Use Case
National Bank of Greece — document intelligence
Documented application · Deployment in Greece
What the evidence establishes
Microsoft’s June 2024 customer story describes NBG using Azure AI Document Intelligence to classify documents and extract information for digital banking. The account reports integration with six customer-facing applications and support for Greek-language documents.
Results and their limits
The customer story reports processing below 0.5 seconds per page and approximately 90% accuracy. These are reported project figures, not guarantees for other document collections.
How to approach a comparable workflow
For a comparable pilot, start with one document family. Separate classification, field extraction and business validation. Route uncertain fields to staff and keep the original document beside the extracted result.
The following is an AIStrategy.gr implementation proposal, not a description of undisclosed project architecture. Begin with a bounded task, a named operational owner and an explicit acceptance decision. Separate data preparation, model output and the final business action so each can be evaluated.
Data and operating prerequisites
Document the origin, permissions, update frequency and known gaps in the input data. Build a representative evaluation set before changing the workflow. Include difficult cases and examples that should be escalated rather than answered automatically.
Risks and human oversight
A page-level accuracy figure can hide errors in critical fields. Personal data, poor scans, changing templates and incorrect document associations require explicit controls.
What to measure
Measure field-level precision, manual correction rate, time to verified completion and cost per accepted document. Evaluate unseen templates separately.
A practical first pilot
- Define the task, affected users and what a correct outcome means.
- Record the current baseline and build the evaluation sample.
- Run in advisory mode with staff reviewing the outputs.
- Review quality, effort, cost and incidents together before expanding.
Decision gate
Proceed only when the responsible team can explain which errors remain, who handles them and whether measured benefit survives the cost of operation and review. Revise or stop the pilot when that case is not established.
Source and review
Source reviewed 6 November 2026. Results are attributed to the publishing organisation or supplier. AIStrategy.gr has not independently audited them.
Use the practical playbooks ↗ · Explore the discussion questions ↗ · Submit evidence or a correction ↗