What problem does it solve?
Many organizations still run on documents that were designed for people: application forms, claims, certificates, supporting evidence, delivery notes, tax documents and letters, arriving as scans, phone photos, PDFs and email attachments. Staff open each one, work out what it is, retype the fields into a system and check them against other records. It is slow, error prone and hard to scale when volumes spike, and the backlog delays decisions that matter to citizens and customers.
Earlier OCR and template tools worked well for fixed layouts and struggled with anything else. Current document AI combines layout aware extraction, handwriting recognition and language models, so it can handle varied layouts, stamps, handwritten notes, tables across pages and several languages, as the Volvo Group deployment on this page shows. The design question is no longer whether AI can read the document but where it may act alone: straight through processing for confident, validated extractions, and human review for the rest, with every value traceable to the place on the page it came from.
How does it work?
- Ingest from every channel. Uploads, scans, email attachments and portal submissions land in one intake, where images are cleaned, rotated and split.
- Classify and separate. Each page or bundle is classified by document type (form, identity document, certificate, statement, invoice) and split into individual documents.
- Extract with confidence. Fields, tables, checkboxes and signatures are extracted, with the location on the page and a confidence score for each value; text in other languages can be translated.
- Validate. Values are checked against business rules (formats, totals, dates) and against source systems (the customer, case or supplier record).
- Route by confidence. Confident, valid documents flow straight into the downstream system; the rest go to a reviewer who sees the page and the extracted value side by side.
- Learn from corrections. Reviewer corrections are logged to improve extraction and to show which document types or sources cause errors.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- API and system to system, Email, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | at least 90% | 1 | 1 vendor |
| Hours saved | Not pooled | 10,000 hours | 1 | 1 vendor |
| Productivity gain | Too few to pool | 10x | 1 | 1 vendor |
Value drivers: Lower cost to serve, Speed and cycle time, Employee productivity, Risk and loss reduction.
Indicative value
An organization that processes 500,000 forms and supporting documents a year
USD 500,000 to USD 2.6 million
Manual document handling cost avoided per year
How this is calculated
Formula: documents * minutesPerDocument * timeSaved * costPerMinute. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Documents processed per year documents, documents per year | 500,000 | 500,000 | The reference organization. |
| Manual handling time per document today minutesPerDocument, minutes per document | 4 | 8 | Editorial assumption for classifying, keying and checking a document. Google Cloud reports that Pupuk Indonesia's data extraction took 5 to 10 minutes before AI. |
| Share of handling time removed, including review of uncertain cases timeSaved, fraction of handling time | 0.5 | 0.8 | Editorial assumption; review of low confidence documents stays with people. |
| Fully loaded processing staff cost costPerMinute, USD per minute | 0.5 | 0.8 | Editorial assumption, replace with your own. |
What it leaves out: Counts only handling time. It leaves out faster decisions for customers and citizens, fewer keying errors, the cost of the platform and integration, and the reviewer capacity needed at peaks.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
U.S. Citizenship and Immigration Services
United States · Government and public sector · 2024
Before this system, every page of an I-539 application (a request to extend or change nonimmigrant status) was scanned and stored as one document, which slowed adjudication and did not meet National Archives records standards. USCIS now uses an intelligent document processing tool to identify, classify and split each application into its component documents, such as the form itself, other USCIS forms, passports, driving licences, marriage certificates and bank statements. Pages the tool cannot identify go to a person. It is listed as deployed since November 2024; no outcome figures are published.
No outcome disclosed.
U.S. Immigration and Customs Enforcement
United States · Government and public sector · 2019
Business units at ICE, part of the Department of Homeland Security, use an intelligent document processing platform (UiPath Suite and Azure AI Document Intelligence) with OCR and machine learning models to verify, extract and classify information from forms, automating repeatable work such as invoice processing and form entry validation. The agency lists it in operation since 2019 and says it saves staff significant time while improving data quality. No figures are published.
No outcome disclosed.
Ancine
Brazil · Government and public sector · 2026
Ancine, which Google Cloud describes as the Brazilian cinema industry regulator, uses Google Cloud AI to extract and structure data from digitized tax documents to automate the accountability analysis of subsidized projects. Google Cloud reports extraction accuracy above 90% and a tenfold increase in analysts' daily processing capacity.
- Accuracy: at least 90%, data extraction accuracy
"This AI implementation achieved over 90% data extraction accuracy, boosting analysts' daily processing capacity by 10x."
Claimed by: vendor - Productivity gain: 10x, analysts' daily processing capacity
"This AI implementation achieved over 90% data extraction accuracy, boosting analysts' daily processing capacity by 10x."
Claimed by: vendor
Pupuk Indonesia
Indonesia · Manufacturing · 2025
Pupuk Indonesia, which Google Cloud describes as Asia's largest fertilizer producer, automated its document processing workflows with Vision AI and Gemini, working with Devoteam. Google Cloud reports that data extraction time fell from 5 to 10 minutes to 40 to 70 seconds, and that one employee now validates the results.
No outcome disclosed.
Volvo Group
Sweden · Automotive · 2023
Volvo Group's automation team built a document processing solution on Azure AI Document Intelligence for its service and financing businesses. It reads emails, digital and scanned PDFs and written bills, including stamps, photographs, handwritten notes over printed text and tables across pages, translates content between languages and outputs XML or CSV for the receiving division, with processing time and success rate tracked in a dashboard. Microsoft reports that the solution has saved 10,000 manual hours since launch, about 850 hours a month.
- Hours saved: 10,000 hours, since launch, about 850 hours per month
"Since launch, the company has saved 10,000 manual hours—about 850-plus manual hours per month."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A catalogue of document types with volumes and the fields each process needs
- A labelled sample per document type to measure field level accuracy
- Business rules and reference data for validation
- Records rules for how originals and extracted data are kept
Systems to integrate
- Scanning, mailroom, email and portal intake
- Case management, ERP or line of business systems that receive the data
- Reference data and customer or supplier master data for validation
- Records and content management for the originals
Complexity: Medium
Extraction from common document types works out of the box. The effort goes into the long tail of layouts and poor scans, validation against source systems, confidence thresholds per field and a review interface that staff can work in quickly.
- 1
Start with volume and pain
Pick the document types with the highest volume and clearest fields, and measure today's handling time and error rate so the baseline is real.
- 2
Measure accuracy per field
Build a labelled test set per document type and measure accuracy per field, not per document. An average field accuracy of 95% can hide a date field that is wrong half the time.
- 3
Set confidence thresholds per field
Decide per field what confidence and which validation checks allow straight through processing, and start conservatively with more human review.
- 4
Design the review screen
Show the page region next to each extracted value and let reviewers correct with one click. Review speed decides most of the business case.
- 5
Validate against systems of record
Check names, numbers and totals against the case, customer or supplier record before data is accepted, and flag mismatches rather than overwrite.
- 6
Watch for drift
Track corrections by document type and source, and retest when forms, suppliers or scanning change.
Guardrails
- Straight through processing only above field level confidence thresholds and after validation checks
- Every extracted value linked to its location in the source document
- Unrecognized pages and documents always go to a person
- Personal data in documents processed and stored under the same controls as the source system
- Extraction informs decisions; eligibility, benefit or credit decisions stay with the owning process and people
KPIs to instrument
- Field level accuracy per document type on a labelled sample
- Straight through processing rate per document type
- Reviewer time per document and correction rate
- End to end time from receipt to data available in the downstream system
- Errors found downstream that originated in extraction
Human in the loop
Reviewers handle every document below the confidence threshold or failing validation, and their corrections are logged. Process owners set thresholds and approve changes to them, and quality teams sample straight through documents regularly to confirm accuracy holds.
Common failure modes
- High average, weak critical field
- Overall accuracy looks good while one field that drives decisions is often wrong. Measure and threshold per field.
- Silent errors in straight through processing
- Confident but wrong values enter systems unchecked. Validate against source systems and sample straight through documents.
- The long tail stalls the program
- Rare layouts consume the project. Route them to people and automate by volume.
- Extraction becomes the decision
- A missing field triggers an automatic rejection of an application. Keep decisions in the owning process with human review.
What are the risks and rules?
EU AI Act
Depends on design
Classifying documents and extracting data for a person or process to use is usually minimal risk. Even inside an Annex III area, a system that only performs a narrow procedural task, such as splitting and classifying documents, can fall outside the high risk category under Article 6(3); the provider must document that assessment and register the system (Article 6(4) and Article 49(2)). The picture changes when extraction materially influences decisions in Annex III areas, such as eligibility for public assistance benefits (point 5(a)), creditworthiness (point 5(b)) or asylum, visa and residence permit applications (point 7), where the whole system must be assessed as potentially high risk. The Article 6(3) exception never applies when the system performs profiling of natural persons.
Rules that apply
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Points 5 and 7 cover public benefits, credit, and migration and asylum decisions that document processing often feeds.
- Article 22 GDPR, automated individual decision making, including profiling (European Union, Europe). Relevant when extraction results trigger automatic decisions with significant effects on people.
Controls to put in place
- Documented accuracy per field and document type before go live and after changes
- Threshold and routing rules under change control
- Audit trail from each extracted value to the source document and reviewer action
- Retention of originals and extracted data aligned with records rules
- Assessment of whether downstream decisions fall in an Annex III area
Frequently asked questions
- How accurate is AI document extraction?
- It depends on the document type and the field, so measure it per field on your own documents. Google Cloud reports that Ancine, Brazil's cinema industry regulator, reached over 90% data extraction accuracy on digitized tax documents and a tenfold increase in analysts' daily processing capacity. Set confidence thresholds per field and keep people on the uncertain cases.
- What volumes and savings do organizations report?
- Microsoft reports that Volvo Group's solution for invoices, credit notes and claims documents has saved 10,000 manual hours since launch, about 850 a month. Google Cloud reports that Pupuk Indonesia cut data extraction time from 5 to 10 minutes to 40 to 70 seconds, with one employee validating the results. USCIS splits and classifies I-539 applications so that adjudicators find each supporting document faster; it publishes no figures.
- How is this different from invoice processing or correspondence triage?
- Invoice processing is one specialized use of document AI, with purchase order matching and posting. Correspondence triage is about routing incoming mail. This page covers the general capability for forms, applications and supporting documents in any process.
How to cite this page
Blits.ai AI Use Case Library, "AI document intelligence for unstructured forms and documents", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/intelligent-document-processing. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published