TechLegion × SoftBees
Reviewing a bank statement for financial monitoring at ASVIO Bank dropped from 11.5 hours of manual work to about one hour, with every finding evidenced and traceable to the transactions behind it.
- Partner
- SoftBees
- Customer
- ASVIO Bank
- Industry
- Banking / Financial Services
- Use case
- AML / financial monitoring
- Claude product
- Claude 3.7 Sonnet via Amazon Bedrock (API)
- Deployment
- Cloud or fully on-premise (Docker container / Java JAR)
The use case
AI pre-screening of bank statements. The agent reads statements in any format — text PDF, spreadsheet, scan or phone photo — reconstructs the transaction table, validates it, and produces an explainable risk assessment with the evidence behind every finding. The compliance officer keeps the final decision.
The problem
Bank statements submitted for financial monitoring run 40–90 pages and arrive in more than ten different formats, frequently as scans or phone photographs. There is no standard layout across banks, so every document was effectively read from scratch by hand — around 11.5 hours per large statement, with the error risk that monotonous work carries. This sat directly on the bank's compliance exposure.
What SoftBees brought
Model selection was an engineering decision rather than a default: SoftBees evaluated a range of models against real banking documents. Local open-source models, without fine-tuning, failed on domain-specific Ukrainian identifiers — MFO routing codes and EDRPOU company numbers read as ordinary digit strings to a general model, and the accuracy of everything downstream depends on getting them right. Claude returned correct results out of the box, with no fine-tuning, which is what made the client's timeline achievable.
The team then built a three-path pipeline so the model is used only where it adds value:
- Text-layer PDFs go through a deterministic Java parser reading the document's own text and table geometry — no model call, no latency.
- Post-processing with Claude is layered on top: it sees each page, its image, and the rows the parser extracted, and fixes what a table parser structurally cannot — rejoining a value wrapped inside its own cell, recognising a bare counterparty code, completing a truncated payment purpose.
- Scans and photos are read by Claude directly, one call per page in parallel, with pages cut into bands to buy resolution where pixels are the only source of digits.
Every path ends the same way: identifiers normalised, IBANs verified against their own mod-97 checksum, totals cross-checked against whatever total the statement prints about itself. Rule-based AML analysis then runs over the structured result across nine categories, and each finding cites the specific transaction rows behind it and states what the officer should verify next. SoftBees deliberately did not build a black box — the agent accelerates the first level of review, and responsibility for the decision stays with the human.
Results
Before → after
- Processing one large statement: 11.5 hours → ~1 hour (−90%)
- Analyst throughput: from 1–2 cases per day to dozens
Processing speed by format
- Spreadsheets and text-layer PDFs: 1–2 seconds
- Scans and photographs: 60–90 seconds
Adoption and scale
- 4,000+ documents processed across 2025–2026
- Live in production for financial monitoring — not a pilot
- Additional banking institutions in deployment discussions
Independently confirmed by the customer
- Cost of processing one document case: reduced seventeen-fold
- Human error in the process: effectively eliminated
Both figures were stated publicly by ASVIO Bank's Chairman of the Board at the DIGITAL HORIZONS: SMART BANKING conference, September 2025.
“There is a great deal of discussion around breakthrough AI technologies in banking at the moment, but still few practical cases. The next step in ASVIO Bank's digitalisation was deploying an AI agent to automate financial monitoring and operational processes. What it gave us — ASVIO's first results: we have effectively eliminated human error; we cut the processing time for a single document by 90%, from 11.5 hours to one hour; and we reduced the cost of processing a case of documents seventeen-fold. Our next step is scaling the AI agent, and our clients will feel it in service quality very soon.”
The SoftBees team
- Tetiana Dashevska — CEO & Co-Founder
- Serhii Ratkov — CTO & Co-Founder, led the technical side of the implementation