Ask a bank’s leadership team whether AI is working and most will say yes. In new research from Camunda, 87% of banking and financial services process decision-makers say AI is making their teams more productive. Among employees doing the work, that falls to 69%.
The gap shows up in daily work. Forty-five percent of employees say they have used AI to artificially comply with their organization’s AI mandate when there was no need to. Adoption dashboards count that activity as progress when it should really be graded on metrics that matter.
The cost lands in failed initiatives, rework and risk nobody designed controls for. Three in four leaders say process-related challenges have caused AI initiatives to fail, at an average cost of $1.73 million per organization.
Multiply that figure against your banks’ AI mandate and you can see how investments will be under even greater pressure to prove their worth.
Bank AI adoption metrics are measuring compliance with a mandate
Logins, prompts and deployed tools show that an AI program is moving. They count output. They can’t show whether work closed: a credit file reaching a decision sooner, or a KYC review needing fewer touches. How satisfied customers are and if they’ll recommend you to a friend – let alone if they plan to switch banks.
Wolters Kluwer’s Q1 2026 Banking Compliance AI Trend Report points the same way: about 31.8% of surveyed institutions have AI or machine learning in production, while 12.2% describe their AI strategy as well-defined and resourced. Deployment is running ahead of the strategy and operating model it depends on.
While it’s important to stay ahead, the gap between expectation and reality only grows without ensuring you are applying agentic AI to solve the right problems.
The rework lands on the people AI was meant to free up
When AI enters a process that wasn’t designed for it, employees absorb the difference. Among those not seeing significant time savings, 39% double-check AI output by hand because they don’t trust it. Nearly half of employees have overridden AI outputs because the process wasn’t set up correctly.
Half of organizations have rolled back AI because it hurt employees’ ability to do their jobs. Commonwealth Bank of Australia showed how far the two views can drift. In 2025 it cut 45 customer service roles after an AI voice bot it said had reduced calls by 2,000 a week. The Finance Sector Union said volumes were rising, with team leaders answering phones. CBA reversed the cuts and called the decision an error.
Every process becomes legacy once agents join the workforce
Most banking processes were designed with a person at every step. A commercial onboarding case moves from relationship manager to KYC to credit to legal, waiting at checkpoints built for human review. Payments exceptions and trade breaks follow the same pattern. A process built last year is legacy if it assumes a human-only workforce.
Add an AI agent and it inherits the same waits and handoffs. Seventy-three percent of decision-makers say implementing AI has exposed broken processes, and 81% say they need to fundamentally redesign them to stay competitive. At a large bank, origination alone runs as a separate version for each product, business line, and jurisdiction. The unit of change is the operating model: how work moves between people, AI agents and systems, and where decisions and controls sit.
Redesign starts with the business object the work changes: the loan application, the payment exception, the KYC case. An agent handed a copy of that object works from a snapshot of a record that keeps moving. An agent given bounded tasks against the live record, with decision rights written into the process, can close the work. Defining each object, which tasks may change it, and who decides is the operating model change.
Supervisors are asking about AI agents before the rulebook exists
In April, US regulators issued SR 26-2, model risk guidance that places generative and agentic AI outside its scope and leaves their controls to each bank’s own governance. This month, Reuters reported that the Fed, OCC and FDIC are already asking banks what limits apply to AI agents, what stops an agent from exceeding its authority, and whether decisions can be audited. Regulators are not yet prescribing answers. The ambiguity does not mean banks aren’t responsible for how their AI agents act. The Bank of England’s Financial Policy Committee has pointed to test incidents in which autonomous models took unexpected actions.
Answers to those questions have to come from the process, where an agent’s decision becomes an action. Thirty-eight percent of organizations had an AI-related compliance or governance issue in the past 12 months, and process-related challenges contributed to 81% of them. When controls sit inside the process, the record answers: which object the agent changed, which task it performed, who approved it.
When they sit outside, someone rebuilds the answer from logs spread out across 14 different systems.
Re-engineering the operating model has to happen at bank scale
Closing the gap takes what Camunda calls The Great Process Re-Engineering: redesigning how work gets done around what AI makes possible, with approvals and guardrails built into the process.
Yet, most banks are underfunding it. Thirty-six percent of banks allocate 10% or less of their AI spend to process redesign and the change management that supports it. Eighty-three percent say their AI investments will be a bust without more investment in process design. Eighty percent expect adapting all their most important processes for AI to take up to five years, four on average. Celent and Temenos research from September found that one in four banking customers worldwide have recently considered switching their primary bank. A four-year redesign timeline gives banks little room to respond.
ProcessOS, now in early access from Camunda, applies AI to the redesign itself. It reads operational telemetry, system logs and existing documentation to reconstruct how a process runs today, then generates a redesign aimed at the outcome you state, scored against cycle time, cost per case and compliance rate. Nothing reaches production without human review of that evidence. Once live, you can use execution data to keep improving the process as priorities change. And because Camunda’s platform is open by design, you can swap models and upgrade logic without interrupting the rhythm of business.
Start with the work your people already override
The 18-point gap between what leaders and employees report about AI productivity is the place to start. At the next AI program review, ask how often employees override the agent, how much rework follows, and which business object the agent changes in each workflow. Token costs and adoption counts supply none of that. Redesign the three processes that have the highest drain that aren’t focused on adding value before adding another agent.
A process that can show, for any case, what an agent changed, under whose authority and against which record gives a supervisor an answer and gives employees less to correct.Token cost and internal adoption metrics confirm AI is present. Cycle time, rework, and control evidence show whether the operating model has changed around it. Start with the processes where employees double-check, re-key, and override AI most often.
Get the full findings.
The AI Process Gap: Banking & Financial Services covers leaders and employees across the US, UK, Germany and France, including which processes organizations plan to redesign first. Download the report.
