Hear it from Banco Macro
The bank's Operations Lead on moving from a team that executed repetitive tasks to one that manages the process.
Client Challenge
A review process that had not changed in 30 years
Banco Macro is Argentina's largest private bank by branch network and equity, and physical checks are still central to how its customers pay. The bank clears roughly 650,000 checks a month, each with its own deposit slip; this adds up to a total of 1.3 million documents a month and 15.6 million a year.
Every document was opened on screen by an analyst who compared the image against COBIS and SPC records by eye: amount, date, drawer CUIT, endorsement, and signature. Values were then re-keyed into the balancing module, and high-value checks required two independent reviewers, which meant reviewer fatigue on decisions worth billions of pesos.
In addition, Banco Macro had acquired the Argentinian operations of Itaú Unibanco, and the integration made scaling that team unviable under a fixed deadline. No bank in Argentina had automated the process, so there was no reference implementation to follow, and any solution had to leave the analysts' workflow undisturbed.
The Solution
A governed document-AI pipeline on Databricks
SunnyData designed and built an end-to-end check-clearing pipeline on Databricks. Images land through an API, key fields are parsed, the signature is verified by a purpose-built computer-vision model, and a rule engine returns an accept or reject decision to SPC, with Unity Catalog governing the lineage of every sensitive image from raw file to the decision written into the core banking system.
Check and deposit-slip images arrive through an API at the bank's own operating frequency, governed in Unity Catalog from the first byte.
ai_parse_document reads amount, date, drawer CUIT, lot number, account and endorsement from handwritten, stamped and low-quality scans.
A purpose-built computer-vision model confirms the signature is present and matches it against the bank's stored signature base at a tuned acceptance threshold.
The rule engine runs the market's clearing logic and returns accept or reject to SPC through Model Serving. Only genuine exceptions reach a person.
The platform runs entirely on Databricks and is deployed as code, so the same architecture ports to the next institution without a rebuild. Two clearing flows × four task types × two check types, each with distinct validation rules, live in one governed engine.
ai_parse_document
The parsing cost
Four document-AI stacks were benchmarked on real check images at the bank's real volume before the architecture was fixed. At 15.6 million documents a year, the price per thousand pages is the difference between a platform that runs for $60K and one that spends $624K on parsing alone.
AI Parse Document reached 91% parsing accuracy at the lowest price point of the four. The design also stayed cloud-agnostic, so the bank kept portability across its existing AWS, GCP, and Azure footprint rather than locking a core banking process into one provider.
The Pipeline at Work
An analyst opened the image, read four fields, compared each against two core systems and re-keyed the result. The pipeline does the same work as a governed sequence, and records every step.
Five stages a person used to run by hand
Key Benefits Achieved
Documented business results, before and after
Baseline figures were measured with the customer. Post-go-live figures are the business case approved by Banco Macro; accuracy is the MLflow evaluation on real check images in the bank's QA environment. Three Model Serving endpoints are running against real images today, reconciled against analyst decisions, with production cutover scheduled for September 2026.
A packaged accelerator for regulated back offices
The check-clearing reference architecture, parse-and-rules pipeline, signature-verification model, and SunnyData's MLOps Quickstart (guardrails, RBAC, monitoring, lineage, AI risk governance) ship together. This build took 16 weeks; the accelerator brings the next stage to 6 to 8 weeks. The same pattern applies to account opening, loan files, insurance claims, and any regulated back office process that still reviews documents by hand.