Libra Internet Bank Automates Financial Analysis With Agentic AI and Scales It Bank-Wide

The Problem

Financial Analysis That Couldn't Scale

Financial analysis is core work at Libra Internet Bank, but the process behind it was slow and manual.

To assess a company, a credit analyst receives a pile of accounting documents: trial balances (balanțe de verificare) and balance sheets (bilanțuri), usually spanning several periods, sometimes as scanned PDFs, sometimes as native spreadsheets, each in whatever format the client's accounting software produces.

Before any analysis can begin, every number in that pile must be entered into the bank's standardized workbook. One row per account in the Romanian chart of accounts; one column per period; debit and credit sides kept straight; analytic sub-accounts rolled up to the correct parents. Done by hand, that is hours of transcription per client. A balance sheet of a few hundred accounts across five periods is a few thousand cells, keyed in one at a time, with the error rate that implies.

The work was repetitive, the calculations were exacting, and quality hinged on getting every figure right. As volumes grew, the only way to keep up was to add analysts. And for an EU-regulated bank, speed alone was never the point. Every output also had to be consistent, accurate, and auditable.

Libra needed a way to turn that pile of documents into a decision-ready workbook in minutes, without giving up accuracy, transparency, or control.

The Solution

A Multi-Agent Workflow for Financial Analysis

Libra Internet Bank partnered with Multimodal to deploy an agentic workflow on AgentFlow that mirrors and augments the analyst's process, from raw documents to a finished credit report.

Rather than a single model, the solution orchestrates specialized agents, each owning one stage of the work:

• Document AI reads incoming accounting documents, runs OCR on scans, and extracts every account across every period, whether the source is a native spreadsheet or a photographed page.
• Decision AI maps those accounts into the bank's standardized workbook: the Romanian chart of accounts, debit and credit sides kept straight, and analytic sub-accounts rolled into their correct parents.
• Report AI builds the 13-section credit report using the same verified numbers, so the analysis and the workbook always agree.

The analyst now reviews the workbook instead of typing it. The workflow was built for control and for incremental adoption: it pauses for human review, so the team can confirm the figures, correct anything that needs it, and approve the result before it moves on. Every execution is transparent and traceable, and approved results flow into Libra's own systems through a live API integration, giving the bank the audit trail a regulated environment requires.

"What differentiates Multimodal is not just the technology, but their ability to understand banking operations and deliver solutions that integrate quickly and effectively." Andreea Manuela Tirlea, Libra Internet Bank

The Results

Adoption That Speaks Loudest

The clearest measure of the workflow's value is how much Libra's analysts actually use it.

Adoption ran far past the plan. When the engagement was scoped, the bank anticipated a gradual level of adoption. The strong results achieved after go-live led to a decision to expand adoption across additional business functions far more quickly than originally anticipated. Within a short period after launch, analysts adopted the Financial Analysis workflow at a level that significantly exceeded initial expectations. The APIA workflow showed a similarly strong adoption pattern soon after go-live. This was not a pilot being used out of obligation, but a tool analysts chose because it substantially reduced the time previously spent on manual data entry.

The work went from hours to minutes. What used to be hours of transcription per client, thousands of cells keyed in one at a time, now arrives as a completed workbook the analyst reviews rather than rebuilds.

Accuracy held up, measured conservatively. Across a 12-client test set of roughly 4,000 account cells over 43 periods, the workflow reproduced 93.6% of the analyst's corrected workbook for native digital documents and 87.1% across all clients, including scanned inputs. That figure is a floor, not an average: the test set was deliberately weighted toward the hardest cases, so typical production accuracy is very likely higher, and the remaining gaps trace back to a short list of specific, already scoped fixes.

The results generalized. On the strength of the work, Libra now runs two workflows in production: Financial Analysis and APIA, both feeding approved results into the bank's systems, and is developing a third workflow for Individual Lending.

Key results

Adoption

• Financial Analysis: adoption significantly exceeded the bank's initial forecast shortly after launch
• APIA: The workflow showed a similarly strong adoption pattern soon after go-live

Speed

• A client's analytical workbook is built in minutes instead of hours of manual transcription
• Analysts review a completed workbook rather than keying thousands of cells by hand

Accuracy

• 93.6% cell-level extraction accuracy on native-digital documents, measured conservatively against human-corrected workbooks
• 87.1% across all 12 clients including scanned inputs, which are handled via OCR
• Measured over ~4,000 account cells across 43 periods and 12 clients

"The pilot delivered measurable efficiency gains and, more importantly, validated the scalability of this approach. Building on these results, we have moved forward with broader implementation, extending these capabilities into our core operations. We view Agentic AI as a foundational capability for the future of banking, and Multimodal is a strong partner in helping us accelerate this transition." Andreea Manuela Tirlea, Libra Internet Bank

About the customer

Libra Internet Bank is a Romanian commercial bank, digital-first, with a niche focus on liberal professions, real estate, and agribusiness. As an EU-regulated institution, it prioritizes accuracy, auditability, and compliance in evaluating and adopting new technology.