The data on AI adoption is contradictory. While 91% of organizations plan to increase AI investment this year, the actual return on that capital is stalling.
According to Deloitte’s 2025 survey, only 6% of organizations see a return on investment in under a year. Even worse, new data from the Voice of the Enterprise reveals that 46% of AI projects are scrapped between proof-of-concept and broad adoption.
What accounts for this gap between AI and ROI? Daniel Pack, Senior Director of Value Engineering for Financial Services at Celonis, calls it “Task Myopia.”
“We’ve validated that GenAI is a powerful engine for knowledge work—mastering search, summarization, and synthesis. But the industry is hitting a wall as it attempts to pivot from a technology of language to a technology of action.”
“We see banks successfully automating isolated cognitive tasks and declaring a win. But systemic failure occurs when they try to scale that to complex, end-to-end execution. If a faster email summary doesn’t actually speed up the loan approval or prevent a client churn, you haven’t generated ROI. You’ve just generated a more expensive technology bill.”
Escaping “Pilot Purgatory”
Financial service firms are currently stuck in pilot purgatory. They have thousands of experiments running, but can’t trust them to run autonomously.
The issue isn’t the model’s intelligence; it’s the model’s context.
AI needs a structured environment to operate within. In data-science terms, this requires a context graph—a framework that organizes disparate data points into meaningful relationships.
But for complex banking operations, a static graph of documents isn't enough. The graph needs to be oriented specifically around processes—one that maps complex system lineage, cross-functional hand-offs, and operational controls.
The Bridge: Reasoning vs. Reality
This is where the architecture needs to evolve.
“Think of the LLM as the reasoning engine, and Process Intelligence as the sensing engine,” Pack says. “An AI agent needs a sensing and contextualizing layer before it can safely act. A generic model doesn’t understand your specific business rules, the interaction points between your siloed legacy systems, or the ‘reality on the ground’—like why a specific loan exception was granted last Tuesday,” Pack explains.“You can’t get reliable, compliant results from a model that lacks the lineage of your data.”
"Right now, many banks are rushing into the 'act' phase without that map. They’re asking AI to execute on workflows without context about system handovers or compliance boundaries. That isn’t automation; that’s risk.”
How to Fix It: Analyze, Design, Operate
To close the gap between a science experiment and a production-grade solution, Pack suggests a three-step framework grounded in reality, not hypothesis:
1. Analyze: “Stop building AI based on how you think your process works (the Happy Path),” Pack warns. “Mine your system logs to see the ‘invisible’ manual work and data breaks that actually happen. If you automate a broken process with AI, you don’t get efficiency—you just get bad results, faster.”
2. Design: “We can’t simply layer Agentic AI on top of legacy workflows and expect transformation,” Pack says. Instead of simply digitizing obvious manual steps, teams should design workflows that support AI with human-in-the loop execution. “By using Process Intelligence to define the business rules and KPI definitions, you create a structure where the agentic solution goes beyond mimicking a human’s clicks, to executing a fundamentally more efficient process.
3. Operate: “We’re at a critical inflection point, where AI is evolving from extrinsic tools that require constant human direction into autonomous allies capable of executing complex, goal-driven workflows,” Pack explains. “That’s why banks need a platform that can monitor and manage the interactions across multiple systems and actors, ensuring that tasks are accomplished not just quickly, but with structured intelligence and correct context.” In this new operating model, Process Intelligence serves as the conductor. It orchestrates the complex hand-offs between a legacy system, an AI agent, and a human in the loop. Moving beyond isolated tasks to intrinsic execution, these agents transform from mere digital assistants into motivated, integrated members of the professional team.
The ROI of "Defined Value"
Funding AI initiatives is no longer a blank check. Executives need to draw a straight line between AI spend and bottom-line impact.
Pack notes that by using context-aware automation, one client achieved a 47% reduction in FX trade processing costs, while another realized 80% savings in manual rework for outgoing payments.
But the biggest ROI might be risk avoidance.
“Proactive governance costs less than reactive remediation,” Pack concludes. “Regulators issue both fines and growth caps for banks that lack robust controls. If you can’t prove transparency in your automated processes, you aren’t just risking a fine—you’re risking your ability to grow. The 'Operational Nervous System' provided by Celonis ensures your AI is not just smart, but safe, scalable, and solvency-enhancing.”
