AI is only as good as its data: How live data context transforms treasury workflows
By Jérémie Feuillette
Commercial banks have spent the last two years deploying AI assistants across treasury and risk functions. The tools are capable. The models are fast. And in most institutions, they are still running blind.
The reason is not a failure of AI, it is a failure of data access. A language model answering questions about intraday liquidity positions, Net Debit Cap (NDC) usage, or counterparty concentration cannot produce accurate answers if it cannot see the underlying data. Most deployments route around this problem by feeding the model spreadsheet exports, end-of-day reports, or manually prepared summaries. The result is a highly capable system answering questions about yesterday, with no mechanism to flag when yesterday's answer no longer holds.
For treasury teams managing real-time settlement flows across multiple rails, that gap is not a minor inconvenience. It is the difference between a tool that accelerates decision-making and one that gives confident answers to the wrong question.
The problem is structural, not technical
Intraday liquidity data is inherently fragmented. A commercial bank's positions at any given moment are distributed across SWIFT payment queues, CLS settlement windows, CCP margin calls, T2 accounts, and increasingly tokenised asset rails operating on different settlement cycles. Batch reporting — the standard model for most treasury management systems — collapses that complexity into a single end-of-day snapshot.
When an AI assistant pulls from that snapshot, it is not accessing intraday intelligence. It is accessing a compressed historical report with all timing, sequencing, and flow-level details stripped out. Asking it to support a live treasury decision on that basis is like asking a navigator to plot a course using last night's weather data.
The structural fix is not to improve the model. It is to change what the model can see.
What ILO MCP does
FNA's Intelligent Liquidity Optimization (ILO) Model Context Protocol (MCP) connects your AI assistant — whether that is Claude, ChatGPT, Gemini, or a proprietary model your institution has approved — directly to live intraday liquidity data via ILO's infrastructure.
The connection works through the Model Context Protocol, a standardized interface that allows AI assistants to query external data sources in real time without requiring the model to store or reproduce that data. When a treasury analyst asks a question, the assistant queries ILO's live data layer, retrieves the relevant positions, and generates a response using current information and the appropriate context, avoiding hallucinations and burning through tokens unnecessarily. . The data does not leave FNA's infrastructure. The audit trail is preserved. The answer is explainable at the transaction level.
In practice, this changes what treasury teams can do in a single working session. Counterparty network graphs, concentration analyses, NDC driver breakdowns, and liquidity impact assessments for pending trades are all available on demand, generated from live positions, and ready to download or share with a CFO or ALCO committee.
What this looks like for a commercial bank treasury team
Three scenarios illustrate the shift.
Before an ALCO meeting, a treasury analyst needs a breakdown of what is driving net debit cap consumption and whether any counterparty exposure has moved materially since the last report. Under a standard workflow, this takes hours of manual data preparation. With ILO MCP, the analyst queries the AI assistant directly: the system pulls live flow data, identifies the structural drivers, and generates a summary with supporting charts — in seconds, with the transaction-level evidence attached.
During a live trading session, the desk wants to assess the intraday liquidity impact of a large pending payment before it is released. Previously, that assessment would require a manual call to treasury operations and a wait for a spreadsheet to be prepared. With ILO MCP, the impact is modelled against current positions the moment the question is asked.
For regulatory reporting, the same data layer that supports the AI queries maps directly to BCBS 248, PRA, and ECB reporting standards. Rather than maintaining separate data pipelines for operational analytics and compliance, the two run from a single source, reducing the risk of discrepancy and cutting the manual effort involved in aligning them.
The infrastructure behind the answer
The accuracy of any AI-generated analysis depends entirely on the reliability of the data layer beneath it. FNA's ILO infrastructure has processed over $358 trillion in payments globally — sitting as an overlay on existing payment infrastructure, requiring no core system replacement or multi-year integration project. Most commercial banks can connect a first currency within two to three weeks, working from just three inputs. ILO's patented algorithms handle the flow orchestration, and the MCP layer exposes the resulting intelligence to whatever AI the institution has deployed.
The model has no access to data outside its query scope. Every response is grounded in the institution's own data, with explainability at the transaction level — which matters both for internal governance and for regulatory audits.
Learn more at fna.fi/products/ilo-commercial-banks or get in touch with the team to book a demo.
