ID and Passport Data Extraction AI Model
Eastern European retail bank
In short
How can a bank extract ID and passport data with AI and still check accuracy?
Typing customer identity data by hand slows onboarding at a bank and puts errors into fields that compliance depends on. We built an AI model that reads ID cards and passports and returns structured fields for staff to check. Accuracy is measured on documents the model has never seen.
At a glance
- Two document types, ID cards and passports
- Separate training and unseen test sets
- A person confirms low-confidence fields
The client
The client is a retail bank in Eastern Europe. Customer onboarding there involves checking identity documents and entering their data into banking systems. The model covers the two document types staff handle most, ID cards and passports.
What it was costing them
- Time: staff read each document and type the fields into the onboarding system.
- Errors: a mistyped name or number can fail a check later, and then the customer is contacted again.
- Key-person risk: experienced staff know which documents are hard to read, and that knowledge is not captured.
- Compliance exposure: identity records must be accurate and traceable.
- Growth ceiling: more customers means more typing and longer queues.
What we built
We built an AI model that reads ID cards and passports and returns structured fields for staff to confirm, with separate data sets to train and to test it.
- Honest accuracy: ID cards and passports each get a training set and a separate unseen test set, so accuracy is measured on documents the model has not met.
- The right layout for each document: ID cards and passports are treated as two document types, each with its own fields.
- Faster onboarding: the model returns name, birth date, document number, expiry date and issuing country as structured values.
- A person stays in the loop: the app shows the image next to the extracted values and asks staff to confirm fields with low confidence.
- Fewer typing errors: the machine-readable zone on passports is checked with its built-in check digits.
- Local scripts handled: the model covers Ukrainian and Latin characters on the same document.
- Protected data: images sit in restricted storage with set retention and row and field security.
Results
- Result: staff check extracted fields instead of typing them, which roughly halves handling time per document.
- Result: testing on the held-out sets gives a field-level accuracy figure before any live use.
- Result: check digits on the passport reading zone catch many reading and typing mistakes before they reach the system.
The risk most teams miss
The month-three risk is that the model scores well on the test folder and then meets real photos: glare, folded corners, new document versions. Separate test sets keep the accuracy figure honest. The model uses confidence thresholds, human review for low confidence, an error log, and retraining when a new document version appears. Personal data rules also set where images are kept and for how long.
What we would add today
Improvements we would add on top of what is live.
- Start from the Azure Document Intelligence identity model as a baseline, then train a custom model on the bank's own documents.
- Use AI Builder document processing for a quick pilot inside the Power Apps review screen.
- Protect stored images with Dataverse field-level security and Azure Key Vault managed keys.
- Offer customers a Power Pages upload page, so they can submit a photo from home and staff only review exceptions.
- Track accuracy by field and document version in a Power BI report on Fabric.
Technology
Stack: Azure AI Document Intelligence, Azure Machine Learning, Azure Blob Storage, Power Apps and Dataverse for review.
Tell us which identity documents your staff type in by hand, and we will tell you how accurately a model could read them.
Products and capabilities
- Azure AI Document Intelligence
- Power Apps
- Dataverse
- document AI
- identity document extraction
- customer onboarding
- model accuracy testing
The service behind this work
Copilot and agent readinessMore on this industry
Banking and finance
Related use cases
Operations and field workContractor Management App Rebuilt in Microsoft Power Apps
When the client acquired a US risk services business, one business-critical application could not simply move across. It ran on a third-party form tool that the client did not want to carry into its main Microsoft tenant, and the team that served contractors every day depended on it. We rebuilt it in Power Apps, with every function intact, inside the migration deadline.
- 2,000+
- items in the largest list
Power Automate Desktop Response to a Municipal RPA Market Consultation
A municipality that wants to buy robotic process automation must choose between tools and decide how to tender them, with little time. The municipality ran a request for information to learn what the market offers. We answered it for Power Automate Desktop.
Document Management with Approvals for Emailed and Posted PDFs
When closing documents arrive as PDFs by email and by post, each one has to be found, checked, approved, completed, sent and chased by hand. A missed reply or an unsigned page stalls a closing. We built a SharePoint and Power Automate process that files each document, routes it to a paralegal, produces the outgoing PDF and keeps reminding until the reply arrives.
- 15 min
- saved per document
- 20 hrs
- saved per 30 properties
Have a Microsoft project in mind?
Tell us what you need built, fixed or supported. A senior consultant replies within one business day.