Show Me the Money: Why AI in Banking is No Longer a Conversation About Technology?
- Published
- 6 min reading
Key takeaways
- Banks are shifting from “where can we use AI” to “how much does it save.”
- A real credit-process case: 1 hour saved per case × 12,000 cases = 12,000 recovered hours.
- Directed toward productive work, that capacity generated roughly €1.3M in annual value.
- A well-executed AI portfolio can lower the cost-to-income ratio by roughly 125 bps over 3 years.
- The bottleneck isn’t the AI model — it’s whether the bank redesigns the process around it.
Here’s how banks are calculating AI ROI in 2026
Moving past the romanticizing of the AI. It’s time for figures
A bank can save thousands of working hours thanks to artificial intelligence — without earning a single penny from it. This might be the most vital lesson from the first wave of AI deployments across the financial sector.
Just a couple of years ago, executive boards asked where they could use AI. Today, leadership teams want to know how much money it can save them.
This shift extends well beyond banking. According to Google Cloud and National Research Group, search queries for "AI ROI" surpassed searches for "how to use AI" for the first time in the spring of 2026. At the same time, 86% of financial sector executives report that they primarily expect AI to provide measurable, cost-effective growth.
This shift matters greatly for banks as well as technology providers.
Why most bank AI use cases don’t show up in the P&L
A press release does not constitute a use case
The initial wave of generative AI naturally focused on areas that were easy to launch and demonstrate, including chatbots, search tools, document summarization, and employee assistance. The greatest economic value often lies deeper within core operations: lending, risk analysis, decision-making, operational automation, and advisor interactions. That deeper operational layer is precisely where standards are rising.
Under the AI Act, several applications linked to individual credit scoring fall under the high-risk category. This classification introduces extra demands regarding data quality, documentation, system controls, and human oversight. The dynamic illustrates a clear paradox for the sector: as AI moves closer to decisions that directly shape bank profits, the burden of proof increases across technological, operational, business, and regulatory fronts. Consequently, the discussion is shifting. Leaders ask fewer questions about future possibilities and focus heavily on current operations, scale, and clear P&L impacts.
Why calculating AI ROI in banking is so hard
The honeymoon has ended — show me the money
Calculating AI ROI in banking is difficult because customer, product, and risk data are fragmented, and returns materialize on different timelines depending on the use case. Here lies the harder part of the problem: controlling departments can easily construct financial models, yet the fundamental structure of digital transformation complicates the math, and software integration almost always coincides with organizational, procedural, and workflow redesigns.
Return horizons also vary significantly. Operational work automation yields visible savings within a few months. By contrast, the true value of an upgraded credit model emerges much later through portfolio performance, credit losses, write-offs, and customer behavior across the entire exposure lifecycle. Attribution adds another layer of complexity. When a process yields better results, isolating the exact financial contribution of AI versus workflow redesign or work reorganization remains difficult.
AI cost savings in action: 12,000 hours, €1.3M in corporate banking
Recovered 12,000 hours: so what comes next?
A real-world process illustrates this ROI challenge effectively. Comarch is implementing the first phase of an AI tool supporting the credit process at a European bank. This solution reduces administrative paperwork for client advisors.
The measured impact shows roughly one hour saved per case file. To understand the broader impact, consider a bank processing 12,000 such applications each year. Multiplying one hour across this many cases yields 12,000 recovered working hours annually.
At this stage, many ROI calculations make a fundamental misstep. Twelve thousand hours represents capacity rather than immediate cost savings or direct revenue. The bank's management decisions alone determine the financial value of that capacity.
If advisors spend those recovered hours driving sales, the existing branch network can process roughly 4,000 additional applications annually. Under baseline process parameters, this creates potential added revenue of EUR 930,000 per year. If AI agents simultaneously handle back-office document verification while staff members shift toward higher-volume tasks such as collateral valuation, the freed capacity roughly matches the capacity of eight full-time positions, generating another EUR 400,000 in annual value.
Combined, these steps create approximately EUR 1.3 million in annual potential impact within a single process during initial implementation. The central element of this calculation relies on the word "if" rather than the headline figure. If none of those 12,000 hours are directed toward productive tasks, the project ends up as an impressive presentation on productivity growth with zero actual impact on the income statement.
Why AI alone doesn’t fix broken banking processes
AI cannot fix what the bank refuses to change
Technology constitutes only one part of the equation. Banks built their first generation of automation largely on robotic process automation, employing bots that mirrored user actions within existing applications. That approach succeeded across stable, highly structured workflows. Complex workloads reveal clear limitations, especially when handling varying document formats, edge cases, interface updates, or data interpretation requirements. User-interface automation becomes fragile under these conditions, with maintenance overhead eating into projected returns.
Modern solutions operate at a deeper systemic level. API integration and microservices enable software agents to work directly on underlying data and workflows rather than mimicking human screen interactions. Well-designed processes can achieve straight-through processing rates at 85–90%.
Even optimized architecture fails to solve core organizational bottlenecks on its own. If AI cuts analysis times from one hour to five minutes while the next workflow phase still waits until the following morning, the financial return remains unrealized. If a model suggests a credit decision, yet an employee must navigate the exact same approval pipeline, the overall impact remains limited. Similarly, giving an advisor an extra hour each day adds little value if targets, client portfolios, and sales structures remain unchanged.
Access to AI models no longer offers a distinct competitive advantage. Almost any financial institution can license identical or comparable models. Competitive advantage stems from rapidly integrating these models into existing data systems, architectures, decision pipelines, and daily operations. This integration capability will likely serve as a key technological differentiator for banks in the coming years.
Measuring AI success with the cost-to-income ratio
The real test: 100 basis points
Ultimately, board members evaluate AI on financial results rather than prompt counts, active agents, process automations, or model deployments. This focus is acute within Polish banking, where costs experience simultaneous pressure from wages, regulatory compliance, cybersecurity, IT infrastructure, and digital customer acquisition.
The cost-to-income ratio serves as one of the best metrics for measuring AI maturity. Moving this ratio long-term presents significant difficulty because large portions of banking expenditure remain fixed or increase regardless of management action. Research by Comarch indicates that a carefully selected and consistently executed portfolio of retail and corporate banking AI applications can lower the cost-to-income ratio by roughly 125 basis points over a three-year horizon. This figure represents a strategic forecast rather than a single implementation output, yet it sets the expected level of ambition. Moving forward, banks will shift away from promoting 25 completed use cases or 200,000 saved hours, focusing instead on quantifiable shifts in business economics.
What changes for banks?
From AI strategy to AI economics
Moving past initial fascination benefits the financial industry. The old question of where to apply AI offered a comfortable baseline, in which nearly every answer seemed correct and innovative. Asking for financial return creates far higher accountability. It demands baseline benchmarks, pre-implementation metrics, dedicated ownership, explicit plans for freed capacity, and the willingness to shut down underperforming initiatives.
This accountability defines the incoming deployment wave. Evaluations move beyond basic software functionality to demonstrate measurable shifts in process economics and overall organizational performance. Seeing the outcomes of Comarch's internal AI transformation provides strong grounds for confidence in this trajectory.
Business Banking Product Manager, Comarch
Originally published in “Miesięcznik BANK”
