Banks must fix their operating model before agentic AI can scale

Written by Russell Pereira

Jun 23, 2026
9 min

Practitioners from NatWest Group, UOB and Teleperformance argue that unclear processes, weak data foundations, governance gaps, poor value measurement and unresolved people impact now constrain agentic AI deployment more than the technology itself

Agentic artificial intelligence (AI) marks a shift from systems that generate text or follow fixed rules to systems that can plan steps, call tools and work across data and workflows. In banking, that difference matters because an agent may operate inside regulated processes that affect customers, controls and accountability. A bank therefore needs more than capable models; it needs clear processes, trusted data, access controls and human handover.
The clearest sign that a bank’s agentic AI programme will struggle to scale is not the quality of its models but the condition of its operating model. Raad Khraishi, head of artificial intelligence research and development at NatWest Group, pointed to the bank’s recent overhaul of its end-to-end complaints handling as an illustration. The team found that staff followed the process inconsistently, documented it poorly and interpreted it differently across the organisation, which meant the AI work exposed a deeper institutional problem before it became a technology problem.
NatWest used the AI build as a diagnostic rather than pausing first for a conventional process overhaul. "Rather than waiting for the process to be fixed," Khraishi noted, "we used our agentic AI solution as a means to understand and really articulate the process and really improve it, make it a bit more homogeneous, a bit more rigorous, a bit more reproducible." The deployment did not wait for institutional readiness; it helped create it by forcing the bank to make an ambiguous process more explicit, standardised and suitable for automation.
That example gives the wider debate its most practical starting point. Banks cannot safely automate a process that the institution itself cannot describe in a consistent way, because formal process maps, frontline workarounds, control requirements and system constraints often diverge. Agentic AI exposes those gaps because it must turn the process into executable steps, and if used poorly, it can scale errors, customer harm and conduct risk as quickly as it scales efficiency.
Models and agentic systems are improving quickly, so banks cannot wait for the technology to settle. Speed only creates value when the institution can preserve trust. That makes the state of the operating model a strategic constraint, not a technical implementation detail.

Readiness now means more than better technology
Alvin Eng, head of enterprise AI at UOB and leader of the bank’s AI Centre of Excellence, described the gap between the slogan and the work. "Having 'AI ready' is a nice slogan," Eng observed, "but there's a lot of work, a lot of preconditions, as you put it, that's required." He framed AI readiness as a recipe that requires leadership alignment, data foundations, governance, operating model redesign, workforce capability and fit-for-purpose technology.
That list matters because no single ingredient can compensate for the absence of the others. Strong leadership will not make poor data usable, better models will not fix unclear process ownership and a modern platform will not remove the need for access controls, auditability and escalation rules. The data requirement is also more demanding than many banks have recognised because agentic systems often need to work across documents, policies, scripts, case notes and other unstructured information, not only structured records such as accounts, balances, transactions and customer attributes.
"You almost need to build a knowledge layer for agentic consumption," Eng argued, pointing to ontologies and knowledge graphs as structures that agents need when they navigate unstructured information reliably. In banking terms, that knowledge layer gives the agent a map of what information means, how different concepts relate to each other and which sources it can trust. Without that layer, banks ask agents to act on an information base that may be fragmented, inconsistent or poorly governed, making the data problem a control problem as well.
Governance also needs to expand beyond traditional model approval. UOB already has mature risk management processes, but Eng said agentic AI introduces additional risks that banks must add to their taxonomy, including live monitoring of model inputs and outputs, as well as access entitlements when an agent calls external tools or uses information from other systems. That matters because an AI agent may retrieve data, trigger a workflow step, recommend an action or hand a case to another system, so the bank must govern what the agent can see, what it can do, how it explains its actions and how staff can challenge or reverse them.

Model choice becomes a governance decision
Once a bank allows agents to operate across processes and data, model choice becomes an economic and risk management decision rather than a technology chase. Khraishi argued that banks should not deploy the latest frontier model simply because it has become available, even though the impulse is understandable. Banks need success criteria, performance metrics and evaluation sets that show when a model change improves a use case and when it creates unnecessary cost, latency, bias, instability or customer risk.
NatWest Group’s research into model switching in multi-turn settings revealed a cost that banks often overlook. When a model is upgraded in the middle of an interaction, the context history generated by the previous model can degrade the new model’s performance. "Switching very frequently might incur hidden costs that might not make sense in many use cases," he observed. "It's more reasonable to have predetermined reflection points or trigger points where you really try to take into account when a switch might make sense."
That finding gives banks a practical rule. They should change models when evidence shows that the switch improves the outcome for a specific workflow, protects the customer experience and justifies the additional cost or operational complexity. Eng reached the same conclusion from a different angle, noting that UOB does not pursue a single best model because frontier models may justify their cost for complex coordination, while cheaper models may be sufficient for summarisation, extraction and other utility tasks.
For banks, the key capability is not access to one leading model but the ability to evaluate, compare and switch models under controlled conditions. "The process to allow you to evaluate new models as they come out, do all your regression testing," Eng noted. "That process is actually more important than the model selection itself." Andy Rangel, chief executive officer of Teleperformance (TP) Malaysia and Thailand, observed the same capability gap in his work with bank clients.
"Have you structured a team to help you manage this fast-moving, ever-changing environment?" Rangel asked. "And if you can't answer that solidly 'yes', then you may not have the right personnel or the right infrastructure." His point widens the issue from technology to organisation because banks need teams that combine banking, risk, operations and AI expertise.

Value only counts when banks can capture it
The same governance discipline must extend into value measurement. Banks cannot defend large AI investments, or expensive frontier model usage, unless they can show where the value comes from and how the institution captures it. This is where the gap between AI ambition and institutional practice becomes most visible.
Eng noted that UOB had deployed Microsoft Copilot to all 35,000 of its staff and measured an 8% to 10% improvement in time saved. The figure is meaningful, but it remains imprecise because the bank does not track every task or every way employees reinvest the time they save. "Surely there is some productivity leakage because I don't quantify and I don't go and track people what they use their time for," Eng acknowledged. "But we know there is a productivity uplift and we accept it."
That distinction matters for banks that deploy broad productivity tools across the workforce. They may know that staff save time, but they still need a clearer mechanism for turning that time into faster service, higher revenue, lower cost, better control or improved customer experience. More targeted AI investments allow harder measurement, such as control groups for revenue-generating use cases and average handling time for cost reduction.
These methods show that return on investment measurement is not separate from AI governance. It helps banks decide which use cases deserve funding, scaling and continued board support. Rangel pointed to a metric that matters especially in customer-facing agentic deployments: contact deflection, or the proportion of customer queries an agent resolves without escalating to a human.
The cost saving is direct because fewer interactions require staff intervention, but the bank must balance that saving against customer satisfaction and trust. A bank that deflects too aggressively may cut cost while frustrating customers who need human help, which means every efficiency gain carries a service-design question. "The sentiments and the analysis that you get from agentic solutions is far more comprehensive than what you get from a human talking to that same customer," Rangel observed. That data can reveal process friction, competitor references and customer satisfaction signals at scale.

The way forward is accountable autonomy
The next phase of agentic AI in banking will depend on whether institutions can move from experimentation to accountable autonomy. That means treating agents as supervised capacity inside the operating model, with clear ownership for the decisions they support, the processes they change and the outcomes they create. Banks that progress fastest will start where process volume, measurable value and controllable risk intersect, then scale only when the agent improves both performance and trust.
That approach changes the role of governance from a gatekeeping function into a continuous management discipline. Rangel described deployment as a clinical trial rather than a product launch, a useful frame for banks because it assumes permanent monitoring, adjustment and accountability. "You have to continue to watch the data," he noted. The moderator’s closing line captured the balance banks now need: "trust is strategy, speed is tactics." Agentic AI gives banks a way to move faster, but the institutions that scale it successfully will make autonomy observable, measurable and answerable inside the bank.