What the two services are
Both are building blocks. Neither is an app an accounts payable clerk opens in the morning. You create a resource in your cloud account, choose a model, send documents to it from your own code, and receive structured results: text, fields, tables and confidence scores. Everything before and after that call, from the scanner to the accounting system, is yours to build.
That is not a criticism. It is what makes them flexible enough to sit inside thousands of different products.
Google Document AI
- Google describes Document AI as a platform that turns unstructured documents into structured data, built on Vertex AI with generative AI.
- Processors include Enterprise Document OCR (printed and handwritten text in more than 200 languages), Layout Parser, Form Parser, and prebuilt processors for invoices, expenses, identity documents, bank statements, W-2s and pay slips.
- Custom Extractor, Custom Classifier and Custom Splitter let you train models for your own documents. The Custom Extractor can use Gemini models for extraction with little or no training data.
- You use it through the API or client libraries in C++, C#, Go, Java, Node.js, PHP, Python and Ruby, and set up custom processors in Workbench.
- Processing happens in the Google Cloud location you choose: the US or EU multi-regions, or one of several single regions. We found no on-premises option.
- The Human-in-the-Loop review feature has been deprecated since January 16, 2024, so a review step is something you build.
- Pricing is per page, with an extra hosting charge for custom processors. Google says it never uses customer data to train Document AI models.
Azure Document Intelligence
- Microsoft now calls it Azure Document Intelligence in Foundry Tools, formerly Azure AI Form Recognizer: a cloud service for building intelligent document processing solutions. The current stable API is version 4.0.
- Models include Read, Layout, and prebuilt models for invoices, receipts, IDs, bank statements, checks, contracts, pay stubs, US tax forms and US mortgage forms.
- Custom models: template models for fixed layouts, neural models for varied documents, composed models, and a custom classifier to identify document types first. Microsoft says training a custom model is always free; analysis is paid.
- For generative extraction, Microsoft points to Azure Content Understanding, a separate service with LLM-powered analyzers.
- You use it through the REST API or SDKs for .NET, Java, JavaScript and Python. Labeling tools are moving from Document Intelligence Studio to the Foundry portal.
- Data is processed in the region of your resource. Containers can run on your own hardware: version 4.0 containers cover Read and Layout. Connected containers report usage to Azure for billing; disconnected containers need an approved request and a commitment plan.
- Pricing is pay as you go per page, with commitment tiers. Microsoft stores submitted data and results for 24 hours after an analysis completes, unless you delete them sooner.
Side by side
Swipe the table sideways to see every column.
| Google Document AI | Azure Document Intelligence | CapturePoint 6 | |
|---|---|---|---|
| What it is | Cloud API and processors for developers | Cloud API and models for developers | A Windows app your team uses |
| Where pages are read | Google Cloud, in the location you choose | Azure, in your resource’s region; Read and Layout also in containers | On your own Windows 10 or 11 PC |
| Document types | Prebuilt processors plus Custom Classifier you train | Prebuilt models plus custom classifier you train | Automatic setup finds the types in a folder of your samples |
| Splitting stacks | Custom Splitter you train | Custom classifier you train returns each document’s page range | Automatic, at barcodes, or at separator sheets |
| Review by people | You build it (Human-in-the-Loop deprecated) | You build it | Built-in review screen with the reason for each doubt |
| Improving over time | Retrain or adjust custom processors | Retrain custom models | Learns from every confirmation and correction |
| Where results go | JSON to your code | JSON to your code | Named folders, Content Central, SharePoint or OneDrive, Google Drive, Dropbox, Nucleus One |
| How you pay | Per page, plus hosting for custom processors | Per page, or commitment tiers | Priced per scanning station, with unlimited scanning |
Google and Microsoft facts are from their own documentation, checked on 4 October 2026 (see sources). Both services change often; their documentation wins.
When building on a cloud API is the better choice
Build on Google or Azure if
- You are adding document reading to your own product or portal, and it has to live in your code.
- You have developers who will own the integration, the review screen and the upkeep.
- Documents already arrive in the cloud (uploads, email, other apps), not at a scanner.
- You need a specific prebuilt model, such as identity documents or US tax and mortgage forms, or OCR in many languages.
- Your data and systems already live in Google Cloud or Azure, and region control matters to you.
Use finished software if
- The goal is getting invoices, claims or forms read and filed, not building software.
- Paper arrives at a scanner and someone has to split, check, name and file it today.
- Documents must be read on your own PC, not sent to a cloud service.
- Nobody on staff can maintain an integration when a model or an API version changes.
- You want a predictable cost per scanning station rather than per page.
Plenty of organizations do both: developers build on a cloud API for documents that arrive inside their own product, while the back office uses finished software for the paper.
What you would build: a checklist for estimating the project
If you are leaning toward an API, list the parts below and put a name and a number of weeks next to each. It is the fastest way to see the real size of the job.
- 01
Intake
Getting pages from scanners, shared folders and mailboxes into storage your code can reach, with retries when something fails. - 02
Splitting and typing
Training and testing a splitter and a classifier on your own mixed stacks, then deciding what to do when they disagree with a person. - 03
Extraction
Choosing prebuilt or custom models per document type, labeling samples, and handling the documents that fit neither. - 04
Checks
Rules of your own: line items that add up to the total, dates that make sense, purchase orders that exist. - 05
Review screen
A page where someone sees the image next to the values, fixes mistakes quickly and sends corrections back for retraining. - 06
Output
File names, folders, searchable PDFs and the hand-off to accounting or a document system. - 07
Running it
Monitoring per-page usage and cost, API version upgrades, model retraining, access control and logs.
CapturePoint 6 ships each of these. Here is what extraction and checks look like, with no labeling:
Where Ademero fits
CapturePoint 6: finished capture on your own PC
A Windows 10 or 11 app for the PC next to your scanner. It scans from TWAIN scanners or imports PDF, TIFF, JPEG, PNG, BMP and GIF files, splits stacks, recognizes document types, reads fields and line-item tables, checks the math, and shows people only the documents it is unsure about. Reading and extraction happen on that PC. It learns from every confirmation and correction. See CapturePoint 6.
Paige: finished, in the cloud
If the cloud suits you but you do not want to build, Paige learns your documents from a few samples, splits, sorts and reads them, pulls fields and line items, and delivers the results by download, SFTP or webhook. It runs on Google Cloud. A webhook gives developers clean data without owning the AI.
Cortexa: a private AI platform of your own
For teams that want AI inside their own walls, the Cortexa AI suite is our private AI platform, on servers you own. It gives your team a private AI chat and your developers an AI API for their own apps, and it connects to Content Central, Nucleus One, CapturePoint 6 and Paige, so AI can work inside your document processes: deciding, looking up, routing and moving information between systems. Our team also builds custom AI tools around how you work.
Where these facts come from
Everything this page says about Google Document AI and Azure Document Intelligence comes from Google's and Microsoft's own documentation, checked on 4 October 2026. Products change; if something here is out of date, the vendor's page wins.
- Google: Document AI overview
- Google: Document AI processors list
- Google: Custom Extractor with generative AI
- Google: Document AI client libraries
- Google: Document AI regions
- Google: Document AI deprecations (Human-in-the-Loop)
- Google: Document AI pricing page (pricing model only)
- Google: Document AI security and data handling
- Microsoft: Document Intelligence overview
- Microsoft: Document Intelligence FAQ (formerly Form Recognizer)
- Microsoft: what are Foundry Tools
- Microsoft: custom classification model (document splitting)
- Microsoft: Content Understanding overview
- Microsoft: SDKs and REST API quickstart
- Microsoft: Document Intelligence Studio
- Microsoft: install and run containers
- Microsoft: disconnected containers
- Microsoft: Document Intelligence service limits (free custom training)
- Microsoft: Document Intelligence pricing page (pricing model only)
- Microsoft: Document Intelligence data, privacy and security
Google Cloud and Document AI are trademarks of Google LLC. Azure is a trademark of Microsoft Corporation. Ademero is not affiliated with Google or Microsoft.