Alicia Ngomo Fernandez
Global AI Lead for Visa Consulting & Analytics, Visa Europe
Agentic AI deployments in financial services are no longer theoretical; they’re becoming a board-level priority.
Autonomous agents, multi-step orchestration and AI systems that don’t simply answer questions but execute work are rapidly moving from experimentation to implementation.
Despite the momentum, ROI remains elusive. Deloitte’s latest State of AI report found that while 74% of organisations hope AI will drive revenue growth, only 20% are seeing that outcome.1
Across client engagements, the success of agentic AI is rarely a model problem. It’s almost always a deployment and workflow design problem. As foundation models become increasingly commoditised, competitive advantage is shifting from raw model capability toward the operational infrastructure required to embed AI safely and effectively into the enterprise.
Too many organisations ask, “How do we deploy agents?” The better question is, “Which workflow should be redesigned to create measurable business value?” If the workflow is poorly selected, inefficient or badly understood, introducing agents simply accelerates dysfunction.
Start with measurable outcomes, not technology
The objective isn’t to “use AI”; it’s to improve the business. This requires disciplined value mapping alongside a realistic understanding of token costs, infrastructure overhead, human review requirements and exception handling. Around 30% of European banks cite cost reduction as their primary AI driver, though the opportunity to create entirely new value pools may prove equally important.2
Redesign the workflow before deploying the agent
Legacy workflows were built for humans navigating system constraints, not AI operating with structured decision logic. They’re often full of inherited approvals, duplicated reviews, unnecessary handoffs and fragmented context. Simply inserting an agent creates inefficiency, not transformation.
Leaders must ask which decisions are repetitive, where the bottlenecks sit, which approvals are necessary and where human oversight must remain essential.
Clear human-agent ownership
Teams need absolute clarity on who owns decisions, approves actions, handles escalation and is accountable when something goes wrong. Employees won’t trust systems where responsibility feels unclear, and leadership will hesitate to scale deployment if accountability is undefined.
Embed delivery resources close to the business
The strongest outcomes come from technical teams embedded directly alongside business operators. Forward-deployed engineers, embedded consultants and cross-functional delivery squads consistently outperform distant implementation models because they understand how work is executed. Agentic AI is not a plug-and-play software rollout; it’s operational transformation.
Security and risk are key
It cannot be treated as a post-deployment checklist. When autonomous agents access financial systems, handle sensitive data or execute transactions, the risk surface expands significantly. Identity, access control, fraud prevention, auditability and decision integrity become considerations from the start.
Clear security boundaries are essential around what agents can access, what actions they’re authorised to take and where human approval remains mandatory. An agent that can approve invoices or place orders without strong controls isn’t innovation; it’s unmanaged risk.
If the workflow is poorly selected, inefficient or badly understood,
introducing agents simply accelerates dysfunction
Continuous outcome tracking
Deployment isn’t the finish line; it’s the start of an optimisation cycle. Organisations need visibility into adoption, operational KPIs, token consumption, exception rates, model reliability and business impact over time.
Without this, it becomes difficult to distinguish transformation from expensive experimentation. The strongest AI programmes behave like optimisation engines, combining deployment with disciplined change management.
Common industry pitfalls
Organisations may consider avoiding pilots in search of a problem. Isolated experiments without executive ownership or measurable outcomes rarely scale. Token economics must also be taken seriously: workflows that look compelling in demonstrations can quickly become financially unsustainable at enterprise scale. CFO scrutiny of AI cost architecture is increasing accordingly.
Governance cannot be replaced by optimism. In regulated industries, teams typically focus on auditability, approval logic, compliance ownership and fallback mechanisms. Research among banking leaders shows that data privacy, security and compliance concerns outweigh even legacy systems as barriers to AI execution. 2
This article is provided for informational purposes only and doesn’t constitute legal, regulatory, financial or other professional advice. Readers should obtain appropriate independent advice for their specific circumstances.
At Visa Consulting & Analytics, we help organisations identify and deploy high-value AI opportunities. Our observation supporting financial institutions is clear: success in agentic AI depends on identifying high-value tasks, redesigning workflows, deploying responsibly, measuring outcomes rigorously and scaling with trust. That is where sustainable competitive advantage will be built.
[1]Deloitte. (2026). The state of AI in the enterprise. https://tinyurl.com/bp5u88kb.
[2] VCA. (2026). From AI promise to AI performance.
