Overcoming Data Fragmentation and the Limits of AI in Transaction Profitability
At a dedicated session of the FF News Virtual Arena, industry specialists gathered to discuss a critical bottleneck in banking operations: how data fragmentation and legacy architecture directly cause financial institutions to lose profitability within their transaction flows.
The discussion featured:
-
Ian Horne, Host at FF News
-
Mariia Komissarova, Data and AI Retail Business Lead at Raiffeisen Bank International
-
Breno Alves De Oliveira, Chief Product Officer at PAYABL
-
Kirill Lisitsyn, Co-Founder and CEO at Torus
The panel explored the hidden operational expenses of un-utilized data collections, the limits of non-deterministic artificial intelligence, and the strategies financial institutions must deploy to transform raw data into a core foundation for market survival.
The Root of Profit Leakage: Legacy and Structured Data Obstacles
For multi-generational financial institutions on the market, such as Raiffeisen Bank International, legacy infrastructure stands out as a primary internal barrier to optimization. Mariia Komissarova explained that the core challenge causing banks to lose profitability in transaction flows is fundamentally a data problem.
Because historical banking applications operate in distinct silos, collecting and structuring corporate transaction data in a transparent, organized format is exceptionally difficult. Without a structured framework, calculating the precise profitability of an individual financial transaction remains virtually impossible.
This breakdown stems from historical data governance and a lack of modern framework deployment. Advanced organizational paradigms, such as the “data mesh” concept, have emerged on the market but remain poorly distributed across large banking enterprises.
As the global financial sector navigates sweeping AI transformations across identity verification and transaction processing, resolving this data layer is no longer a luxury. Establishing a clean data foundation has escalated into an absolute requirement for long-term corporate survival.
The Hidden Cost Grid of Mass Data Ingestion
A common pitfall for legacy institutions is the assumption that capturing higher volumes of data naturally yields higher business value. Five to seven years ago, traditional industry playbooks focused on collecting as many varied data points as possible, including feeding data from social media networks into corporate servers.
The modern transaction ecosystem has outgrown this mindset. Financial institutions are finding that simply storing and maintaining humongous quantities of unstructured information incurs immense server and data-engineering expenses.
“This amount of data, big amount of data to collect them and to store them, it’s quite costly and if you not make use out of it, you also kind of start losing in this pricing game…”
When a firm triggers heavy operational storage costs without actively extracting commercial value from that data, it falls behind in the competitive pricing game. It cannot offer optimal rates to its merchants because its base infrastructure costs are artificially inflated.
As Kirill Lisitsyn highlighted, modern data strategy must focus on extracting real value from existing data assets first. Only when an explicit business use case is established should an institution invest capital to acquire additional data streams, thereby avoiding unnecessary operational hurdles and cost accumulation.
The Non-Deterministic Trap: Why LLMs Cannot Fix Bad Data
As institutions work to unify legacy systems that speak entirely different software languages and utilize non-standardized data formats, many turn to Artificial Intelligence and Large Language Models (LLMs) to automate code and data transformation. Breno Alves De Oliveira noted that fintechs excel at ingesting complex data and reorganizing it into easily digestible formats, a process heavily accelerated by AI tools.
However, Komissarova issued a strong technical caution regarding over-reliance on generative algorithms for core transactional infrastructure. LLMs are inherently non-deterministic, meaning their outputs are probability-based rather than absolute, exposing them to the systemic risk of algorithmic hallucinations.
In the transactional world, where errors directly impact financial ledgers, dropping below total precision is unacceptable. Feeding inaccurate or unstructured data into an LLM significantly increases the likelihood of generating incorrect calculations, potentially costing financial institutions millions of dollars.
The panel agreed that there is no technological silver bullet; companies cannot simply throw disorganized datasets at a generative model and expect flawless business logic. Building a reliable data layer requires a disciplined investment of time and capital, alongside skilled internal specialists who can structure the data pipeline correctly.
Balancing the Equation: The Deterministic Hybrid Core
To safely capture the speed of modern AI without sacrificing absolute financial accuracy, the panellists proposed a hybrid structural architecture. This model balances deterministic processing engines with flexible language interfaces to ease the workflow of the end user:
-
The Deterministic Foundation: The core data layer must remain strictly deterministic. Specialized intelligence platforms, such as Torus, intentionally construct their backend logic to focus on total mathematical accuracy rather than an “80% probability” model, ensuring that scheme fees and transaction records are perfectly reconciled.
-
The Conversational Interface: Once a baseline of verified data integrity is established, institutions can layer LLMs on top to interpret the data, simplifying user interactions and accelerating analytical tasks.
This structured foundation allows institutions to leverage concepts like data lakes to formulate and test commercial hypotheses. Historically, discovering a processing trend or evaluating a pricing variable required massive manual database queries.
With a unified hybrid core, product teams can rapidly test hypotheses to evaluate their probability of success. Ultimately, this framework enables banks to analyze their internal statistics, competitor landscapes, and macro market shifts simultaneously. This data-driven approach guides targeted adjustments across conversion flows, transaction routing, and product experiences, transforming necessary capital investments into predictable drivers of corporate profitability.
Key Highlights from the Virtual Arena Panel:
-
The Data Structure Bottleneck: Collecting data across legacy systems that utilize different formats makes accurate transaction profitability tracking highly complex.
-
The High Cost of Data Stagnation: Storing mass quantities of data without clear use cases inflates operational overhead, making banks less competitive in merchant pricing.
-
Value Over Volume: Modern data intelligence prioritizes extracting maximum utility from existing assets before purchasing external data streams.
-
The Danger of Non-Deterministic AI: Because generative AI models are probability-based, using them on unstructured core data risks financial calculation errors.
-
The Hybrid System Blueprint: Successful architectures combine a 100% accurate, deterministic data layer with conversational LLM tools on top for user interpretation.
-
Hypothesis-Driven Innovation: Re-engineering core data frameworks allows teams to quickly validate processing changes, de-risking capital investments.
The post Overcoming Data Fragmentation and the Limits of AI in Transaction Profitability appeared first on FF News | Fintech Finance.
Ayrıca Şunları da Beğenebilirsiniz
Best Mobile User Retention Tools in 2026
Elon Musk trial on track as Washington securities case
