Azure Document Intelligence and Textract alternatives for Europe: four options and when to switch
TL;DR
Below about 10,000 pages a month with no residency constraint, the hyperscaler is usually the right tool — keep it. Above roughly 60,000 pages, or once your DPO says the files stay in the EU, a pipeline on local models is cheaper per page and easier to defend.
An EU region does not change who operates the service. The US CLOUD Act reaches a US-headquartered provider wherever the server sits, and the EU–US transfer framework has fallen twice since 2015.
Against a document API at €0.01 per page, a €600-a-month reserved EU GPU crosses over near 60,000 pages; against a €0.03 tier with tables and custom fields, near 20,000.
Custom models trained in a vendor’s format cannot be exported. Three years of corrections become an asset you rent.
Do not migrate by reproducing the vendor’s JSON. Freeze the fields your ERP consumes, score both pipelines on 300–500 human-corrected pages, and cut over one document family at a time.
Questions people ask
What are the alternatives to Azure Document Intelligence in Europe?
Four categories: EU-hosted document APIs from European providers such as Mistral OCR; self-hosted open-weight models like PaddleOCR-VL, DeepSeek-OCR and Qwen3-VL on your own GPU; managed pipelines that run those models with a schema, validation and delivery; and European extraction SaaS such as Rossum or Parseur. Which fits depends on monthly page volume, residency rules and whether you have an engineering team to run the stack.
Is Azure Document Intelligence GDPR compliant if I use an EU region?
An EU region keeps the bytes on European soil but does not change who operates the service. Microsoft, Amazon and Google remain US-headquartered and subject to the US CLOUD Act, so your DPO still needs a transfer basis under GDPR Art. 44–49 — one that has been rewritten twice since Safe Harbor fell in 2015. For many workloads that is enough; teams handling Art. 9 health data often prefer a design where the question never arises.
At what volume is a local model cheaper than Textract or Azure Document Intelligence?
In the article’s illustrative scenario, a reserved EU A100-class GPU at about €600 a month crosses a document API priced at €0.01 per page near 60,000 pages a month, and a €0.03-per-page tier with tables and custom fields near 20,000 pages. Both crossovers come earlier than against a token-billed LLM, because a document API charges the full page rate even for a clean, cheap page. Engineering time is excluded on both sides.
How do I migrate from Textract or Azure Document Intelligence to another provider?
Do not try to reproduce the vendor’s JSON. Freeze the fields your ERP actually consumes, export 300–500 pages per document family with the human-corrected values as ground truth, write one normalizer for the old output so you can score it fairly, run the candidate on the same sample, set a confidence gate, parallel-run for two to four weeks, and cut over one family at a time — keeping the hyperscaler for any family it still wins on.
When should a European team stay on Azure, Textract or Google Document AI?
When volume is under about 10,000 pages a month and the DPO has signed off on the EU region and transfer mechanism; when the documents match the prebuilt catalog, such as US tax forms or English receipts; when the API is already wired into that cloud’s storage and queues; when volume is spiky rather than steady; or when there is no engineering capacity and no appetite for a managed provider.
Which open-weight models can replace Azure Document Intelligence?
PaddleOCR-VL, a compact Apache 2.0 vision-language model with wide language coverage; DeepSeek-OCR, released under MIT, which turns a page into markdown; and Qwen3-VL, available in Apache-licensed sizes from 2B parameters upward for schema-driven extraction. All are small enough to run on one card — about €0.0015 per page of compute on a rented EU A100-class GPU. The cost that is not small is being the team on call for the queue.
Want this worked out on your documents?
We will price a real sample against OpenAI or Anthropic and tell you whether Bulk document processing or a local model on your hardware is the cheaper first step.
For a European team, the realistic alternatives to Azure AI Document Intelligence, Amazon Textract and Google Document AI fall into four groups: EU-hosted document APIs, self-hosted open-weight models such as PaddleOCR-VL, DeepSeek-OCR and Qwen3-VL, managed pipelines that run those models for you, and European extraction SaaS. Two numbers decide which group fits. Below about 10,000 pages a month with no residency constraint, the hyperscaler is usually the right tool and you should keep it. Above roughly 60,000 pages a month — or at any volume once your DPO says the files stay in the EU — a pipeline on local models is cheaper per page and easier to defend in a data-protection review.
We do not quote vendor list prices below; they change and are published per region. Where we do cost math it is an illustrative scenario, with the assumptions stated in the table caption.
What Azure, Textract and Document AI offer today
The three products are more alike than their marketing suggests. Each exposes an OCR layer, a set of prebuilt extractors for common document types, and a way to train a custom model on your own labeled samples. Each returns JSON, bills per page, and is available in at least one EU region.
Azure AI Document Intelligence (formerly Form Recognizer): read and layout models; prebuilt models for invoices, receipts, identity documents, contracts and a set of US tax and banking forms; custom extraction (template and neural) and custom classification. Output is an analyzeResult JSON with typed fields, confidence scores and bounding regions. Billing is per 1,000 pages, tiered by model class. Container images exist for several models, but they still meter usage back to Azure.
Amazon Textract: DetectDocumentText for plain OCR; AnalyzeDocument with Forms, Tables, Queries, Signatures and Layout; AnalyzeExpense for invoices and receipts; AnalyzeID for identity documents; Custom Queries adapters trained on your samples. Billing is per page and per feature, so a page analyzed for forms and tables is charged for both. Output is a flat list of Blocks (page, line, word, key-value set, table, cell) linked by IDs.
Google Document AI: Document OCR, Form Parser and Layout Parser; specialized processors for invoices, expenses and identity documents; a Custom Extractor built on a foundation model, with optional fine-tuning; a custom classifier. Billing is per page by processor type. Output is a Document object whose entities point into the page text by character offset. An EU multi-region endpoint keeps processing inside the EU.
On price, the prebuilt invoice and receipt endpoints at all three are published at around one US cent per page at entry volumes at the time of writing; forms, tables, queries and custom or generative extraction are billed as add-ons or at higher per-page tiers. Check the vendor’s pricing page for your region — a realistic page with tables and key-value pairs often costs several times the headline rate.
Why European teams look for alternatives
A US-owned provider, even in an EU region
Choosing West Europe, Frankfurt or the Google EU endpoint keeps the bytes on European soil. It does not change who operates the service. The US CLOUD Act (2018) lets US authorities compel a US-headquartered provider to produce data it holds, regardless of where the server sits. Microsoft, Amazon and Google publish transparency reports, contest requests they consider overbroad, and have built EU Data Boundary and sovereign-cloud offerings in response; none of that is wrongdoing, and for many workloads it is enough. But the transfer story your DPO writes under GDPR Art. 44–49 has been rewritten twice already: Safe Harbor fell in 2015, Privacy Shield fell with Schrems II in 2020, and the 2023 EU–US Data Privacy Framework is the third attempt. Teams in DACH finance, Nordic public sector and anything touching Art. 9 health data increasingly prefer a design where the question never arises — the argument in GDPR-compliant document AI.
Per-page bills that track volume
Per-page pricing is easy to start and hard to stop. At 5,000 invoices a month the bill is pocket money. At 100,000 pages a month with tables and custom fields it is a budget line that grows with the business and never gets cheaper per page. A reserved EU GPU is the opposite shape: about €600 a month whether you send it 20,000 pages or 200,000.
Illustrative scenario, assumptions stated here: a document API at an assumed blended €0.01 per page (prebuilt model only) and €0.03 per page (with tables and custom fields), versus a reserved EU A100-class GPU at €600 per month. Vendor tiers at the time of writing differ by region and feature — check the vendor. Engineering time excluded on both sides.
Volume / month
API at €0.01 / page
API at €0.03 / page
Reserved GPU
10,000 pages
€100
€300
€600
20,000 pages
€200
€600
€600
60,000 pages
€600
€1,800
€600
100,000 pages
€1,000
€3,000
€600
250,000 pages
€2,500
€7,500
€600
Illustrative scenario, assumptions stated here: a document API at an assumed blended €0.01 per page (prebuilt model only) and €0.03 per page (with tables and custom fields), versus a reserved EU A100-class GPU at €600 per month. Vendor tiers at the time of writing differ by region and feature — check the vendor. Engineering time excluded on both sides.
The crossover against the cheap tier sits near 60,000 pages a month; against a richer tier it is near 20,000. Both come earlier than the 85,000-page crossover against GPT-4.1 in the 100,000-page invoice cost model, because a document API charges the full page rate even for a clean, cheap page, while a token bill shrinks with it.
Lock-in to proprietary output formats
Textract Blocks, Azure’s analyzeResult and Google’s Document object are three different ways of saying “here is text with coordinates.” If your ERP integration, validation rules and test fixtures are written against one of them, switching vendors means rewriting all three. The custom models are worse: the labeled training sets live in vendor-specific formats, and the trained model cannot be exported and run elsewhere. Three years of corrections become an asset you rent.
Gaps in local languages and handwriting
The OCR layers cover most European scripts. The prebuilt invoice and receipt models are trained and validated on a narrower list of locales, and handwriting support is narrower still — Textract, for example, documents handwriting for English only at the time of writing. A Latvian, Estonian or Slovenian supplier invoice is often “supported” in the sense that the characters are read, and unsupported in the sense that VAT and totals are mislabeled. Check the supported-locale page for the exact model, then test 50 of your own pages before believing it.
Four categories of alternatives
Alternative categories to hyperscaler document APIs for EU teams. Each row describes the category; individual vendors vary, and their claims should be verified against their own documentation.
Category
What it is
Data location
Best for
Watch out for
EU-hosted cloud document APIs
Per-page OCR and extraction APIs from EU-headquartered providers (Mistral OCR is one example)
Provider’s EU infrastructure; verify sub-processors
Teams that want API convenience under an EU contract
Still a transfer to a processor; feature depth varies; formats are proprietary too
Self-hosted open-weight models
PaddleOCR-VL, DeepSeek-OCR or Qwen3-VL running on your GPU
Your hardware or your EU tenant
Engineering teams with steady volume and residency rules
You own the model and the ops: queues, retries, monitoring, upgrades
Managed pipelines on local models
A provider runs open-weight models with schema, validation and delivery (what Ækora does)
Provider’s EU-hosted stack or your premises
Ops and finance teams without an ML team
Fewer prebuilt document types than a hyperscaler catalog; per-page tiers
European SaaS extraction tools
Products such as Rossum or Parseur with a UI, templates and connectors
Vendor’s hosting; check region and DPA
AP teams and SMBs who want a validation screen and integrations
Volume-tiered pricing; the model stays with the vendor
Alternative categories to hyperscaler document APIs for EU teams. Each row describes the category; individual vendors vary, and their claims should be verified against their own documentation.
EU-hosted cloud document APIs
The contract is European, the infrastructure is European, the model is still someone else’s. It is a reasonable middle step: it removes the CLOUD Act question without asking you to run anything. Ask two things before signing. Where is the endpoint actually hosted — a European brand on a US hyperscaler is a different answer — and is the output schema documented well enough to normalize into yours?
Self-hosted open-weight models
The open-weight document models that matter in 2026 are small enough to run on one card. PaddleOCR-VL is a compact vision-language model under an Apache 2.0 license with wide language coverage; DeepSeek-OCR is released under MIT and turns a page into markdown; Qwen3-VL ships Apache-licensed sizes from 2B parameters upward, so you can pick a general VLM for schema-driven extraction. The model-by-model view is in the 2026 guide to local OCR and document models. The compute is cheap — about €0.0015 per page on a rented EU A100-class card at €1.80 an hour and 20 pages a minute. What is not cheap is being the team on call for the queue.
Managed pipelines on local models
This is the category we sell, so weigh the description accordingly. A managed pipeline takes your document families, agrees a schema per family, runs open-weight models on EU infrastructure or your own hardware, validates the output against rules, and delivers CSV, JSON or an API call. Ækora’s published tiers are €999 a month for 10,000 pages and €4,999 a month for 75,000, with per-page overage — priced against pages that arrive, not tokens or features. Bulk document processing lists what is included; local document models is the same stack on your premises, the pattern described in what on-premise document AI means.
European SaaS extraction tools
Rossum, Parseur and similar tools sell a user interface: a validation screen, template editors, mailbox intake, ERP connectors. If your accounts-payable team wants to see and correct each invoice before it posts, that UI is the product and the model behind it is secondary. Data location depends on the vendor’s hosting; ask the same question as for the EU-hosted APIs.
How the options compare on the dimensions that matter
Qualitative comparison across the six dimensions European buyers ask about most. ‘Exit’ means what you keep if you leave: your data and your schema are always yours; the model and the labeled corrections often are not.
Dimension
Hyperscaler API
EU-hosted API
Self-hosted open-weight
Managed local pipeline
European SaaS
Data location
EU region of a US provider
EU provider; verify hosting
Your hardware or tenant
EU stack or your premises
Vendor hosting; verify
Pricing model
Per page, per feature
Per page
Fixed GPU plus your time
Fixed monthly tier plus overage
Per-document or per-seat tiers
Custom schemas
Custom models, labeled in vendor format
Varies; often prompt-based
Any schema; prompt or fine-tune
One schema per family, agreed up front
Template or field editor
Languages
Broad OCR; narrower prebuilt locales
Varies by provider
Model-dependent; wide in Qwen3-VL and PaddleOCR-VL
Tested per family before go-live
Varies; check
Ops burden
Integration only
Integration only
Full: GPUs, queues, upgrades, on-call
Provider runs it
None
Exit
Rewrite parsers; models stay
Rewrite parsers; models stay
You keep everything
You keep schema, samples and outputs; models are open-weight
Export data; templates and models stay
Qualitative comparison across the six dimensions European buyers ask about most. ‘Exit’ means what you keep if you leave: your data and your schema are always yours; the model and the labeled corrections often are not.
A migration path from Textract or Azure JSON to a schema-first pipeline
The mistake in most migrations is trying to reproduce the vendor’s JSON. Do not. Reproduce the twelve fields your ERP actually consumes.
Freeze the target schema. List the fields each downstream system reads from the current output — supplier name, VAT ID, invoice number, dates, net, VAT, gross, currency, line items. Give each a type and a validation rule. Everything else in the vendor JSON is noise.
Export a ground-truth sample. Pull 300–500 pages per document family with the raw API JSON and, more importantly, the values as finally posted after human correction. The corrected values are the truth; the API output is not.
Write one normalizer for the old output. A small script that maps Blocks or analyzeResult into the schema. You need it to score the incumbent fairly and to keep the API as a fallback later.
Run the candidate on the same sample. Same pages, same schema, field-level comparison against the corrected values. Count exact matches for amounts and dates; count normalized matches for names and addresses.
Set the confidence gate. Decide the threshold below which a page goes to a human or to the old API. Local models for the routine 95%, an exception path for the rest — that hybrid is what keeps the migration boring.
Parallel-run for two to four weeks. Both pipelines process live traffic; only the incumbent posts to the ERP. Compare straight-through rate and touch rate per family.
Cut over per family. Move invoices first, keep the hyperscaler for any family it still wins on, and cancel what is left once the parallel run shows nothing.
If the incumbent has custom models, step 2 is where you recover their value: the human-corrected outputs are your labeled set now, in your format, usable to prompt or fine-tune any open-weight model.
When to stay on the hyperscaler
Volume is under about 10,000 pages a month and your DPO has signed off on the EU region, the DPA and the transfer mechanism. The bill is smaller than the salary fraction it would take to replace it.
Your documents are what the prebuilt catalog was built for — US tax forms, US identity documents, English-language expense receipts.
You are already deep in that cloud and the document API feeds that vendor’s storage, queues and functions. Moving one link in the chain buys little.
Volume is spiky rather than steady. Hardware wants a floor; a burst of 300,000 pages once a year is what a metered API is for.
You have no engineering capacity and do not want a managed provider. The hyperscaler console is a working product today.
The EU AI Act does not change this list for most extraction work; the high-risk obligations that started applying on 2 August 2026 bite on specific uses, not on reading invoices. EU AI Act obligations for document processing covers which uses need more than a DPA.
Measure three things this quarter: pages per month per document family, the fully loaded per-page cost on last month’s invoice (including the features you forgot you enabled), and the answer your DPO gives when asked where the pages go. If the first is above 60,000, the second is above €0.01, or the third is uncomfortable, run the seven steps above on 300 of your own pages. The European buyer’s guide to document processing on local models lists the questions to put to every vendor on the shortlist.