The operational context imperative: Why banks are failing to scale AI (and how they can)

Here’s a modern cautionary tale. A bank decides to use AI to process its payment exceptions more efficiently. It begins by introducing an AI agent to handle payment repairs.

But the agent doesn’t understand the steps its human investigators take. It’s also unaware that payments from this customer, on this channel, typically feature the same handful of issues. As a result, human beings keep having to override the agent’s suggestions.

The bank struggles to know whether AI is actually making payment repairs more efficient— though it’s keenly aware of AI’s spiraling costs, now that the vendor has introduced consumption-based pricing. Quietly, plans to scale AI agents to other parts of the payment journey, and the wider business, begin to stall.

This is a story with an important moral, especially as we all head to Sibos 2026 to discuss, in real, practical detail, how AI is reshaping global finance.

Banks continue to make substantial investments in Enterprise AI. Large US institutions will, on average, invest $170 million over the next twelve months, according to a recent study by KPMG. But most banks still have no way to answer some fundamental questions:

• Is our AI agent making the process better?
• Have we reduced risk, or shifted it?
• At the end of the day, is this agent saving or costing us money?

“When banks can answer questions about process improvement, risk, and ROI, they can begin to use AI, at scale, to deliver against pressing objectives,” said Chris Johnston, SVP and Head of Global Banking at Celonis. “They can use AI to help drive T+1 settlement. They can use it to find new operational efficiencies, and make demonstrable progress against ESG commitments.”

But banks don’t get to that point by investing in AI alone.

A newer (and often more expensive) frontier model may improve an agent’s ability to reason, but it won’t show an agent how the bank operates. It won’t make the agent’s decisions easier to audit, or its costs easier to manage and justify. And it won’t help the bank to build a harmonious, mutually enabling workforce of AI agents and expert people.

At this moment in time, banks don’t need more capable AI models. They need AI-ready architecture.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:5a5fa9f4-7b65-4091-8d93-5fb12ef859cf/original/as/2026_celonis_329_Sibos-Banner-Social-Image_01_04.png

Legacy data architecture is holding banks back

Today’s core banking systems were designed decades ago. They were built for the needs and priorities of that era: for processing batches and controlling risks, for stability rather than for speed. They weren’t built for agility, adaptability, and the pressures of modern banking. And they certainly weren’t built for AI agents.

Implementing AI agents directly on top of legacy systems can cause a bank’s AI spend to explode. Every deployment requires bespoke integrations and, in the absence of a mechanism for gaining oversight of agent actions, it’s easy for token usage to soar.

What’s more, putting a data lake or warehouse between your AI agents and your core banking systems isn’t an answer, at least, not if you want your agents to be effective and accountable.

Data warehouses work well for human analysts, who can use their experience to navigate ambiguity. They can draw the right conclusions from different data points, because they understand the context. AI agents, on the other hand, don’t have that context built in. They must be shown how a process like payment repairs actually runs, as a series of sequential steps, variations and all.

“Banks, like all companies using AI and even those selling it, are waking up to the realization that AI needs operational context to deliver meaningful value,” Johnston said. “Without that context, it’s simply not enterprise-ready.”

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:c934078b-1470-49e5-b7c1-14db7bbd188d/original/as/banking_ceo_ebook_promo_card.png
https://www.celonis.com/insights/ebooks/create-a-banking-center-of-excellence

What banks need to create (and act on) operational context

Some banks are already solving this problem by adding a new, contextual layer to their data architecture. This layer combines a real-time, operational digital twin with agent guardrails, orchestration, and observability.

A real-time digital twin of banking operations

The digital twin captures every process, dependency, control point, and variance. It’s created by continuously mining event logs and user interactions across the business. For end-to-end payment processes, that can mean payment origination systems, payment rails, settlement platforms (i.e., every system the process touches).

The digital twin shows AI agents how payment processes have run in the past, how they’re running now, and what will happen if they take this or that action. It also helps banks to identify and prioritize their lowest-risk, highest-value AI use cases.

Agent guardrails and orchestration

The contextual layer also helps banks to circumscribe and coordinate agent activity. Data privacy controls ensure agents only see what they need to see when monitoring payments. Guardrails help to define the limits of their autonomy, mandating, for example, when payment repair decisions need to be referred to a human for review.

But banks also need to be able to orchestrate agents, simple automations, humans, and workflows in harmony. As a single source of truth for end-to-end processes, this contextual layer is the natural location for that control tower (not least, because it captures the actions of agents themselves).

Continuous agent observability

This is crucial. It’s possible to observe agents in real time, within the contextual layer, because they’re part of the operational reality being captured by the digital twin. Every action they take leaves a trace that can be mined alongside every other step in the process.

When banks can observe the actions of their AI agents, in real time, they can see the moment they deviate from a defined process. They can make sure agent actions are auditable, at scale, helping banks continue to comply with the principles of BCBS 239, and stay on the right side of evolving AI and model risk management guidance from regulators. And, importantly, they can evaluate whether any AI agent is genuinely profitable, by linking its actions and token usage to downstream effects.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:8e7a1f96-7433-4b1e-acc9-959bfbbdc488/original/as/sibos_2026_chris_ann_lauren_daniel_headshots_updated.jpg

Five challenges operational context solves

Once banks have put this contextual layer in place, they’re not only able to deploy agents to support payment repairs, insufficient-funds failures, and the other payment exceptions that litter the end-to-end payment lifecycle. They’re able to use AI to optimize customer onboarding, to identify fee and commission mismatches, and to bring AI to other parts of the business, from lending to wealth management.

As they begin to realize AI’s transformative potential, at scale, they’re able to move the needle on some urgent business objectives.

1. Deploying AI agents consistently and cost-effectively. When agents have a shared contextual layer, and bespoke integrations are minimized, banks can dramatically shorten time to production for AI initiatives. At the same time, shared governance frameworks help to reduce overall TCO.
2. Accelerating agent ROI. Through their real-time digital twin, banks can identify workflows and processes that should be optimized before AI is introduced, as well as those that represent low-hanging fruit for automation. The result? Process efficiency gains, and AI pilots that rapidly demonstrate significant returns.
3. Aligning agent outcomes with business needs. Business unit leaders understand their workflows better than anyone else. With central IT providing a single, enterprise-wide platform that’s built for low or no-code agent configuration, the right people can own AI transformation as it spreads throughout the bank.
4. Shifting to continuous compliance. T+1 settlement, instant cross-border payments (in modern banking, there’s less and less time to identify and address payment issues). But by monitoring agent actions and process execution in the moment, with real-time KPIs, banks can shift from reactive to continuous risk management and compliance. That means settlement exceptions decline, and capital velocity rises.
5. Showing progress on ESG targets. With real-time operational visibility and effective AI agents, banks can minimize AI’s environmental impact, while streamlining processes to reduce their overall carbon footprint. Such internal efficiencies augment improvements in financed emissions, bolstering Scope 3 disclosures and helping banks to demonstrate the depth of their commitment to sustainability goals.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:158990f8-b091-44d6-bdef-6ab19e742cb2/original/as/banking_datasheet_promo_card.png
https://www.celonis.com/insights/data-visualizations/banking-ai-datasheet

Building banks for today and tomorrow

We’re already on this journey with many of the world’s leading banks.

They share similar priorities, maintaining auditability and governance, preparing human teams to work alongside agents while building customer trust, ensuring AI makes banking faster and greener, without driving up costs.

But they also share an understanding. The banks that lead the way in the coming months won’t be the banks with the best AI models. They’ll be the banks with the most AI-ready architecture.

Find us at Sibos 2026

Celonis will be out in force at Sibos 2026. If you’re there too, come and connect with us.