-office and digital cheque processing within reach.²

As complexity grows, conventional gains shrink

Payment operations teams have been running efficiency programs for decades, and most have done the work. That's exactly the problem.

What remains are the exceptions, the investigations, the repairs and the judgment calls that sit between stages. Our recent assessments of mature payment operations have found that further process re-engineering accounts for low single digits of the remaining addressable value, traditional automation for roughly 15%.³

Meanwhile, the payments ecosystem continues to evolve through SWIFT mandates, Real-Time Rail, Lynx, and ISO 20022. Richer and mandatory data will eventually improve screening, reconciliation, and STP, however in the near-term it means parallel standards, changes to reference data, and continued tuning of detection logic.

Agentic AI is the solution for the large majority of what remains.

Redesign the process before pointing agents at it

Agentic AI can increase the efficiency of payment operations, but only against a process that has been mapped from end to end. A pilot scoped to a single stage optimizes that stage and inherits every constraint around it. An end-to-end view is also the only way to take a defensible position on human intervention.  

Three levers for operating model redesign

Using this end-to-end view, we assess payment operations through three levers, applied in sequence. Each addresses a different kind of opportunity and delivers a different level of uplift.

1. Re-engineering

This lever yields 5–15% in potential uplift and is the prerequisite for everything that follows. It is also the most cost-effective source of value.⁴

A process redesigned for one line of business can typically hold for others with modest variation, because the underlying payment stages are the same. Designing the first implementation with horizontal expansion in mind turns a single-function return into a portfolio return.

2. Automation

Most banks have already used this lever extensively, but some opportunities still remain. Individually, each remaining opportunity is too small to fund, but the economics change when they are assessed together as part of an end-to-end redesign.

3. Agentic AI

This is where most remaining value sits because it is the only lens that reaches work requiring interpretation: exceptions, investigations, repairs, enrichment decisions and the constant reconciliation of imperfect data between systems. An agent can read an unstructured instruction, retrieve relevant context, apply policy, decide, and explain the decision.

Reuse makes it scale. The single most important design decision is how the agents are built as reusable capabilities. After a use case has been built as a reusable service, the subsequent use cases cost a fraction of the first.

Specialist agents alone do not scale. What turns a collection of capable agents into a platform is an orchestration layer. Shared enterprise services including rules and policy, audit and logging, identity and entitlements answer control questions at the platform level, allowing each agent to work together efficiently and effectively along with humans in the loop.

Controls belong at the design table. Banks can already build agents capable of running payment operations; the harder problem is getting them past the risk committee. An agent designed against the bank's stated risk appetite and control requirements is designed with the evidence required for approval.

Agents and the models they use should be designed against a control framework of observability, guardrails, orchestration, knowledge management, and access controls.

How Deloitte can help

The opportunity in payments is too significant to ignore: 30–60% of manual effort is addressable, with most reachable through agentic AI.⁵ The next step is finding where your bank sits inside this range and what it would take to move toward its upper limit.

Deloitte helps banks move from opportunity to enterprise-wide impact through a three-step process:

1. Ideate: Using the Deloitte Imagine tool to identify and reimagine the target state for payment operations.
2. Size: Assess the value and complexity of opportunities across payment processes to develop a defensible business case.
3. Scale: Build a reusable platform that extends across payments and adjacent domains to advance enterprise-wide objectives, supported by Deloitte’s Payments Workbench.

We also understand Canada's unique context, addressing Lynx and ISO 20022, the Real-Time Rail, ACSS and Interac. Elements of the regulatory perimeter like RPAA supervision, OSFI expectations, and SWIFT mandatories shape what can be delegated to an agent and the evidence required to support it.

Ready to see the size of your agentic opportunity? Let's continue the conversation.